text
stringlengths 47
469k
| meta
dict | domain
stringclasses 1
value |
|---|---|---|
---
bibliography:
- 'references.bib'
---
HU Berlin-EP-03/41\
hep-th/0307229
[**On the propagator of a scalar field in ${\text{AdS}}\times\text{S}$ and in its plane wave limit**]{}
Harald Dorn, Mario Salizzoni and Christoph Sieg [^1]
Humboldt–Universität zu Berlin, Institut für Physik\
Newtonstraße 15, D-12489 Berlin\
[**Abstract**]{}
We discuss the scalar propagator on generic ${{\text{AdS}}_{d+1}\times \text{S}^{d'+1}}$ backgrounds. For the conformally flat situations and masses corresponding to Weyl invariant actions, the propagator is powerlike in the sum of the chordal distances with respect to ${\text{AdS}}_{d+1}$ and $\text{S}^{d'+1}$. In these cases we analyze its source structure. In all other cases the propagator depends on both chordal distances separately. There an explicit formula is found for certain special mass values. For pure AdS we show how the well known propagators in the Weyl invariant case can be expressed as linear combinations of simple powers of the chordal distance. For $\text{AdS}_5\times\text{S}^5$ we relate our propagator to the expression in the plane wave limit and find a geometric interpretation of the variables occurring in the known explicit construction on the plane wave. As a byproduct of comparing different techniques, including the KK mode summation, a theorem for summing certain products of Legendre and Gegenbauer functions is derived.
Introduction
============
The AdS/CFT correspondence [@Maldacena:1998re] relates $\mathcal{N}=4$ super Yang-Mills gauge theory in Minkowski space to type ${\text{II}\,\text{B}}$ string theory in ${\text{AdS}}_5\times\text{S}^5$ with some RR background flux. In the supergravity approximation one handles the fields in a Kaluza-Klein mode expansion with respect to the $\text{S}^5$. For calculations on the supergravity side the propagators of the whole spectrum of fields in the ${\text{AdS}}_5$ background are an essential technical ingredient. The simplest case to start with is of course the well known scalar propagator [@Burgess:1985ti; @D'Hoker:2002aw]. Explicit tests of the AdS/CFT correspondence, beyond the supergravity approximation, remain a difficult task, since the relevant string spectrum in general is not available. In a limit of this correspondence, proposed by Berenstein, Maldacena and Nastase (BMN limit) [@Berenstein:2002jq], the ${\text{AdS}}_5\times\text{S}^5$ background itself is transformed via a Penrose limit [@Penrose:1976] to a certain plane wave background found by Blau, Figueroa-O’Farrill, Hull and Papadopoulos [@Blau:2001ne; @Blau:2002dy; @Blau:2002mw]. In this background, for brevity called ‘the plane wave’, string theory is exactly quantizable [@Metsaev:2001bj], and thus enables independent checks of the duality, including string effects. In this plane wave background the separation between the ${\text{AdS}}_5$ and the $\text{S}^5$ part breaks down, and one has to take the limit on full 10-dimensional ${\text{AdS}}_5\times\text{S}^5$ objects.
One of the crucial unsolved questions in this setting concerns the issue of holography [@Berenstein:2002sa; @Das:2002cw; @Leigh:2002pt; @Kiritsis:2002kz; @Dobashi:2002ar]. In the Penrose limiting process the old 4-dimensional conformal boundary is put beyond the new plane wave space, which by itself has a one-dimensional conformal boundary. In [@Dorn:2003ct] we started to investigate this issue and found, for each point remaining in the final plane wave, a degeneration of the cone of boundary reaching null geodesics into a single direction. To continue this program beyond geometric properties, we now want to study the limiting process for field theoretical propagators.
The so called bulk-to-boundary propagator plays an essential role in the holographic description of the AdS/CFT correspondence and is therefore a quantity of particular interest. However, as follows from the results in [@Dorn:2003ct], a reasonable Penrose limit cannot be taken due to the fact that one of its legs ends at the boundary and hence lies outside the region of convergence to the plane wave. The situation is different for the bulk-to-bulk propagator. One has the choice to let both legs end within the region that converges to the plane wave in the limit. Although this propagator first seems to be of minor importance for the realization of holography and even for the computation of correlation functions [@D'Hoker:1999ni], it might be very useful for defining a bulk-to-boundary propagator in the plane wave. A hint that this could be a promising direction is, that in pure ${\text{AdS}}_{d+1}$ the knowledge of the bulk-to-bulk propagator is sufficient for deriving the corresponding bulk-to-boundary propagator.
Due to the existence of such a relation it seems to be worthwhile to study also in ${\text{AdS}}\times\text{S}$ the the bulk-to-bulk propagator to get information about the bulk-boundary correspondence in the plane wave limit. For brevity when we talk about ‘the propagator’ in the following we always understand it as the bulk-to-bulk one.
The scalar propagator in the plane wave has been constructed in [@Mathur:2002ry] by a direct approach leaving the issue of its derivation via a limiting process from ${\text{AdS}}_5\times\text{S}^5$ as an open problem.
Motivated by the above given questions, and because it is an interesting problem in its own right, we will study in this paper the construction of the scalar propagator on ${{\text{AdS}}_{d+1}\times \text{S}^{d'+1}}$ spaces with radii $R_1$ and $R_2$, respectively. Allowing for generic dimensions $d$ and $d'$ as well as generic curvature radii $R_1$ and $R_2$ is very helpful to understand the general mechanism for the construction of the propagator. Of course only some of these spacetimes are parts of consistent supergravity backgrounds.
In Section \[waveeqapp\] we will focus on the differential equation defining the scalar propagator. Within this Section we will be able to find the propagator in conformally flat situations, i.e. $R_1=R_2$ and for masses corresponding to Weyl invariant actions.
The next two Sections mainly serve as a kind of interpretation of the results of Section \[waveeqapp\]. Using Weyl invariance, we map patchwise to flat space in Section \[weylana\] and globally to the Einstein Static Universe (ESU) in Section \[relESU\]. This includes a discussion of global aspects of the solutions, like their boundary conditions and $\delta$-source structure.
With the hope to get the propagator for generic masses, in Section \[modesummation\] we study its KK mode sum. We will be able to perform the sum for a linear relation between the conformal dimension of the KK mode and the quantum number parameterizing the eigenvalue of the Laplacian on the sphere. Beyond the cases treated in the previous Sections this applies to certain additional mass values, but fails to solve the full generic problem. As a byproduct, the comparison with the result of Section \[waveeqapp\] yields a theorem on the summation of certain products of Gegenbauer and Legendre functions.
In Section \[pwlimit\] we will discuss the plane wave limit of ${\text{AdS}}_5\times\text{S}^5$ in brief. We will explicitly show that the massless propagator on the full spacetime indeed reduces to the expression of [@Mathur:2002ry]. Furthermore, we will present the limit of the full differential equation which is fulfilled by the propagator of massive scalar fields given in [@Mathur:2002ry]. Finally, our results will be summarized in Section \[concl\].
In Appendix \[bulkboundproprel\] a detailed derivation of the relation between the bulk-to-bulk and the bulk-to-boundary propagator in the case of pure Euclidean ${\text{AdS}}_{d+1}$ is presented. Appendix \[userel\] summarizes some of the relevant formulae for hypergeometric functions and spherical harmonics we needed for the analysis. Furthermore we will sketch an independent proof of the theorem that has been extracted from Sections \[waveeqapp\] and \[modesummation\]. Appendix \[chordaldistpw\] contains a short review of the embedding of the plane wave in flat spacetime.
The differential equation for the propagator and its solution {#waveeqapp}
=============================================================
The scalar propagator on ${{\text{AdS}}_{d+1}\times \text{S}^{d'+1}}$ {#AdSSderiv}
---------------------------------------------------------------------
The scalar propagator is defined as the solution of[^2] $$\label{waveeq}
(\Box_z-M^2)G(z,z')=\frac{i}{\sqrt{-g}}\delta(z,z'){~,}$$ with suitable boundary conditions at infinity. $\Box_z$ denotes the d’Alembert operator on ${{\text{AdS}}_{d+1}\times \text{S}^{d'+1}}$, acting on the first argument of the propagator $G(z,z')$. In the following we denote the coordinates referring to the ${\text{AdS}}_{d+1}$ factor by $x$ and those referring to the $\text{S}^{d'+1}$ factor by $y$, i.e. $z=(x,y)$. We first look for solutions at $z\neq z'$ and discuss the behaviour at $z=z'$ afterwards.
${\text{AdS}}_{d+1}$ and $\text{S}^{d'+1}$ can be interpreted as embeddings respectively in $\mathds{R}^{2,d}$ and in $\mathds{R}^{d'+2}$ with the help of the constraints[^3] $$\label{defeq}
-X_0^2-X_{d+1}^2+\sum_{i=1}^d
X_i^2=-R_1^2{~,}\qquad\sum_{i=1}^{d'+2}Y_i^2=R_2^2{~,}$$ where $X=X(x)$, $Y=Y(y)$ depend on the coordinates $x$ and $y$, respectively. We define the chordal distances on both spaces to be $$\label{uv}
u(x,x')=(X(x)-X(x'))^2{~,}\qquad v(y,y')=(Y(y)-Y(y'))^2{~.}$$ The distances have to be computed with the corresponding flat metrics of the embedding spaces that can be read off from . The chordal distance $u$ is a unique function of $x$ and $x'$ if one restricts oneself to the hyperboloid. On the universal covering it is continued as a periodic function. For later use we note that on the hyperboloid and on the sphere the antipodal points $\tilde x$ and $\tilde y$ to given points $x$ and $y$ are defined by changing the sign of the embedding coordinates $X$ and $Y$ respectively. From one then finds with $\tilde u=u(x,\tilde x')$, $\tilde v=v(y,\tilde y')$ $$\label{uvaprel}
u+\tilde u=-4R_1^2{~,}\qquad v+\tilde v=4R_2^2{~.}$$ Using the homogeneity and isotropy of both ${\text{AdS}}_{d+1}$ and $\text{S}^{d'+1}$ it is clear that the propagator can depend on $z,z'$ only via the chordal distances $u(x,x')$ and $v(y,y')$. Strictly speaking this at first applies only if ${\text{AdS}}_{d+1}$ is restricted to the hyperboloid. Up to subtleties due to time ordering (see the end of Section \[relESU\]) this remains true also on the universal covering. The d’Alembert operator then simplifies to $$\label{dalemchordal}
\begin{aligned}
\Box_z&=\Box_x+\Box_y{~,}\\
\Box_x&=2(d+1)\Big(1+\frac{u}{2R_1^2}\Big){\frac{\partial}{\partial u}}+\Big(\frac{u^2}{R_1^2}+4u\Big){\frac{\partial^2}{\partial {u}^2}}{~,}\\
\Box_y&=2(d'+1)\Big(1-\frac{v}{2R_2^2}\Big){\frac{\partial}{\partial v}}-\Big(\frac{v^2}{R_2^2}-4v\Big){\frac{\partial^2}{\partial {v}^2}}{~.}\end{aligned}$$
One can now ask for a solution of at $z\neq z'$ that only depends on the total chordal distance $u+v$. Indeed, using , it is easy to derive that such a solution exists if and only if $$\label{conditions}
R_1=R_2=R{~,}\qquad M^2=\frac{d'^2-d^2}{4R^2}{~.}$$ Furthermore, it is necessarily powerlike and given by $$G(z,z')\propto (u+v)^{-\frac{d+d'}{2}}{~.}$$ Extending this to $z=z'$ we find just the right power for the short distance singularity to generate the $\delta$-function on the r.h.s. of . Hence after fixing the normalization we end up with $$\label{AdSSprop}
G(z,z')=\frac{\Gamma(\frac{d+d'}{2})}{4\pi^{\frac{d+d'}{2}+1}}\frac{1}{(u+v+i\varepsilon(t,t'))^\frac{d+d'}{2}}{~.}$$ Note that due to besides the singularity at $z=z'$ there is another one at the total antipodal point where $z=\tilde z'=(\tilde x', \tilde y')$. We have introduced an $i\varepsilon$-prescription by replacing $u\to
u+i\varepsilon$, where $\varepsilon$ depends explicitly on time. We will comment on this in Section \[relESU\]. In particular, we will see that on the universal covering of the hyperboloid the singularity at the total antipodal point does not lead to an additional $\delta$-source on the r.h.s. of .
Scalar fields with mass $m^2$ in ${\text{AdS}}_{d+1}$ via the ${\text{AdS}}/\text{CFT}$ correspondence are related to CFT fields with conformal dimension $$\label{conformaldim}
\Delta_\pm(d,m^2)=\frac{1}{2}\Big(d\pm\sqrt{d^2+4m^2R_1^2}\Big){~.}$$ Note that the exponent of $(u+v)$ in the denominator of the propagator is just equal to $\Delta_+(d,M^2)$. From the ${\text{AdS}}_{d+1}$ point of view the $(d+d'+2)$-dimensional mass $M^2$ is the mass of the KK zero mode of the sphere. We will say more on these issues in Section \[modesummation\].\
For completeness let us add another observation. Disregarding for a moment the source structure, under the conditions there is a solution of , that depends only on $(u-v)$. The explicit form is $$\label{AdSSmirprop}
\tilde G(z,z')\propto\frac{1}{(u-v+4R^2+i\varepsilon(t,t'))^\frac{d+d'}{2}}{~.}$$ It has the same asymptotic falloff as . But due to it has singularities only at the semi-antipodal points where $z=z'_\text{s}=(x',\tilde y')$ and $z=\tilde z'_\text{s}=(\tilde x',y')$. We will say more on $\tilde G(z,z')$ in Sections \[weylana\] and \[relESU\].\
At the end of this Subsection we give a simple interpretation of the conditions . The equality of the radii is exactly the condition for conformal flatness of the complete product space ${{\text{AdS}}_{d+1}\times \text{S}^{d'+1}}$ as a whole. Describing the AdS metric in Poincaré coordinates ($x=(x^0,x^1,\dots,x^{d-1},x_{\perp })$) one finds $$\label{AdSmetric}
{\operatorname{d}\!}s^2=\frac{R_1^2}{x_{\perp}^2}\big(-({\operatorname{d}\!}x^0)^2+{\operatorname{d}\!}x_{\perp}^2+{\operatorname{d}\!}\vec x^2\big){~.}$$ For ${{\text{AdS}}_{d+1}\times \text{S}^{d'+1}}$ one thus obtains $$\label{AdSSmetric}
{\operatorname{d}\!}s^2=\frac{R_1^2}{x_{\perp}^2}\left ({-({\operatorname{d}\!}x^0)^2+{\operatorname{d}\!}x_{\perp}^2+{\operatorname{d}\!}\vec
x^2}+\frac{R_2^2}{R_1^2}x_{\perp}^2{\operatorname{d}\!}\Omega_{d'+1}^2\right ){~,}$$ which is obviously conformally flat if $R_1=R_2$. That this is also necessary for conformal flatness follows from an analysis of the corresponding Weyl tensor. Furthermore, the mass condition just singles out the case of a scalar field coupled in Weyl invariant manner to the gravitational background. The corresponding $D$-dimensional action is $$S=-\frac{1}{2}\int{\operatorname{d}\!}^Dz\,\sqrt{-g}\Big[g^{\mu\nu}\partial_\mu\phi\partial_\nu\phi+\frac{D-2}{4(D-1)}\mathcal{R}\phi^2\Big]{~.}$$ Inserting the constant curvature scalar $\mathcal{R}$ for ${{\text{AdS}}_{d+1}\times \text{S}^{d'+1}}$ with equal radii one gets for the mass just the value in .
Altogether in this Subsection we have constructed the scalar ${{\text{AdS}}_{d+1}\times \text{S}^{d'+1}}$ propagator for the case of Weyl invariant coupling to the metric in conformally flat situations. The Weyl invariant coupled field is the natural generalization of the massless field in flat space.
A remark on the propagator on pure ${\text{AdS}}_{d+1}$ {#AdSderiv}
-------------------------------------------------------
Having found for ${{\text{AdS}}_{d+1}\times \text{S}^{d'+1}}$ such a simple expression for the scalar propagator, one is wondering whether the well known AdS propagators can also be related to simple powers of the chordal distance.
The general massive scalar propagator on pure ${\text{AdS}}_{d+1}$ space corresponding to the two distinct conformal dimensions $\Delta_{\pm}$ defined in with generic mass values is given by [@Burgess:1985ti; @D'Hoker:2002aw] $$\label{AdSprop}
G_{\Delta_\pm}(x,x')=\frac{\Gamma(\Delta_\pm)}{R_1^{d-1}2\pi^\frac{d}{2}\Gamma(\Delta_\pm-\frac{d}{2}+1)}\Big(\frac{\xi}{2}\Big)^{\Delta_\pm}{F\big(\tfrac{\Delta_\pm}{2},\tfrac{\Delta_\pm}{2}+\tfrac{1}{2};\Delta_\pm-\tfrac{d}{2}+1;\xi^2\big)}{~,}\quad\xi=\frac{2R_1^2}{u+2R_1^2}{~.}$$
Again, here a powerlike solution of (but now using only the AdS d’Alembert operator of and replacing $M^2$ by the AdS mass $m^2$) exists for the Weyl invariant coupled mass value[^4] $$\label{AdSWeylmass}
m^2=\frac{1-d^2}{4R_1^2}{~.}$$ The related value for the conformal dimension from is then $\Delta_{\pm}=\frac{d\pm 1}{2}$. The powerlike solution is given by $$\label{AdSprop2}
G(x,x')=\frac{\Gamma(\frac{d-1}{2})}{4\pi^{\frac{d+1}{2}}}\frac{1}{(u+i\varepsilon(t,t'))^\frac{d-1}{2}}{~.}$$In contrast to the ${{\text{AdS}}_{d+1}\times \text{S}^{d'+1}}$ case here the exponent of $u$ is given by $\Delta_-(d,m^2)$. We have again kept the option of a time dependent $i\varepsilon(t,t')$ and will comment on it in Section \[relESU\].
The above solution can indeed be obtained from by taking the sum of the expressions for $\Delta_+$ and $\Delta_-$. In addition one finds another simple structure by taking the difference. They are given by $$\label{AdSsuperpos}
\begin{aligned}
\frac{1}{2}(G_{\Delta_-}+G_{\Delta_+})&=\frac{\Gamma(\frac{d-1}{2})}{4\pi^{\frac{d+1}{2}}}\frac{1}{(u+i\varepsilon(t,t'))^\frac{d-1}{2}}{~,}\\
\frac{1}{2}(G_{\Delta_-}-G_{\Delta_+})&=\frac{\Gamma(\frac{d-1}{2})}{4\pi^{\frac{d+1}{2}}}\frac{1}{(u+4R^2+i\varepsilon(t,t'))^\frac{d-1}{2}}{~.}\end{aligned}$$ Both expressions are derived by using and of Appendix \[userel\]. The first combination has the right short distance singularity to be a solution of . The second combination resembles . We will say more on this in Sections \[weylana\] and \[relESU\].
Comment on masses and conformal dimensions on ${\text{AdS}}_{d+1}$ {#BFbounds}
------------------------------------------------------------------
On AdS spaces one has to respect the Breitenlohner-Freedman bounds [@Breitenlohner:1982bm; @Breitenlohner:1982jf]. To get real values for $\Delta_\pm$ requires $$\label{BFbound}
m^2\geq -\frac{d^2}{4R_1^2}{~.}$$ Furthermore, the so called unitarity bound requires $$\label{splitbound}
\Delta >\frac{d-2}{2}{~.}$$ This implies that for $-\frac{d^2}{4R_1^2}\leq m^2 < \frac{4-d^2}{4R_1^2}$ both, $\Delta _+$ and $\Delta _-$ are allowed. On the other side for $\frac
{4-d^2}{4R_1^2}\leq m^2$ only $\Delta _+$ is allowed.
The masses for Weyl invariant coupling are $\frac{1-d^2}{4R_1^2}$ and $\frac{d'^2-d^2}{4R_1^2}$ for ${\text{AdS}}_{d+1}$ and ${{\text{AdS}}_{d+1}\times \text{S}^{d'+1}}$, respectively. Hence in our Weyl invariant cases for pure AdS $\Delta _+$ and $\Delta _-$ are allowed while for ${{\text{AdS}}_{d+1}\times \text{S}^{d'+1}}$ with $d'>1$ only $\Delta _+$ is allowed.
Derivation of the propagator from the flat space one {#weylana}
====================================================
In the previous Section we have shown that a simple powerlike solution of can be found if the underlying spacetime is ${\text{AdS}}_{d+1}$ or a conformally flat product space ${{\text{AdS}}_{d+1}\times \text{S}^{d'+1}}$ and if the corresponding scalar field is Weyl invariant coupled to the curvature of the background. Both properties allow for a mapping of the differential equation, the scalar field and the propagator to flat space. The other way around, one can use Weyl invariance in this special case to construct the propagator of Weyl invariant coupled fields on conformally flat backgrounds from the flat space massless propagator.
We will use this standard construction to rederive the ${{\text{AdS}}_{d+1}\times \text{S}^{d'+1}}$ expressions and from the flat space solutions.
The relevant Weyl transformation in a $D$-dimensional manifold is $$\label{weyltraf1}
g_{\mu\nu}\to \varrho\,g_{\mu\nu}{~,}\qquad
\phi\to\phi'=\varrho^\frac{2-D}{4}\phi{~.}$$ If then the metric is of the form $g_{\mu\nu}(z)=\varrho(z)\,\eta_{\mu\nu}$ one finds the following relation between the propagator in curved and flat space $$G(z,z')=\big(\varrho(z)\,\varrho(z')\big)^{\frac{2-D}{4}}\,G_\text{flat}(z,z'){~,}\qquad
G_\text{flat}(z,z')=\frac{\Gamma(\frac{D-2}{2})}{4\pi^{\frac{D}{2}}}\frac{1}{((z-z')^2+i\epsilon)^{\frac{D-2}{2}}}
{~.}$$ It can be derived either by formal manipulations with the corresponding functional integral or by using the covariance properties of the defining differential equation.[^5]
Applying the formula first to pure AdS one gets in Poincaré coordinates $$G(x,x')=\frac{\Gamma(\frac{d-1}{2})}{R_1^{d-1}4\pi^{\frac{d+1}{2}}}
\Big(\frac{1}{x_\perp x'_\perp}\big[(x_\perp-x'_\perp)^2-(x^0-x'^0)^2+(\vec
x-\vec x')^2+i\epsilon\big]\Big)^{\frac{1-d}{2}}{~.}$$ Using the relation between Poincaré coordinates and the coordinates in the embedding space, see e.g. [@Aharony:1999ti], it is straightforward to verify that this with is equal to .\
The Poincaré patch of pure ${\text{AdS}}_{d+1}$ is conformal to a flat half space with $x_\perp\ge0$. $x_\perp=0$ corresponds to the conformal boundary of AdS. Let us first disregard that the flat half space represents only one half of ${\text{AdS}}_{d+1}$ and discuss global issues later. We can then implement either Dirichlet or Neumann boundary conditions by the standard mirror charge method. To $x=(x_\perp,x^0,x^1,\dots,x^{d-1})$ we relate the mirror point[^6] $$\label{mirror}
\tilde x=(-x_\perp,x^0,x^1,\dots,x^{d-1})$$ and the mirror propagator by $$\tilde G_\text{flat}(x,x')=G_\text{flat}(x,\tilde x'){~.}$$ Then $\frac{1}{2}(G_{\Delta_-}-G_{\Delta_+})$ in the second line of turns out to be just the Weyl transformed version of $\tilde G_\text{flat}(x,x')$. Equivalently we can state, that $G_{\Delta_+}$ and $G_{\Delta_-}$ are the Weyl transformed versions respectively of the Dirichlet and Neumann propagator in the flat halfspace.
The situation is different for ${{\text{AdS}}_{d+1}\times \text{S}^{d'+1}}$ spacetimes. According to , $x_{\perp}\ge0$ becomes a radial coordinate of a full $(d'+2)$-dimensional flat subspace of a total space with coordinates $$\label{zcoord}
z=\big(x_0,\vec x,x_\perp\tfrac{\vec Y}{R}\big){~,}$$ where $\vec Y^2=R^2$ are the embedding coordinates of $S^{d'+1}$. The boundary of the AdS part is mapped to the origin of the $(d'+2)$-dimensional subspace. Similarly to the pure AdS case, $G(z,z')$ from is the Weyl transform of $G_\text{flat}(z,z')$. To see this one has to cast the length square on the $(d'+2)$-dimensional subspace, which appears in the denominator of the propagator, into the form $$\frac{1}{R^2}(x_\perp\vec Y-x'_\perp\vec
Y')^2=x_\perp^2+x'^2_\perp-2\frac{x_\perp x'_\perp}{R^2}\vec Y\vec Y'
=(x_\perp-x'_\perp)^2+\frac{x_\perp x'_\perp}{R^2}v{~.}$$ In addition, with $$\label{twistapzcoord}
z_\text{s}=
\big(x_0,\vec x,-x_\perp\tfrac{\vec Y}{R}\big){~,}\qquad\tilde
G_\text{flat}(z,z')=G_\text{flat}(z,z'_\text{s})$$ we find that the second simple solution is the Weyl transformed version of $\tilde G_\text{flat}(z,z')$.
The coordinates and are related by replacing $\vec Y$ by $-\vec Y$, i.e. $z_\text{s}$ is related to $z$ by going to the antipodal point in the sphere, according to the definition of $z_\text{s}$ after . The two points $z$, $z_\text{s}$ are elements of $\mathds{R}^{d+d'+2}$ lying in the first Poincaré patch where $x_\perp\ge0$.
As we mentioned before, one has to be careful with global issues. We work in the Poincaré patch that only covers points with $x_\perp\ge0$. It is easy to see that the coordinates of $z$ and of $z_\text{s}$ remain unchanged if one simultaneously replaces $x_\perp$ by $-x_\perp$ and $\vec Y$ by $-\vec Y$. This operation switches from $z$ and $z_\text{s}$ respectively to the total antipodal positions $\tilde z$ and $\tilde z_\text{s}$, that are covered by a second Poincaré patch with $x_\perp<0$. Thus the latter points, being elements of the complete manifold, are not covered by the first Poincaré patch. In the context of pure ${\text{AdS}}_{d+1}$, the mirror point $\tilde x$ in related to $x$ is outside of the first Poincaré patch but it is still a point in ${\text{AdS}}_{d+1}$ covered by the second Poincaré patch. Hence $\tilde x$ is not an element of the flat half space that is conformal to the first Poincaré patch. We will now analyze the global issues more carefully by working with the corresponding ESU.
Relation to the ESU {#relESU}
===================
${\text{AdS}}_{d+1}$ and ${{\text{AdS}}_{d+1}\times \text{S}^{d'+1}}$ with $R_1=R_2$ are conformal to respectively one half and to the full ESU of the corresponding dimension. This can be easily seen in a certain set of global coordinates where the metric of ${\text{AdS}}_{d+1}$ assumes the form $$\label{AdSmetricglobalcoord}
{\operatorname{d}\!}s_{\text{AdS}}^2=R_1^2\sec^2\bar\rho\big(-{\operatorname{d}\!}t^2+{\operatorname{d}\!}\bar\rho^2+\sin^2\bar\rho{\operatorname{d}\!}\Omega_{d-1}^2\big){~,}$$ where $0\le\bar\rho<\frac{\pi}{2}$. The corresponding ESU has the topology $\mathds{R}\times\text{S}^d$ and its metric is given by the expression in parentheses. The conformal map between ${\text{AdS}}_{d+1}$ and $\text{ESU}_{d+1}$ has been used in [@Avis:1978yn] at $d=3$ to find consistent quantization schemes on ${\text{AdS}}_4$. In case of the Weyl invariant mass value the quantization prescription on the ESU leads to two different descriptions for pure AdS. One can either choose transparent boundary conditions or reflective boundary conditions at the image of the AdS boundary. The reflectivity of the boundary is guaranteed for either Dirichlet or Neumann boundary conditions. This is realized by choosing a subset of modes with definite symmetry properties, whereas in the transparent case all modes are used. Quantization in the reflective case leads one to the solutions $G_{\Delta_\pm}$. These results motivate why we will work on the ESU in the following. We will find the antipodal points and see how the mirror charge construction works. Then we will discuss what this implies for the well known propagators in ${\text{AdS}}_{d+1}$ and our solutions for ${{\text{AdS}}_{d+1}\times \text{S}^{d'+1}}$ in the Weyl invariant cases. In the coordinates a point $\tilde x$ antipodal to the point $x=(t,\bar\rho,x_\Omega)$ in ${\text{AdS}}_{d+1}$ is given by $$\label{AdSantipodglobalcoord}
\tilde x=(t+\pi,\bar\rho,\tilde x_\Omega){~,}$$ where $x_\Omega$ denotes the angles of the $(d-1)$-dimensional subsphere of ${\text{AdS}}_{d+1}$ with embedding coordinates $\omega_i$, which then obey $$\label{AdSsubsphereantipodalrel}
\omega_i(\tilde x_\Omega)=-\omega_i(x_\Omega){~.}$$ The above relation must not be confused with the relation between two points that are antipodal to each other on the sphere of the ESU at fixed time.
We now want to visualize the above relation on the sphere of the ESU. For convenience we choose ${\text{AdS}}_2$ such that the ESU has topology $\mathds{R}\times\text{S}^1$. The subsphere of ${\text{AdS}}_2$ is given by $\text{S}^0=\{-1,1\}$ such that we have $\omega=\pm1$. Hence, the transformation of $x_\Omega$ as prescribed in becomes a flip between the two points of the $\text{S}^0$. The information contained in $\text{S}^0$ can be traded for an additional sign information of $\bar\rho$, and therefore the transformation from $x_\Omega$ to $\tilde x_\Omega$ simply corresponds to an reflection at $\bar\rho=0$. We will now describe the time shift. After the transformation of the spatial coordinates is performed, one has found the antipodal event at time $t+\pi$. To relate it to an event at the original time $t$ one simply travels back in time along any null geodesics that crosses the spatial position of the antipodal event. On the ESU these null geodesics are clearly great circles. They meet at two points on the sphere. One is at the spatial position of the event and the other point is the antipodal point on the sphere of the ESU. The time it takes for a massless particle to travel between these two points is given by $\pi$, see Fig. \[fig:antipodalpntinESU\].
In this way one now arrives at an event that can have caused the event at later time $t+\pi$, and that has the same time coordinate as $x$, and its coordinate value $\bar\rho$ is given by a reflection at $\bar\rho=\frac{\pi}{2}$ on $\text{S}^1$. As $\bar\rho=\frac{\pi}{2}$ is the position of the AdS boundary, the mirror image to $x$ is situated outside of the region that corresponds to AdS. The effect of the original source at $x$ in combination with the mirror source either at $\tilde x$ as given in or at equal times mirrored at the boundary is that a light ray that travels to the boundary of AdS is reflected back into the interior.
Let us now discuss what happens in the case of ${{\text{AdS}}_{d+1}\times \text{S}^{d'+1}}$. The point $z=(t,\bar\rho,x_\Omega,y)$ possesses the total antipodal point $\tilde z$ and the two semi-antipodal points $z_\text{s}$ and $\tilde z_\text{s}$ given by $$\label{AdSSantipodglobalcoord}
\tilde z=(t+\pi,\bar\rho,\tilde x_\Omega,\tilde y){~,}\qquad
z_\text{s}=(t,\bar\rho,x_\Omega,\tilde y){~,}\qquad
\tilde z_\text{s}=(t+\pi,\bar\rho,\tilde x_\Omega,y){~,}$$ where $x_\Omega$ is as in the pure ${\text{AdS}}_{d+1}$ case and fulfills and $y$ are all angle coordinates of $S^{d'+1}$. In Fig. \[fig:antipodalpntinESU\] the case of ${\text{AdS}}_2\times\text{S}^0$, is shown. The effect of the factor $\text{S}^0$ can be alternatively described by adding to the range $0\le\bar\rho\le\frac{\pi}{2}$ the interval $\frac{\pi}{2}\le\bar\rho\le\pi$. This is possible because in the ESU at $\bar\rho=\frac{\pi}{2}$ the $\text{S}^0$ shrinks to a point. The complete ESU is now covered by the image of ${\text{AdS}}_2\times\text{S}^0$. The map to an antipodal position within the ${\text{AdS}}_2$ factor is as before, one finds the spatial coordinates by reflecting at $\bar\rho=0$. Within the $\text{S}^0$ factor, the antipodal position is found by reflecting at $\bar\rho=\frac{\pi}{2}$. Using this, it can be seen that w.r.t. the point $z$, the point $\tilde z$ is at the antipodal position on the $\text{S}^1$ of the ESU. Traveling back in time from $t+\pi$ to $t$ along a null geodesic, one arrives at $z$ from where one started. In the same way, the two semi-antipodal points $z_\text{s}$, $\tilde z_\text{s}$ are connected with each other by light rays. On the sphere of the ESU the $z$ and $z_\text{s}$ are related by a reflection at $\bar\rho=\frac{\pi}{2}$. Here, in contrast to the case of ${\text{AdS}}_2$, even the mirror events at equal times are situated within the image of ${\text{AdS}}_2\times\text{S}^0$. The above results are straightforwardly generalized to arbitrary dimensions.
Coming back to the discussion in Section \[weylana\], we can now make more precise statements about the mirror charge method to impose definite boundary conditions at $\bar\rho=\frac{\pi}{2}$. A linear combination of the two solutions like in does not necessarily generate additional $\delta$-sources on the r.h.s. of the differential equation , although both powerlike solutions in have singularities within ${\text{AdS}}_{d+1}$, the expression in the first line has one at $x=x'$ and the expression in the second line has one at $x=\tilde x'$. The singularity of the second expression only appears at $t=t'+\pi$, and its contribution to the r.h.s. of the differential equation depends on the time ordering prescription. In the cases where the $\theta$-function used for time ordering has an additional step at $t=t'+\pi$, a second $\delta$-function is generated (see [@Avis:1978yn] for a discussion of ${\text{AdS}}_4$). With the standard time ordering one finds that $G_{\Delta_\pm}$ are solutions with a source at $x=x'$ only. For ${\text{AdS}}_4$ this was obtained in [@Dullemond:1985bc].
The situation is different for ${{\text{AdS}}_{d+1}\times \text{S}^{d'+1}}$, where the propagator has singularities at $z=z'$, $z=\tilde z'$ and the second solution has singularities at $z=z'_\text{s}$, $z=\tilde z'_\text{s}$. Again, whether the singularities at $z=\tilde z'$ and $z=\tilde z'_\text{s}$ appear as $\delta$-sources on the r.h.s. of the differential equation , depends on the chosen time ordering. However in contrast to the pure ${\text{AdS}}_{d+1}$ case, the singularity of the second solution at $z=z'_\text{s}$ always leads to a $\delta$-source on the r.h.s. of but at the wrong position. This result corresponds to the above observation on the ESU that the mirror sources at equal times are not part of the image of ${\text{AdS}}_{d+1}$ but of ${{\text{AdS}}_{d+1}\times \text{S}^{d'+1}}$.
At the end let us give some comments on the $i\varepsilon(t,t')$-prescription. First of all, one has to introduce it in all expressions , and , since all of them have singularities at coincident or antipodal positions. Secondly, as worked out for ${\text{AdS}}_4$, a time independent $\varepsilon(t,t')=\epsilon$ refers to taking the step function $\theta(\sin(t-t'))$ for time ordering [@Avis:1978yn] which is appropriate if one restricts oneself to the hyperboloid. Standard time ordering with $\theta(t-t')$, being appropriate on the universal covering, yields a time dependent $\varepsilon(t,t')=\epsilon\operatorname{sgn}((t-t')\sin(t-t'))$ [@Dullemond:1985bc]. As mentioned in Section \[waveeqapp\], due to the time dependence of $\varepsilon(t,t')$, the coordinate dependence of the solutions is not entirely included in $u$ and $v$.
Mode summation on ${{\text{AdS}}_{d+1}\times \text{S}^{d'+1}}$ {#modesummation}
==============================================================
In this Section we will use the propagator on pure $\text{AdS}_{d+1}$ given by and the spherical harmonics on $\text{S}^{d'+1}$ to construct the propagator on ${{\text{AdS}}_{d+1}\times \text{S}^{d'+1}}$ via its mode expansion, summing up all the KK modes. We will be able to perform the sum only for special mass values where the conformal dimensions $\Delta_\pm$ of the scalar modes are linear functions of $l$, with $l$ denoting the $l$th mode in the KK tower. Even a mixing of several scalar modes of this kind is allowed. The mixing case is interesting because it occurs in supergravity theories on ${\text{AdS}}_{d+1}\times\text{S}^{d+1}$ backgrounds [@Corley:1999uz; @Michelson:1999kn; @Deger:1998nm; @Kim:1985ez; @Gunaydin:1985fk]. For example in type ${\text{II}\,\text{B}}$ supergravity in $\text{AdS}_5\times\text{S}^5$ the mass eigenstates of the mixing matrix for scalar modes [@Kim:1985ez; @Gunaydin:1985fk] correspond to the bosonic chiral primary and descendant operators in the ${\text{AdS}}/\text{CFT}$ dictionary [@Lee:1998bx]. For these modes $\Delta_\pm$ depend linearly on $l$.
The main motivation for investigating the mode summation was the hope to find the propagator for generic mass values. But forced to stay in a regime of a linear $\Delta_\pm$ versus $l$ relation we can give up the condition of conformal flatness, but remain restricted to special mass values. We nevertheless present this study since several interesting aspects are found along the way. Furthermore, in the literature it is believed that an explicit computation of the KK mode summation is too cumbersome [@Mathur:2002ry]. We will show how to deal with the mode summation by discussing the ${\text{AdS}}_3\times\text{S}^3$ case first, allowing for unequal radii but necessarily a special mass value. The result will then be compared to the expressions in the previous Sections by specializing to equal embedding radii.
Having discussed this special case we will comment on the modifications which are necessary to deal with generic ${{\text{AdS}}_{d+1}\times \text{S}^{d'+1}}$ spacetimes.
The results of the previous Sections in connection with the expression for the mode summation in the conformally flat and Weyl invariant coupled case lead to the formulation of a summation rule for a product of Legendre functions and Gegenbauer polynomials. An independent proof of this rule is given in Appendix \[userel\]. With this it is possible to discuss the results in generic dimensions without doing all the computations explicitly. Furthermore, the sum rule might be useful for other applications, too.
For the solution of we make the following ansatz[^7] $$\label{AdSSpropmodeexp}
G(z,z')=\frac{1}{R_2^{d'+1}}\sum_{I}G_I(x,x')Y^I(y)Y^{\ast I}(y'){~,}$$ where we sum over the multiindex $I=(l,m_1,\dots,m_{d'})$ such that $l\ge
m_1\ge\dots\ge m_{d'-1} \ge | m_{d'}|\ge 0$, $Y^I$ denote the spherical harmonics on $\text{S}^{d'+1}$, and ‘$\ast$’ means complex conjugation. Some useful relations for the spherical harmonics can be found in Appendix \[userel\].
The mode dependent Green function on ${\text{AdS}}_{d+1}$ then fulfills $$\Big(\Box_x
-M^2-\frac{l(l+d')}{R_2^2}\Big)G_I(x,x')=\frac{i}{\sqrt{-g_\text{AdS}}}\delta
(x,x'){~,}$$ which follows when decomposing the d’Alembert operator like in and using . The solution of this equation was already given in , into which the (now KK mode dependent) conformal dimensions enter. They were already defined in , and the AdS mass is a function of the mode label $l$ $$\label{mkksphere}
m^2=M^2+m_\text{KK}^2=M^2+\frac{l(l+d')}{R_2^2}{~.}$$
In the following as a simple example we will present the derivation of the propagator on ${\text{AdS}}_3\times\text{S}^3$ via the KK mode summation. Compared to the physically more interesting ${\text{AdS}}_5\times\text{S}^5$ background the expressions are easier and the general formalism becomes clear.
Evaluating for $d=d'=2$ the ${\text{AdS}}_3$ propagator for the $l$th KK mode is given by $$\label{AdS3propexpl}
G_\Delta(x,x')=\frac{1}{R_12^{\Delta+1}\pi}\xi^\Delta{F\big(\tfrac{\Delta}{2},\tfrac{\Delta}{2}+\tfrac{1}{2};\Delta;\xi^2\big)}=\frac{1}{R_14\pi}\frac{1+\sqrt{1-\xi^2}}{\sqrt{1-\xi^2}}\Big[\frac{\xi}{1+\sqrt{1-\xi^2}}\Big]^\Delta{~.}$$ From , and one finds that the mode dependent positive branch of the conformal dimension reads $$\label{AdS3S3conformaldim}
\Delta=\Delta_+=1+\frac{R_1}{R_2}\sqrt{\frac{R_2^2}{R_1^2}+l(l+2)+M^2R_2^2}{~.}$$
The spherical part follows from of Appendix \[userel\] where we discuss it in more detail and is given by $$\sum_{\scriptscriptstyle m_1\ge| m_2|\ge0 } ^l Y^I(y) Y^{\ast I}(y') =
\frac{(l+1)}{2 \pi^2}
C_l^{(1)}(\cos\Theta){~,}\qquad\cos\Theta=\frac{Y\cdot
Y'}{R_2^2}=1-\frac{v}{2R_2^2}{~.}$$ Remember that the $C_l^{(\beta)}$ denote the Gegenbauer polynomials and $Y$, $Y'$ in the formula for $\Theta$ are the embedding space coordinates of the sphere, compare with and . One thus obtains from $$G(z,z')=\frac{1}{8\pi^3R_1R_2^3}\frac{1+\sqrt{1-\xi^2}}{\sqrt{1-\xi^2}}\sum_{l=0}^\infty(l+1)\Big[\frac{\xi}{1+\sqrt{1-\xi^2}}\Big]^\Delta
C_l^{(1)}(\cos\Theta){~.}$$ In this formula $\Delta$ is a function of the mode parameter $l$ and we can explicitly perform the sum only for special conformal dimensions which are linear functions of $l$ $$\Delta=\Delta_+=\frac{R_1}{R_2}l+\frac{R_1+R_2}{R_2}{~,}$$ following from after choosing the special mass value $$\label{AdS3S3specmass}
M^2=\frac{1}{R_2^2}-\frac{1}{R_1^2}{~.}$$
The sum then simplifies and can explicitly be evaluated by a reformulation of the $l$-dependent prefactor as a derivative and by using $$\label{sumcalc}
\sum_{l=0}^\infty(l+1)q^l
C_l^{(1)}(\eta)=\Big(q{\frac{\partial}{\partial q}}+1\Big)\sum_{l=0}^\infty q^l
C_l^{(1)}(\eta)=\frac{1-q^2}{(1-2q\eta+q^2)^2}{~.}$$ With the replacements $$\label{qetadef}
q=\Big[\frac{\xi}{1+\sqrt{1-\xi^2}}\Big]^\frac{R_1}{R_2}{~,}\qquad
\eta=\cos\Theta$$ one now finds after some simplifications $$G(z,z')=\frac{1}{8\pi^3R_1R_2^3}\frac{1}{\sqrt{1-\xi^2}}\xi^{1+\frac{R_1}{R_2}}\frac{(1+\sqrt{1-\xi^2})^\frac{R_1}{R_2}-(1-\sqrt{1-\xi^2})^\frac{R_1}{R_2}}{\Big[(1+\sqrt{1-\xi^2})^\frac{R_1}{R_2}-2\xi^\frac{R_1}{R_2}\cos\Theta+(1-\sqrt{1-\xi^2})^\frac{R_1}{R_2}\Big]^2}{~.}$$ For the conformally flat case $R_1=R_2=R$, where becomes the mass generated by the Weyl invariant coupling to the background, the above expression simplifies to $$G(z,z')=\frac{1}{4\pi^3R^4}\frac{\xi^2}{(2-2\xi\cos\Theta)^2}=\frac{1}{4\pi^3}\frac{1}{(u+v+i\varepsilon(t,t'))^2}{~,}$$ where we have restored the $i\varepsilon(t,t')$-prescription. This result exactly matches .
The way to perform the KK mode summation on generic ${{\text{AdS}}_{d+1}\times \text{S}^{d'+1}}$ backgrounds is very similar to the one presented above. One finds a linear relation between $l$ and $\Delta$ $$\Delta_\pm=\pm\frac{R_1}{R_2}l+\frac{dR_2\pm d'R_1}{2R_2}$$ at the $(d+d'+2)$-dimensional mass value $$M^2=\frac{d'^2R_1^2-d^2R_2^2}{4R_1^2R_2^2}{~.}$$ This expression is a generalization of and it reduces to in the conformally flat case. For generic dimension the way of computing the propagator is very similar to the one presented for the ${\text{AdS}}_3\times\text{S}^3$ background. However the steps to express the hypergeometric function in the AdS propagator and to compute the sum become more tedious. For dealing with the hypergeometric functions see the remarks in Appendix \[userel\]. The sum generalizes in the way, that higher derivatives and more terms enter the expression .\
Next we discuss the mode summation in the conformally flat case $R_1=R_2$ at the Weyl invariant mass value but for generic $d$ and $d'$. In this case with the corresponding conformal dimensions $$\label{conformaldimconf}
\Delta=\Delta_+=l+\frac{d+d'}{2}{~,}$$ using and , the propagator is expressed as $$\label{modesumprop}
\begin{aligned}
G(z,z')&=\frac{\Gamma(\frac{d'}{2})}{4\pi}\Big(\frac{\xi}{2\pi
R^2}\Big)^{\frac{d+d'}{2}}\\
&\phantom{={}}
\times\sum_{l=0}^\infty\frac{\Gamma(l+\frac{d+d'}{2})}{\Gamma(l+\frac{d'}{2})}\Big(\frac{\xi}{2}\Big)^l{F\big(\tfrac{l}{2}+\tfrac{d+d'}{4},\tfrac{l}{2}+\tfrac{d+d'}{4}+\tfrac{1}{2};l+\tfrac{d'}{2}+1;\xi^2\big)}C_l^{(\frac{d'}{2})}(1-\tfrac{v}{2R^2}){~.}\end{aligned}$$ This equality together with the solution has lead us to formulate a sum rule for the above given functions at generic $d$ and $d'$. The above series should exactly reproduce . In Appendix \[userel\] we give an independent direct proof of the sum rule.
Considering the mode summation one finds an interpretation of the asymptotic behaviour of observed in Subsection \[AdSSderiv\]. The asymptotic regime $u\to\infty$ corresponds to $\xi\to0$. As the contribution of the $l$th mode is proportional to $\xi^{\Delta_+}\sim\xi^l$, the conformal dimension of the zero mode determines the asymptotic behaviour.
Note also that the additional singularity of at the total antipodal position $z=\tilde z'$ can be seen already in . Under antipodal reflection in ${\text{AdS}}_{d+1}$ the pure AdS propagator fulfills $G_{\Delta_\pm}(x,\tilde x')=(-1)^{\Delta_\pm}G_{\Delta_\pm}(x,x')$. On the sphere the spherical harmonics at antipodal points are related via $Y^I(y)=(-1)^lY^I(\tilde y)$. Hence, in case that $\Delta_\pm$ is given by , replacing $z'$ by the total antipodal point $\tilde z'$ leads to the same expression for the mode sum up to an $l$-independent phase factor.
One final remark to the choice of $\Delta_+$. What happens if one performs the mode expansion with AdS propagators based on $\Delta_-$? First in any case for high enough KK modes $\Delta_-$ violates the unitarity bound . But ignoring this condition from physics one can nevertheless study the mathematical issue of summing with $\Delta_-$. The corresponding series is given by after replacing $q$ by $q^{-1}$. It is divergent since for real $u$ the variable $q$ in obeys $|q|\le 1$ (case $R_1=R_2$). One can give meaning to the sum by the following procedure. $q$ as a function of $\xi$ has a cut between $\xi=\pm 1$. If $|q|\le 1$ on the upper side of the cut, then $|q|\ge 1$ on the lower side. Hence it is natural to define the sum with $\Delta_-$ as the analytic continuation from the lower side. By this procedure we found both for ${\text{AdS}}_3\times\text{S}^3$ and ${\text{AdS}}_5\times\text{S}^5$ up to an overall factor $-1$ the same result as using $\Delta_+$. The sign factor can be understood as a consequence of the continuation procedure.
The plane wave limit {#pwlimit}
====================
The plane wave background arises as a certain Penrose limit of ${\text{AdS}}_5\times\text{S}^5$. The scalar propagator in the plane wave has been constructed in [@Mathur:2002ry]. In this Section we study how this propagator in the massless case arises as a limit of our ${\text{AdS}}_5\times\text{S}^5$ propagator .
This approach is in the spirit of [@Dorn:2003ct], where one follows the limiting process instead of taking the limit before starting any computations. One finds a simple interpretation of certain functions of the coordinates introduced in [@Mathur:2002ry].
As an additional consistency check we take the $R\to\infty$ limit of the differential equation using to obtain the equation on the plane wave background and find that it is fulfilled by the massive propagator given in [@Mathur:2002ry].
Taking the aforementioned Penrose limit of ${\text{AdS}}_5\times\text{S}^5$ means to focus into the neighbourhood of a certain null geodesic which runs along an equator of the sphere with velocity of light. The metric of ${\text{AdS}}_5\times\text{S}^5$ in global coordinates[^8] $${\operatorname{d}\!}s^2=R^2\big(-{\operatorname{d}\!}t^2\cosh^2\rho+{\operatorname{d}\!}\rho^2+\sinh^2\rho
{\operatorname{d}\!}\Omega_3^2+{\operatorname{d}\!}\psi^2\cos^2\vartheta+{\operatorname{d}\!}\vartheta^2+\sin^2\vartheta
{\operatorname{d}\!}\hat\Omega_3^2\big)$$ via the replacements $$\label{vartraf}
t=z^++\frac{z^-}{R^2}{~,}\qquad\psi=z^+-\frac{z^-}{R^2}{~,}\qquad\rho=\frac{r}{R}{~,}\qquad\vartheta
=\frac{y}{R}$$ in the $R\to\infty$ limit turns into the plane wave metric.
The relation between global coordinates and the embedding space coordinates, see e. g. [@Aharony:1999ti], yields ($\omega_i$, $\hat\omega_i$, $i=1,\dots,4$ with $\vec\omega^2=\vec{\hat\omega}^2=1$ are the embedding coordinates of the two unit 3-spheres) $$\begin{aligned}
u&=2R^2\Big[-1+\cosh\rho\cosh\rho'\cos(t-t')-\sinh\rho\sinh\rho'\,\omega_i\omega'_i\Big]\\
v&=2R^2\Big[+1-\cos\vartheta\cos\vartheta'\cos(\psi-\psi')-\sin\vartheta\sin\vartheta'\,\hat\omega_i\hat\omega'_i\Big]{~.}\end{aligned}$$ Applying one gets at large $R$ up to terms vanishing for $R\to\infty$ $$\label{chordalpw}
\begin{aligned}
u&=2R^2\Big[-1+\cos\Delta z^++\frac{1}{R^2}\Big(-(\vec x^2+\vec
x{\hspace{0.5pt}}'^2)\sin^2\frac{\Delta z^+}{2}+\frac{(\vec x-\vec
x{\hspace{0.5pt}}')^2}{2}-\Delta z^-\sin\Delta z^+\Big)\Big]\\
v&=2R^2\Big[+1-\cos\Delta z^++\frac{1}{R^2}\Big(-(\vec y^2+\vec
y{\hspace{1pt}}'^2)\sin^2\frac{\Delta z^+}{2}+\frac{(\vec y-\vec
y{\hspace{1pt}}')^2}{2}-\Delta z^-\sin\Delta z^+\Big)\Big]{~,}\end{aligned}$$ where $\Delta z^\pm=z^\pm-z'^\pm$ and $\vec x=r\vec\omega$, $\vec
y=y\vec{\hat\omega}$. In the $R\to\infty$ limit the sum of both chordal distances is thus given by $$\label{totalchordalpw}
\Phi=\lim_{R\to\infty}(u+v)=-2(\vec z^2+\vec
z{\hspace{1pt}}'^2)\sin^2\frac{\Delta z^+}{2}+(\vec z-\vec
z{\hspace{1pt}}')^2-4\Delta z^-\sin\Delta z^+{~,}$$ where $\vec z=(\vec x,\vec y)$, $\vec z{\hspace{1pt}}'=(\vec x{\hspace{0.5pt}}',\vec y{\hspace{1pt}}')$ and $\Phi$ refers to the notation of [@Mathur:2002ry]. $\Phi$ is precisely the $R\to\infty$ limit of the total chordal distance on ${\text{AdS}}_5\times\text{S}^5$, which remains finite as both $\sim R^2$ terms in cancel. This happens due to the expansion around a *null* geodesic. In Appendix \[chordaldistpw\] it is shown that $\Phi$ is the chordal distance in the plane wave.
The massless propagator in the plane wave background in the $R\to\infty$ limit of with $d=d'=4$ thus becomes $$G_\text{pw}(z,z')=\frac{3}{2\pi^5}\frac{1}{(\Phi+i\varepsilon(z^+,z{\hspace{1pt}}'^+))^4}{~,}$$ which agrees with [@Mathur:2002ry].
In addition we checked the massive propagator of [@Mathur:2002ry] which fulfills the differential equation on the plane wave background. This equation can be obtained from and by taking the $R\to\infty$ limit. In the limit the sum of both chordal distances is given in . The difference is given by $$\label{diffchordalpw}
\lim_{R\to\infty}\frac{u-v}{R^2}=4(\cos\Delta z^+-1)$$ this has to be substituted into . Finally, one obtains the differential equation $$\Big[4\cos\Delta
z^+\Big(5{\frac{\partial}{\partial \Phi}}+\Phi{\frac{\partial^2}{\partial {\Phi}^2}}\Big)+4\sin\Delta
z^+{\frac{\partial}{\partial \Phi}}{\frac{\partial}{\partial \Delta
z^+}}-M^2\Big]G_\text{pw}(z,z')=\frac{i}{\sqrt{-g_\text{pw}}}\delta
(z,z'){~,}$$ which is fulfilled by the expression given in [@Mathur:2002ry]. As already noticed in Section \[waveeqapp\], in contrast to the massless propagator the massive one depends not only on the total chordal distance $\Phi$ but in addition on .
Conclusions {#concl}
===========
In this paper we have focussed on the propagator of scalar fields on ${{\text{AdS}}_{d+1}\times \text{S}^{d'+1}}$ backgrounds. We have argued that for an investigation of holography in the plane wave, in a first step one should study this propagator instead of the bulk-to-boundary one, since only for the former the Penrose limit is well defined.
We have first discussed solutions of the defining propagator equation at points away from possible singularities. On conformally flat backgrounds and for Weyl invariant coupled fields, both in ${{\text{AdS}}_{d+1}\times \text{S}^{d'+1}}$ and in pure ${\text{AdS}}_{d+1}$, exist two solutions, which are powerlike in the chordal distances. For ${{\text{AdS}}_{d+1}\times \text{S}^{d'+1}}$ these two solutions are powers either in the sum or difference of the chordal distances with respect to the ${\text{AdS}}_{d+1}$ and $\text{S}^{d'+1}$ factor. The first solution has a singularity if both points coincide or if they are at antipodal positions to each other. The second solution has singularities where both points are semi-antipodal to each other.
Whether, acting with the d’Alembert operator, $\delta$-sources are generated at the locations of these singularities, depends on the time ordering prescription. For ${\text{AdS}}_{d+1}$, being the universal cover of the embedded hyperboloid, standard time ordering is appropriate. Then for both solutions source terms arise only at coinciding times. This implies that the first solution develops just the right source to solve the full propagator equation. But the second solution necessarily has a source term away from the coincidence of the two points. Therefore it cannot be used to form different propagators via linear combinations with the first solution.
This is in contrast to the pure ${\text{AdS}}_{d+1}$ case. There the second solution has a singularity at the position where both points are antipodal to each other on the hyperboloid, i.e. there is no singularity at coinciding time coordinates. Hence, this singularity does not lead to a $\delta$-source. Thus, taking the sum and difference of the two solutions, propagators, which obey respectively Neumann and Dirichlet boundary conditions, can be constructed.
In addition, for ${{\text{AdS}}_{d+1}\times \text{S}^{d'+1}}$ we have investigated the KK decomposition of the propagator using spherical harmonics. We have noted that the summation can be performed even in non conformally flat backgrounds, but only for special mass values. The relevant condition is that the conformal dimension of the field mode is a linear function of the KK mode parameter.
In the conformally flat case for a Weyl invariant coupled field the uniqueness of the solution of the differential equation in combination with the KK decomposition led to the formulation of a theorem that sums up a product of Legendre functions and Gegenbauer polynomials. It was independently proven in Appendix \[userel\].
For ${\text{AdS}}_5\times\text{S}^5$ we explicitly performed the Penrose limit on our expression for the propagator to find the result on the plane wave background. We found agreement with the literature [@Mathur:2002ry] and got an interpretation for the spacetime dependence of the result. It simply depends on the $R\to\infty$ limit of the sum of both chordal distances on ${\text{AdS}}_5\times\text{S}^5$, which was shown to be the chordal distance in the plane wave. In the general massive case there is an additional dependence on the suitable rescaled difference of both chordal distances. We formulated the differential equation in the limit and checked that the massive propagator on the plane wave background given in [@Mathur:2002ry] is a solution.
Clearly future work is necessary to construct the propagator for the case of generic mass values. But already with our results one should be able to address the issue of defining a bulk-to-boundary propagator in the plane wave limit.\
[**Acknowledgements**]{}\
The work was supported by DFG (German Science Foundation) with the “Schwerpunktprogramm Stringtheorie" and the “Graduiertenkolleg 271". We thank James Babington, Niklas Beisert, Danilo Diaz, Johanna Erdmenger, George Jorjadze and Nicolaos Prezas for useful discussions.
Relation of the bulk-to-bulk and the bulk-to-boundary propagator {#bulkboundproprel}
================================================================
We will show for a scalar field in an Euclidean space how the bulk-to-boundary propagator is related to the bulk-to-bulk propagator, if the boundary has codimension one w.r.t. the bulk, like in the case of the ${\text{AdS}}/\text{CFT}$ correspondence. The bulk-to-bulk propagator $G(x,x')$ of a scalar field with mass $m$ is defined as Green function that fulfills[^9] $$\label{Gfuncdiffeq}
(\Box_x -m^2)G(x,x')=-\frac{1}{\sqrt{g}}\delta (x,x'){~,}$$ with appropriate boundary conditions. Here $\Box_x$ is the Laplace operator on the $(d+1)$-dimensional Riemannian manifold $M$ with a $d$-dimensional boundary which we denote with $\partial M$. The coordinates are $x^i$, the metric is $g_{ij}$ and its determinant is $g$. The propagator $G(x,x')$ corresponds to a scalar field $\phi(x)$ which should obey $$\label{phidiffeq}
(\Box_x -m^2)\phi(x)=J(x){~,}\qquad\lim_{x_\perp\to 0}\phi(x)x^a_\perp=\bar\phi(\bar x){~,}$$ where $J(x)$ are sources for the field $\phi$ in the interior. We have split the coordinates like $x=(x_\perp,\bar x)$ with the boundary at $x_\perp=0$. Boundary values $\bar\phi$ for the field $\phi$ are specified with a nontrivial scaling with $x^a_\perp$ for later convenience. The bulk-to-boundary propagator $K(x,\bar x')$ is defined as the solution of the equations $$\label{bulkboundpropdefeq}
(\Box_x -m^2)K(x,\bar x')=0{~,}\qquad\lim_{x_\perp\to 0}K(x,\bar
x')x^a_\perp=\delta(\bar x,\bar x'){~,}$$ where the second equation implements the necessary singular behaviour at the boundary. A solution of the equations with $J(x)=0$ is then given by $$\label{boundvalsol}
\phi(x)=\int_{\partial M}{\operatorname{d}\!}^d\bar x' K(x,\bar x')\bar\phi(\bar x'){~.}$$ Since we deal with the problem in Euclidean signature [@Witten:1998qj; @Freedman:1998tz; @D'Hoker:2002aw], we will denote $K(x,\bar x')$ as the Poisson kernel. It is not independent from the Green function defined via as we will now show.
With one can write an identity for the field $\phi$ that reads $$\phi(x)=-\int_M{\operatorname{d}\!}^{d+1}x'\sqrt{g}\phi(x')(\Box_{x'}-m^2)G(x,x'){~.}$$ After applying partial integration twice and using it assumes the form $$\label{boundsourcesol}
\phi(x)=\int_{\partial M}{\operatorname{d}\!}A'_\mu\sqrt{g}g^{\mu\nu}\big[(\partial'_\nu\phi(x'))G(x,x')-\phi(x')\partial'_\nu
G(x,x')\big]-\int_M{\operatorname{d}\!}^{d+1}x'\sqrt{g}J(x')G(x,x'){~,}$$ where ${\operatorname{d}\!}A'_\mu$ denotes the infinitesimal area element on $\partial M$ which points into the outer normal direction and $\partial'_\mu$ denotes a derivative w.r.t. $x'_\mu$. If one has the additional restriction that $G(x,x')=0$ for $x'\in\partial M$ ($x'_\perp=0$) the first term in the above boundary integral is zero. One then arrives at the ‘magic rule’ which for the boundary value problem in presence of a source $J(x)$ in the interior can be found in [@Barton:1989]. Here, however, we have to be more careful. In we have allowed for a scaling of the boundary value with $x^a_\perp$ as written down in . For $a>0$ the vanishing $G(x,x')$ at the boundary can be compensated and the first term in the boundary integral of then contributes.
Considering ${\text{AdS}}_{d+1}$, this is indeed the case, because the field $\phi$ with conformal dimension $\Delta$ represents the non-normalizable modes, which scale as given in with $a=\Delta-d$. Indicating the corresponding propagator with the suffix $\Delta$, one finds $G_\Delta(x,x')=0$ at $x'\in\partial M$ but the vanishing is compensated by the singular behaviour of the non-normalizable modes in the limit $x'_\perp\to 0$.
We now formulate on Euclidean ${\text{AdS}}_{d+1}$ in Poincaré coordinates with $J(x')=0$, where one has $${\operatorname{d}\!}A_\mu=-{\operatorname{d}\!}^d\bar
x\delta_\mu^\perp{~,}\qquad\sqrt{g}=\Big(\frac{R_1}{x_\perp}\Big)^{d+1}{~,}\qquad g^{\perp\perp}=\Big(\frac{x_\perp}{R_1}\Big)^2{~.}$$ The minus sign in the area element stems from the fact that the $x_\perp$-direction points into the interior of ${\text{AdS}}_{d+1}$, but one has to take the outer normal vector. Now reads $$\phi_\Delta(x)=-R_1^{d-1}\int{\operatorname{d}\!}^d\bar
x'x'^{1-d}_\perp\big[(\partial'_\perp\phi(x'))G_\Delta(x,x')-\phi(x')\partial'_\perp G_\Delta(x,x')\big]{~.}$$ Using now with $a=\Delta-d$, one finds that the relation of the bulk-to-bulk and the bulk-to-boundary propagator is given by $$\label{bulktoboundarypropagatorrel}
K_\Delta(x,\bar x')=-R_1^{d-1}
\big[(d-\Delta)x'^{-\Delta}_\perp-x'^{1-\Delta}_\perp\partial'_\perp
\big] G_\Delta(x,x')\big|_{x'_\perp=0}$$ If we now insert the explicit expression for $G_\Delta(x,x')$, we see what we already mentioned: in approaching the boundary ($x'_\perp\to 0$), $G_\Delta(x,x')$ itself goes to zero like $x'^\Delta_\perp$ but this is compensated by the singular behaviour of the prefactor in the first term of . Hence, in contrast to the situation of the ’magic rule’ [@Barton:1989], it contributes to the bulk-to-boundary propagator. One then finds with $$\label{xiinPcoord}
\xi=\frac{2x_\perp x'_\perp}{x_\perp^2+x'^2_\perp+(\bar x-\bar x')^2}{~.}$$ in Poincaré coordinates and with ${F\big(a,b;c;0\big)}=1$ that effectively $$K_\Delta(x,\bar x')=-R_1^{d-1}(d-2\Delta)x'^{-\Delta}_\perp G_\Delta(x,x')\big|_{x'_\perp=0}{~,}$$ which is in perfect agreement with the explicit expressions given in [@D'Hoker:2002aw].
Useful relations for hypergeometric functions and spherical harmonics {#userel}
=====================================================================
Most of the relations we present here can be found in [@Bateman:1953; @Abramowitz:1972; @Gradshteyn:1980] or are derived from there. The hypergeometric functions in the propagators with $\Delta_\pm=\frac{d\pm1}{2}$ (at the mass value generated by the Weyl invariant coupling) become ordinary analytic expressions $$\label{hypergeom1}
\begin{aligned}
{F\big(a,a+\tfrac{1}{2};\tfrac{1}{2};\xi^2\big)}&=\frac{1}{2}\big[(1+\xi)^{-2a}+(1-\xi)^{2a}\big]{~,}\\
{F\big(a+\tfrac{1}{2},a+1;\tfrac{3}{2};\xi^2\big)}&=-\frac{1}{4a\xi}\big[(1+\xi)^{-2a}-(1-\xi)^{-2a}\big]{~.}\end{aligned}$$ Setting $a=\frac{d-1}{4}$ one finds .
To find the hypergeometric functions relevant for the propagators in higher dimensional AdS spaces one can use a recurrence relation (Gauß’ relation for contiguous functions) $$\label{recurrencerel}
{F\big(a,b;c-1;z\big)}=\frac{c[c-1-(2c-a-b-1)z]}{c(c-1)(1-z)}{F\big(a,b;c;z\big)}+\frac{(c-a)(c-b)z}{c(c-1)(1-z)}{F\big(a,b;c+1;z\big)}{~.}$$ where the hypergeometric functions relevant in lower dimensional AdS spaces enter.
For odd AdS dimensions (even $d$) the relevant hypergeometric functions can be expressed with the above recurrence relation in terms of ordinary analytic functions. This happens because of the explicit expressions $$\begin{aligned}\label{hypergeomstart}
{F\big(a,a+\tfrac{1}{2};2a;z\big)}&=\frac{2^{2a-1}}{\sqrt{1-z}}\big[1+\sqrt{1-z}\big]^{1-2a}{~,}\\
{F\big(a,a+\tfrac{1}{2};2a+1;z\big)}&=2^{2a}\big[1+\sqrt{1-z}\big]^{-2a}{~.}\end{aligned}$$ One has to apply $n$ times to compute the AdS propagator at generic $\Delta$ in $d+1=3+2n$ dimensions. For ${\text{AdS}}_3$ one simply uses the first expression in . The ${\text{AdS}}_5$ case is of particular importance and therefore we give the explicit expression for the needed hypergeometric function $${F\big(\tfrac{\Delta}{2},\tfrac{\Delta}{2}+\tfrac{1}{2};\Delta-1;z\big)}=\frac{1}{2(1-z)^\frac{3}{2}}\Big[\frac{2}{1+\sqrt{1-z}}\Big]^{\Delta-1}\Big[\sqrt{1-z}+\frac{\Delta-1}{\Delta-2}(1-z)+\frac{1}{\Delta-1}\Big]{~.}$$
Spherical harmonics $Y^I(y)$ on $\text{S}^{d'+1}$ are characterized by quantum numbers $$\label{multindex}
I = (l, m_1, \dots, m_{d'}){~,}\qquad l \ge m_1 \ge \dots \ge m_{d'-1} \ge |
m_{d'}| \ge 0$$ and form irreducible representations of $\text{SO}(d'+2)$. They are eigenfunctions with respect to the Laplace operator on the sphere $$\label{SHcasimir}
\Box_y Y^I(y)=-\frac{l(l+d')}{R_2^2} Y^I(y)$$ and satisfy the relation $$\label{SHcomplete}
\sum_{\scriptscriptstyle m_1\ge\dots\ge
m_{d'-1}\ge|m_{d'}|\ge0}^lY^I(y)Y^{\ast I}(y') = \frac{(2l+d')\Gamma
(\frac{d'}{2})}{4 \pi^{\frac{d'}{2}+1}}
C_l^{(\frac{d'}{2})}(\cos\Theta){~,}\qquad\cos\Theta=\frac{\vec Y\cdot\vec
Y'}{R_2^2}=1-\frac{v}{2R_2^2}{~.}$$ The $C_l^{(\frac{d'}{2})}$ are the Gegenbauer Polynomials which can be defined via their generating function $$\label{genfunc}
\frac{1}{(1-2q\eta+q^2)^\beta}=\sum_{l=0}^\infty q^lC_l^{(\beta)}(\eta){~.}$$ Using and for $\Delta_+=l+\frac{d+d'}{2}$ (leading to ), one finds the solution if the following relation holds for $\alpha\ge\beta>0$, $2\alpha,2\beta\in\mathds{N}$ $$\label{hyperid}
\sum_{l=0}^{\infty}\frac{\Gamma(l+\alpha)}{\Gamma(l+\beta)}\Big(\frac{\xi}{2}\Big)^l{F\big(\tfrac{l}{2}+\tfrac{\alpha}{2},\tfrac{l}{2}+\tfrac{\alpha}{2}+\tfrac{1}{2};l+\beta+1;\xi^2\big)}C_l^{(\beta)}(\eta)
=\frac{\Gamma(\alpha)}{\Gamma(\beta)}\frac{1}{(1-\xi\eta)^\alpha}{~,}$$ with the interpretation $\alpha=\frac{d+d'}{2}$, $\beta=\frac{d'}{2}$. We could not find the above formula in the literature. It is in fact a summation rule for a product of a special hypergeometric function for which so called quadratic transformation formulae exist and which can be expressed in terms of a Legendre function [@Abramowitz:1972] and of a Gegenbauer polynomial. The identity can therefore be re-expressed in the following way $$\label{polyid}
\Big(\frac{2}{\xi}\Big)^\beta(1-\xi^2)^\frac{\beta-\alpha}{2}\sum_{l=0}^{\infty}\Gamma(l+\alpha)(l+\beta)P_{\alpha-\beta-1}^{-l-\beta}\big(\tfrac{1}{\sqrt{1-\xi^2}}\big)C_l^{(\beta)}(\eta)=\frac{\Gamma(\alpha)}{\Gamma(\beta)}\frac{1}{(1-\xi\eta)^\alpha}{~.}$$ The simplest way to prove[^10] is to use the orthogonality of the Gegenbauer polynomials $$\int_{-1}^{1}d\eta~
(1-\eta^2)^{\beta-\frac{1}{2}}C_m^{(\beta)}(\eta)C_n^{(\beta)}(\eta)=\frac{2^{1-2\beta}\pi\Gamma(n+2\beta)}{\Gamma(n+1)(n+\beta)\Gamma(\beta)^2}\delta_{mn}$$ to project out a term with fixed $l$ from the sum in . The Gegenbauer polynomials in the integral on the r.h.s. of should then be expressed via Rodrigues’ formula $$C_n^{(\beta)}(\eta)=(-1)^n2^{-n}\frac{\Gamma(\beta+\frac{1}{2})\Gamma(n+2\beta)}{\Gamma(n+1)\Gamma(2\beta)\Gamma(n+\beta+1)}(1-\eta^2)^{\frac{1}{2}-\beta}\frac{d^n}{d\eta^n}(1-\eta^2)^{n+\beta-\frac{1}{2}}{~.}$$ Repeated partial integration to shift all the above derivatives to the first factor under the integral and a suitable variable transformation at the end leads to an integral that can be expressed via a hypergeometric function. This hypergeometric function is than connected to the one on the left hand side of via the quadratic transformation formula $${F\big(a,a+\tfrac{1}{2};c;z^2\big)}=(1+z)^{-2a}{F\big(2a,c-\tfrac{1}{2};2c-1;\tfrac{2z}{1+z}\big)}{~.}$$ Both sides of than match and the proof is complete. The relation is therefore valid not only for $\alpha\ge\beta>0$, $2\alpha,2\beta\in\mathds{N}$ but for all $\beta>0$.
The chordal distance in the plane wave {#chordaldistpw}
======================================
It has been shown by Penrose [@Penrose:1965] that it is impossible to globally embed the plane wave spacetimes into a pseudo-Euclidean spacetime. However, an isometric embedding of the $D$-dimensional CW spaces with metric $$\label{CWmetric}
{\operatorname{d}\!}s^2=-4{\operatorname{d}\!}z^+{\operatorname{d}\!}z^-+H_{ij}z^iz^j({\operatorname{d}\!}z^+)^2+{\operatorname{d}\!}\vec
z^2{~.}$$ in $\mathds{R}^{2,D}$ is possible [@Blau:2002mw]. The flat metric of $\mathds{R}^{2,D}$ via the coordinate transformations $$Z_+^1=\frac{1}{2}(Z_0+Z_d){~,}\quad Z_-^1=\frac{1}{2}(Z_0-Z_d){~,}\quad Z_+^2=\frac{1}{2}(Z_{d+1}+Z_{d-1}){~,}\quad Z_-^2=\frac{1}{2}(Z_{d+1}-Z_{d-1})$$ can be transformed to $${\operatorname{d}\!}s^2=-4\sum_{k=1}^2{\operatorname{d}\!}Z_+^k{\operatorname{d}\!}Z_-^k+\sum_{i=1}^D{\operatorname{d}\!}Z_i^2{~.}$$ If the hypersurface is defined as $$\sum_{k=1}^2Z_+^kZ_+^k=1{~,}\qquad H_{ij}Z^iZ^j+4\sum_{k=1}^2Z_+^kZ_-^k=0$$ and parameterized as follows $$\begin{aligned}
Z_+^1&=\cos z^+{~,}\\
Z_-^1&=-z^-\sin z^+-\frac{1}{4}H_{ij}z^iz^j\cos z^+{~,}\\
Z_+^2&=\sin z^+{~,}\\
Z_-^2&=z^-\cos z^+-\frac{1}{4}H_{ij}z^iz^j\sin z^+{~,}\\
Z_i&=z_i{~,}\end{aligned}$$ one finds that the induced metric is given by . The chordal distance in the plane wave reads $$\label{pwchordaldist}
\begin{aligned}
\Phi(z,z')&=-4\sum_{k=1}^2(Z_+^k(z)-Z_+^k(z'))(Z_-^k(z)-Z_-^k(z'))+\sum_{i=1}^D(Z_i(z)-Z_i(z'))^2\\
&=-4(z^--z'^-)\sin(z^+-z'^+)+2H_{ij}(z^iz^j+z'^iz'^j)\sin^2\frac{z^+-z'^+}{2}+(\vec
z-\vec z')^2{~.}\end{aligned}$$ In our case where $H_{ij}=-\delta_{ij}$, this result matches .
[^1]: [email protected], [email protected], [email protected]
[^2]: This normalization is consistent with the definition in [@D'Hoker:2002aw], because in a continuation to Euclidean space the factor $i$ in front of the $\delta$-source becomes $-1$ as in .
[^3]: More precisely, ${\text{AdS}}_{d+1}$ is the universal covering of the hyperboloid in $\mathds{R}^{2,d}$.
[^4]: Pure AdS spaces are conformally flat.
[^5]: Of course, the discussion has to be completed by considering also the boundary conditions.
[^6]: Using $x_\perp<0$ for parameterizing the second Poincaré patch the mirror point is at the antipodal position on the hyperboloid.
[^7]: This ansatz is designed to generate a solution that corresponds to . If one wants to generate a solution corresponding to one has to replace either $y$ or $y'$ by the corresponding antipodal coordinates $\tilde y$ or $\tilde y'$.
[^8]: The ${\text{AdS}}_5$ coordinates are related to the ones in via $\cosh\rho=\sec\bar\rho$.
[^9]: This equation is the analytic continuation of .
[^10]: We thank Danilo Diaz for delivering a simplification of our proof in the previous version, that allows for an extension to more general values of $\alpha$ and $\beta$.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'We consider the problem of scheduling and power allocation for the downlink of a 5G cellular system operating in the millimeter wave (mmWave) band and serving two sets of users: users typically seen in device-to-device (D2D) communications, and users, or high data rate services. The scheduling objective is the weighted sum-rate of both and users, and the constraints ensure that active users get the required rate. The weights of the objective function provide a trade-off between the number of served users and the resources allocated to users. For mmWave channels the virtual channel matrix obtained by applying fixed discrete-Fourier transform (DFT) beamformers at both the transmitter and the receiver is sparse. This results into a sparsity of the resulting multiple access channel, which is exploited to simplify scheduling, first establishing an interference graph among users and then grouping users according to their orthogonality. The original scheduling problem is solved using a graph-coloring algorithm on the interference graph in order to select sub-sets of orthogonal users. Two options are considered for users: either they are chosen orthogonal to users or non-orthogonal. A waterfilling algorithm is then used to allocate power to the users.'
author:
- '[^1]'
title: 'Beamforming and Scheduling for mmWave Downlink Sparse Virtual Channels With Non-Orthogonal and Orthogonal Multiple Access'
---
5G; Beamforming; mmWave; Non-orthogonal Multiple Access; Scheduling.
Introduction
============
MmWave transmission systems operating with multiple users will pose many design challenges, related to the need of precise beamforming to overcome the strong attenuation at GHz frequencies [@Erkip-14]. On the other hand, the use of mmWave has been advocated for fifth-generation (5G) systems in order to satisfy on ten-fold achievable data rate with respect to the previous mobile communication generation: in this scenario the use of extremely high-frequency bands becomes even more problematic due to the presence of multiple users and the need to design suitable multiple access schemes. From a paradigm based on orthogonality among users to maximize the sum rate – mainly pursued in 4G systems – the 5G trend is now shifting back to non-orthogonal multiple access (NOMA) schemes, better suited to heterogeneous devices envisioned in the 5G network, e.g., sensors and actuators of the Internet of Things (IoT) [@Gupta-15]. All these problems are further complicated by the use of a huge number of antennas, the massive regime, again proposed for 5G systems in order to propel data rate and cell user density. In this context, a radical review of both beam-forming and scheduling is needed, taking into account also the need of keeping both the signaling exchange on air at the minimum (for a higher data-rate) and the signal processing complexity under control, especially on the terminal side.
Focusing on a downlink scenario where a base station (BS) equipped with many antennas aims at serving multiple mobile terminals (MTs) on the mmWave band, many solutions are available in the literature for beam-forming and scheduling. For example, orthogonal user transmission can be considered in order to maximize the rate, thus (regularized) zero-forcing beamforming can be used at the BS, together with water-filling power allocation and scheduling [@Spencer-04; @Yang-13]. Similar approaches include diagonalization and block-diagonalization of the channel [@Wong-03]. Indeed, it has been shown that in a massive MIMO regime, eigenbeamforming and matched filter precoders are optimal, and the transmit power can be made arbitrarily small [@Hoydis-13]. However, these results leverage orthogonality among users and have as target the maximization of the sum rate. These assumptions must be revised in 5G systems used also for device-to-device (D2D) communication in the IoT context, typically having requirements: for these services the target becomes serving the largest number of users, providing them the required fixed data rate. In the NOMA context, two approaches have been proposed in the literature: superposition coding and successive interference cancellation. With the first approach users suffer from the interference but can still decode the intended signal as long as the signal to noise plus interference ratio (SNIR) is above a given threshold; with the latter approach the interfering signal is decoded and then canceled before the useful signal is decoded. For superposition coding a large literature is available, for the maximization of the weighted sum-rate and energy-efficient solutions (see [@Fang-16; @Zte] for surveys). In [@Ding-16] beamforming and power allocation have been designed using as metrics the sum rate and outage probability of MIMO-NOMA systems. In [@Shi-08] optimal beamforming and power allocation solutions have been derived for the weighted sum-rate maximization under a total power constraint and a max-min balancing of user rate under a total power constraint: in this work however no user is present. A simple approach has been proposed in [@Nassar1; @Nassar2], where and users are allocated orthogonal resources, then using further orthogonal resources for the users and non-orthogonal resources for the users. In [@Naspatent] an scheduling considering also the pilot overhead needed for channel estimation, adapted to the user channels characteristics has been considered.
In this report we focus on the problem of scheduling and power allocation for the mixed user characteristics of [@Nassar1]: first, we formulate the problem of maximizing the weighted sum-rate of both and users, with constraints ensuring that active users get the required rate. This turns out to be a mixed integer programming problem, where the integer variables dictate which users are active, and continuous variables provide the power allocation for all users. The weights of the objective function provide a trade-off between the number of served users and the resources allocated to variable users. We then exploit the peculiarity of mmWave channels which can be suitably represented as a sparse matrix in a dual domain [@Schniter]. The sparse channel is obtained by applying fixed discrete Fourier transform (DFT) beamformers at both the transmitter and the receiver, thus providing a low-complexity implementation. On the other hand, the sparsity of the resulting multiple access channel is exploited to simplify scheduling, first establishing an interference graph among users and then grouping users according to their orthogonality. In particular, the original scheduling problem is solved using a graph-coloring algorithm on the interference graph in order to select sub-sets of orthogonal users. Two options are considered for users: either they are chosen orthogonal to users or they are allowed to interfere with the other set. A waterfilling algorithm is then used to allocate power to the users, while the power needed to achieve the required rate is allocated to users. The rest of the report is organized as follows. Firstly we derive the system model for the described scenario and give a mathematical description of the power allocation and scheduling problem. Secondly we propose three solutions: the direct power allocation and scheduling, a solution based on , where both and users are clustered in different sets in order to eliminate interference, and a solution where only users are clustered to eliminate interference. Lastly we analyse the performance of the proposed solutions in terms of achievable sum-rate when the optimization target moves from users to users and show how the number of antennas at the BS impacts on the proposed solutions.
System Model
============
We consider a massive downlink transmission from a BS with a large number $N_{a}$ of antennas organized in a uniform planar array (UPA) to a set of $N_u$ single-antenna users, with $N_u << N_a$. Assuming an flat-fading channel model among each antenna couple, the resulting $N_u \times N_a$ channel matrix is denoted as $\bm{H}$. For a transmission in the mmWave band, each BS-user link is described by $L$ paths, with $L$ typically being small. Moreover, considering the UPA configuration at the BS and the single antenna MT, as derived in [@Sayeed-02] the row channel vector $\bm{h}_k$ of user $k$ can be transformed into a virtual channel vector $\bm{h}_{{\rm v},k}$ by the transformation $$\bm{h}_k = \bm{h}_{{\rm v},k} \bm{U}_t^*\,,$$ where $\bm{U}_t$ is the $N_a \times N_a$ unitary DFT matrix. By stacking the rows $\bm{h}_{{\rm v}, k}$ into the matrix $\bm{H}_v$ we can write the channel matrix as $$\bm{H}= \bm{H}_{\rm v}\bm{U}_t^*\,.
\label{eq:hv}$$ When the number of antennas at the BS tends to infinity, the virtual channel matrix $\bm{H}_v$ becomes sparse, asymptotically having $L \cdot N_u$ non-zero entries associated to the $L$ paths of each one of the $N_u$ users. Also, when angles of departures are aligned with DFT direction [@Schniter] again $\bm{H}_v$ becomes sparse with $L \cdot N_u$ non-zero complex entries, with probability density function $$p_{[\bm{H}_v]_{j,k}}(a) = (1-\eta)\delta(a)+\frac{\eta}{\pi \sigma^2_L}\exp\left(-\frac{|a|^2}{\sigma_L^2}\right)\,,
\label{eq:chent}$$ where $\eta=L/ N_a$ is the probability that a virtual channel matrix entry is different from zero, $\delta(\cdot)$ is the delta function and $\sigma_L^2$ is the average channel gain.
Received signal is affected by complex , independent at each user, with zero mean and variance $\sigma^2$.
We consider two different types of users:
- $K$ users having as target the maximization of their ;
- $N$ users aiming at being active, i.e., transmitting with a fixed data rate $R_{\rm fix}$, corresponding to fixed $\gamma_{\rm fix}$ with $$R_{\rm fix} = \log_2 (1+\gamma_{\rm fix})\,.$$
The total number of users is $K+N = N_u$. A user assumes only two possible states: transmission with a imposed by the requirement or turn off. In particular, the first $K$ users are the users, while next $N$ users are the users.
We denote by $p_j$, $j=1, \ldots, N_u$ the power assigned to user $j$. We also collect into the $N_u$ column vector $\bm{p}$ all user powers. We consider a total power constraint, i.e., $$\label{powerconst}
p_k \geq 0, \quad \sum_{k=1}^{N_u}p_k \leq 1\,.$$
According to the beamformer adopted by the BS, as described in the following section, the resulting for user $k=1, \ldots, N_u$ is denoted as $\gamma_k(\bm{p})$, depending on the power allocation. For the beamforming and scheduling problem we aim at maximizing the weighted $$\mathcal{R}(\bm{p}) = \sum_{k=1}^{K} w_{k} \log_{2}(1+\gamma_{k}(\bm{p}))+ \sum_{k=K+1}^{K+N}\rho_k x_k R_{\rm fix}\,,
\label{eq:sr}$$ where $\rho$ is the weight of users , $w_k$ is the weight of the of user $k$ and $x_k$ is a binary variable which assumes value 1 if the considered user is active and 0 otherwise.
Beamforming and Scheduling
==========================
Considering the properties of the virtual channel, and that it can be obtained by the inverse matrix applied at the BS, we consider as BS transmit beamformer matrix $\bm{U}_t$, therefore the equivalent channel seen by the users is the sparse virtual channel $\bm{H}_v$ having $N_a$ virtual transmit antennas and $N_u$ virtual receive antennas. We will use one transmit virtual antenna per user, and in particular for user $k$ we select in $\bm{h}_{{\bm v},k}$ the entry with largest gain, i.e., $$\ell_k = {\rm argmax}_i |[\bm{h}_{{\bm v},k}]_i|^2\,,$$ and we indicate with $$G_{k,k} = |[\bm{h}_{{\bm v},k}]_{\ell_k}|^2$$ the corresponding gain. With this choice, user $k$ will suffer from the the interference of all other user signals, based on his virtual channel $\bm{h}_{{\bm v},k}$, i.e. the gain of the interference channel from user $j$ to user $k$ is $$G_{j,k} = |[\bm{h}_{{\bm v},k}]_{\ell_j}|^2\,,$$ corresponding to the use of virtual antenna $\ell_k$ by user $j$ through the channel of user $k$. Therefore $N_u \times N_u$ matrix $\bm{G}$ having entries $G_{j,k}$ defines the useful gain and interference gain for all users. With this choice, user $k$ experiences a $$\gamma_k(\bm{p}) = \frac{G_{k,k}p_k}{\sum_{j \neq k}G_{j,k}p_j+\sigma^2}\,.
\label{eq:sinr}$$
Let us consider the weighted ASR maximization problem with variables $p_k$ and $x_k$. Our aim is to select a subset of users and to allocate power to both users and active users in order to maximize the weighted ASR. Let $\bm{x} = [x_{1}, \ldots, x_{K+N}]$. The problem can be modeled as
$$\begin{split}
\label{eq:max}
\underset{\bm{p}, \bm{x}}{\text{max}}
\sum_{k=1}^{K} w_k \log_{2}\bigg (1+\frac{G_{k,k} p_{k}}{\sum_{j=1, j \neq k}^{K+N}G_{j,k} p_{j} x_j +\sigma^2}\bigg )+ \\
+\sum_{k=K+1}^{K+N}\rho_k R_{\rm fix}x_k
\end{split}$$
and $$\label{const1}
\frac{G_{k,k} p_{k}}{\sum_{j=1, j \neq k}^{K+N}G_{j,k} p_{j} x_j +\sigma^2} \geq \gamma_{\rm fix} \quad k = K+1,...,K+N$$ $$\quad x_k = 1 \quad k = 1,...,K$$ $$\label{const2}
\quad x_k \in \{0,1\} \quad k = K+1,...,K+N$$
This problem of joint power allocation and scheduling is modelled as a mixed-integer optimization problem, which belongs to the NP-hard class of problems. We will hence first analyse the solution and then discuss two different methods that can reduce the computational complexity of the solution while ensuring the same weighted ASR value.
Direct solution
---------------
In order to solve the mixed-integer programming problem we divide it into two problems: the problem of finding the set of active users (thus solving on $\bm{x}$) and the problem of allocating the power (thus solving on $\bm{p}$). In particular we exhaustively consider all possible sets of active users, and for each set we solve the power allocation problem (if feasible) maximizing the weighted ASR. For each set of active users $\bm{x}$ the resulting power allocation problem is hence $$\begin{split}
\label{7a1}
\underset{\bm{p} }{\text{max}}
\sum_{k=1}^{K} w_k \log_{2}\bigg (1+\frac{G_{k,k} p_{k}}{\sum_{j=1, j \neq k}^{K+N}G_{j,k} p_{j} x_j +\sigma^2}\bigg )+ \\
+\sum_{k=K+1}^{K+N}\rho_k R_{\rm fix}x_k
\end{split}$$ s.t. (\[powerconst\]), (\[const1\]) and (\[const2\]).
We notice that the resulting optimization problem is non-convex due to the argument of the the log function of (\[7a1\]) and we approximate the target function to make it convex. In particular, we consider a high regime and perform the change of variables $q_k = \log_2 p_k$, obtaining the convex maximization problem
$$\begin{aligned}
&\underset{\bm{p}}{\text{max}}
& & \sum_{k=1}^{K} w_k \log_{2}\bigg (\frac{G_{k,k} e^{q_{k}}}{\sum_{j=1, j\neq k}^{K+N}G_{j,k} e^{q_{j}} x_j +\sigma^2}\bigg )+
\end{aligned}$$
$$\sum_{k=K+1}^{K+N}\rho_k R_{\rm fix}x_k$$
$$\text{s.t.}
\quad e^{q_k} \geq 0, \quad \sum_{k=1}^{K+N}e^{q_k} \leq 1$$
$$\label{const1}
\frac{G_{k,k} e^{q_k}}{\sum_{j=1, j \neq k}^{K+N}G_{j,k} e^{q_j} x_j +\sigma^2} \geq \gamma_{\rm fix} \quad k = K+1,...,K+N.$$
The resulting maximization problem is hence convex and can be solved with standard optimization solutions. We then pick the set of active users for which the problem is feasible and provides that maximum weighted .
solution
---------
The second solution aims at reducing the computational complexity of the optimization problem by clustering the users in different groups, in which users transmit without interference. Then we transmit only to one set of non-interfering users, allocating powers in order to maximize the weighted . The set of active users is the one that maximizes the weighted among all sets.
In order to perform the clustering operation we use the interference matrix $\bm{G}$. We notice that, since the mmWave virtual channel results sparse, the resulting interference matrix computed over the matrix of the gains of the virtual channel $\bm{G}$ will have mostly small or zero off-diagonal entries. In particular, as the ratio $N_a/N_u$ of transmit and receive antennas tends to infinity, the interference matrix itself will become sparse. Indeed, in sparse massive- channels the channel vectors associated to different users become orthogonal as the number of antennas goes to infinity [@Gao]. Our system behaves in a similar manner. Considering the virtual channel model with $L$ non-zero paths and a growing number of antennas at the BS, we have that, asymptotically, users do not interfere. Therefore the interference matrix becomes a diagonal matrix and the graph colouring leads to all users labelled with the same colour.
From the interference matrix we build the interference graph, i.e. a directed graph composed by a node for each user and with edges connecting interfering users: edges are directed from the interfering user toward the user suffering the interference. No edge between nodes $k$ and $n$ is present if and only if $G_{k,n} = 0$. Nodes are then clustered into groups of non-interfering users: this corresponds to color the graph, where each color represents a group and nodes with the same colors must not be connected by edges. We aim at finding the minimum number of colors needed for the graph, and we resort to the solution proposed in [@Segundo]. Fig. \[fig:graph\] shows an example of interference graph obtained for a particular realization of the interference matrix, where, after graph colouring, each node (corresponding to user $U_k$) has been assigned a colour red or green.
\(a) at (1,1) [$U_1$]{}; (b) at (2.5,2.5) [$U_2$]{}; (c) at (2.5,1) [$U_3$]{}; (d) at (4,4) [$U_4$]{}; (a1) at (1,2.5) [$U_5$]{}; (a2) at (4,1) [$U_6$]{}; (a3) at (1,4) [$U_7$]{}; (a) to node\[above\][$G_{1,3}$]{} (c) ; (a) to node\[left\][$G_{1,5}$]{} (a1); (a1) to node\[left\][$G_{5,7}$]{} (a3); (c) to node\[above\][$G_{3,6}$]{} (a2); (a3) to node\[right\][$G_{7,2}$]{} (b); (b) to node\[above\][$G_{2,5}$]{} (a1);
Since users having the same color do not interfere, assuming to transmit only to users of the same color the of user $k$ can be written as $$\gamma_k = \frac{G_{k,k}p_k}{\sigma^2}\,,
\label{eq:snr}$$ which coincides with the .
Let $\mathcal C$ the set of colors, $\mathcal F_c$ be the set of indices of users with color $c \in \mathcal C$, and let $\mathcal V_c$ the set of users with color $c$. To select the subset of active users (with the same color) we resort to the exhaustive search approach, thus computing the weighted corresponding to each active set. For the users the allocated power is $$p_k = \gamma_{\rm fix} \frac{\sigma^2}{G_{k,k}}x_k\,, \quad k \in \mathcal F_c
\label{eq:poF}$$ For users since no interference is present, the optimal power allocation strategy is provided by the waterfilling algorithm
$$\begin{aligned}
&\underset{ \{p_k, k \in \mathcal V_c\}}{\text{max}}
& & \sum_{k\in \mathcal V_c} w_k \log_{2}\bigg (1+\frac{G_{k,k} p_{k}}{\sigma^2}\bigg)
\end{aligned}$$
$$\text{s.t.}
\quad p_k \geq 0, \quad \sum_{k\in \mathcal V_c}p_k \leq 1- \sum_{k \in \mathcal F_c} x_k p_k\,.$$
The graph-coloring solution is summarized in Algorithm 1.
$c^* = {\rm argmax}_c \mathcal R_c$
$\bm{p} = \bm{p}_{c^*}$
$\bm{x} = \bm{x}_{c^*}$
solution
---------
The third solution applies grouping (i.e., ) only on users, while the users are allowed to interfere with the users and among themselves. This allows a more flexible allocation of users, while at the same time due to their low-rate request we expect that the resulting interference on the users is reduced.
Therefore, in this case we apply the only on the interference graph of the users, while we must perform the exhaustive search of the the active users among all users. Therefore, for each color $c \in \mathcal C$ we must solve the optimal scheduling and power allocation problem for the set of users $\mathcal V_c$ and for all users. With respect to the global direct approach we have reduced the set of users, but on the other hand we must run the optimization algorithm for each color.
Numerical Results
=================
We present now numerical results obtained with the different solutions described in previous sections. We consider $N_u=7$ users where $K=4$ are users and $N=3$ are users. We assume that the virtual channel matrix $\bm{H}_v$ is composed by $L=100$ non-zero entries and that they are distributed as in (\[eq:chent\]) and the SINR requirement for users is $5$ dB while leads to an average of 17 dB.
We start verifying the assumption that the interference matrix is sparse, due to the large number of BS antennas. Let us consider matrix $\bm{D}$, defined as the matrix containing off-diagonal elements of $\bm{G}$, i.e., $\bm{D} = \bm{G} - {\rm diag}(\bm{G})$. Fig. \[fig:as\] shows the average number of non-zero entries of $\bm{D}$ vs the number of antennas at the BS. We observe that the average number of non-zero entries sharply decreases as the number of antennas increases, and at $N_a =200$ we have on average about 3 non-zero off-diagonal entries. In the following we consider $N_a =200$.
![Average number of non-zero entries in $\bm{D}$ vs. number of antennas at the BS.[]{data-label="fig:as"}](as.eps){width="1\hsize"}
We now consider the scheduling and power allocation algorithms described in the paper. The weight $w_k$ has been chosen to be equal to $1-\rho$ for each user, i.e. $w_k = 1-\rho \quad \forall k=\{1,...,K\}$ and $\rho \in \{0,1\}$. With this choice we do not distinguish among the users in the two sets, but only use $\rho$ to balance the resources uses for and users. In particular, when $\rho \approx 0$, we maximize the of users, whereas for $\rho \approx 1$ we maximize the of users. Therefore we consider as performance metric the average for both and users. Fig. \[fig:all\] shows the average (over various channel realizations) ASR vs. $\rho$, for the different scheduling and power allocation solutions. The direct algorithm is denoted as [direct]{}, and then we have the and algorithms. As we expect, with increasing values of $\rho$, VR users ASR decreases while FR users ASR increases. Note also that even if $\rho = 1$ we still have a non-zero ASR for users, since all are served and the remaining power is allocated to users.
We notice that the direct solution to power allocation and scheduling is well approximated by both and solutions for VR users, whereas, for FR users, the solution is the best performing. Worse ASR performance for FR users in GC solution are due to the fact that VR users could be labelled with different colours and hence can not transmit together. Notice that the GC solution for VR users outperforms the direct solution. This is due to the fact that even if the maximization problem is focused on FR users ASR, they could be labelled with different colours and hence in GC solution they can not transmit all together, resulting in a power allocation for VR users with higher values than direct solution. This results in higher ASR values for VR users when $\rho$ tends to 1.
![ vs. users for different $\rho$ values.[]{data-label="fig:all"}](noSpars.eps){width="1\hsize"}
Conclusions
===========
In this work we considered a mmWave massive system and we solved the problem of power allocation and users scheduling when two different sets of users transmit: users and users with a required level for transmission. We analysed the direct solution and proposed two algorithms, GC and PGC, that present similar ASR performance for both users and users while reducing the computational complexity. Then we analysed the performance of all the proposed solutions in terms of ASR when the weight $\rho$ (and hence $w_k$) assigned to users changes and showed that the proposed algorithms present the same performance obtained with the direct solution. We also showed that with an increasing number of antennas at the BS the interference matrix becomes diagonal and hence that the low-complexity GC solution achieves optimal results.
[99]{} A. M. Sayeed, “Deconstructing multiantenna fading channels,” [*IEEE Transactions on Signal Processing*]{}, vol. 50, no. 10, pp. 2563-2579, Oct 2002. J. Papandriopoulos and J. S. Evans, “Low-Complexity Distributed Algorithms for Spectrum Balancing in Multi-User DSL Networks”, in Proc. [*IEEE International Conference on Communications*]{}, vol.7, pp. 3270-3275, June 2006. P. S. Segundo “A new DSATUR-based algorithm for exact vertex coloring”, [*Computers and Operations Reasearch*]{}, no.39, pp. 1724-1733, 2012. J. Mo, P. Schniter, N. G. Prelcic and R. W. Heath Jr. “Channel Estimation in Millimiter Wave MIMO Systems with One-Bit Quantization”, in Proc. [*48th Asilomar Conference on Signals, Systems and Computers*]{}, pp. 957-961, Nov. 2014. Z. Gao, L. Dai, C. Yuen, Z. Wang, “Asymptotic Orthogonality Analysis of Time-Domain Sparse Massive MIMO Channels”, [*IEEE Communications Letters*]{}, vol. 19, no. 10, Oct. 2015 S. Rangan, T. S. Rappaport and E. Erkip, “Millimeter-Wave Cellular Wireless Networks: Potentials and Challenges,” [*Proceedings of the IEEE*]{}, vol. 102, no. 3, pp. 366-385, Mar. 2014. A. Gupta and R. K. Jha, “A Survey of 5G Network: Architecture and Emerging Technologies,” [*IEEE Access*]{}, vol. 3, pp. 1206-1232, 2015. K. K. Wong, R. D. Murch, and K. B. Letaief, “A joint-channel diagonalization for multiuser MIMO antenna system,” [*IEEE Trans. Wireless Commun.*]{}, vol. 2, pp. 773-786, Jul. 2003. Q. H. Spencer, A. L. Swindlehurst, and M. Haardt, “Zero-forcing methods for downlink spatial multiplexing in multiuser MIMO channels,” [*IEEE Trans. Signal Process.*]{}, vol. 52, no. 2, pp. 461-471, Feb. 2004. H. Yang and T. L. Marzetta, “Performance of Conjugate and Zero-Forcing Beamforming in Large-Scale Antenna Systems,” [*IEEE Journal on Selected Areas in Communications*]{}, vol. 31, no. 2, pp. 172-179, Feb. 2013. J. Hoydis, S. ten Brink and M. Debbah, “Massive MIMO in the UL/DL of Cellular Networks: How Many Antennas Do We Need?,” [*IEEE Journal on Selected Areas in Communications*]{}, vol. 31, no. 2, pp. 160-171, Feb. 2013. F. Fang, H. Zhang, J. Cheng and V. C. M. Leung, “Energy-Efficient Resource Allocation for Downlink Non-Orthogonal Multiple Access Network,” in [*IEEE Transactions on Communications*]{}, vol. 64, no. 9, pp. 3722-3732, Sept. 2016. S. Shi, M. Schubert and H. Boche, “Rate Optimization for Multiuser MIMO Systems With Linear Processing,” [*IEEE Transactions on Signal Processing*]{}, vol. 56, no. 8, pp. 4020-4030, Aug. 2008. Z. Ding, F. Adachi, and H. V. Poor, “The application of MIMO to non-orthogonal multiple access,”[*IEEE Trans. Wireless Commun.*]{}, vol. 15, no. 1, pp. 537-552, Jan. 2016. Wei Zhiqiang, Yuan Jinhong, Derrick Wing Kwan Ng, Maged Elkashlan, and Ding Zhiguo, “A Survey of Downlink Non-Orthogonal Multiple Access for 5G Wireless Communication Networks”, [*ZTE Communications*]{}, Vol.14 No. 4, Oct. 2016. N. Ksairi, S. Tomasin and M. Debbah, “A Multi-Service Oriented Multiple Access Scheme for M2M Support in Future LTE,” [*IEEE Communications Magazine*]{}, vol. 55, no. 1, pp. 218-224, Jan. 2017. N. Ksairi, S. Tomasin and M. Debbah, “A multi-service oriented multiple-access scheme for next-generation mobile networks,” in Proc. [*2016 European Conference on Networks and Communications (EuCNC)*]{}, Athens, 2016, pp. 355-359. N. Ksairi, S. Tomasin, and B. Tomasi, “Systems and Methods for Scheduling of Resources and Pilot Patterns to User Terminals in a Multi-User Wireless Network,” Patent PCT/EP2016/052202, Filed on February 2, 2016.
[^1]: This work has been supported in part by Huawei Technologies, Milan, Italy.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'In this work we have employed Bethe-Salpeter equation (BSE) under covariant instantaneous ansatz (CIA) to study electromagnetic decays of ground state equal mass vector mesons: $\rho$, $\omega$, $\phi$, $\psi$ and $Y$ through the process $V\rightarrow\gamma*\rightarrow e^+ + e^-$. We employ the generalized structure of hadron-quark vertex function $\Gamma$ which incorporates various Dirac structures from their complete set order-by-order in powers of inverse of meson mass. The electromagnetic decay constants for the above mesons are calculated using the leading order (LO) and the next-to-leading order (NLO) Dirac structures. The relevance of various Dirac structures in this calculation is studied.'
author:
- Shashank Bhatnagar$^1$
- Jorge Mahecha$^2$
- Yikdem Mengesha
title: 'Relevance of various Dirac covariants in hadronic Bethe-Salpeter wave functions in electromagnetic decays of ground state vector mesons'
---
[$^1$ Corresponding author. [email protected]]{}\
[$^2$ [email protected]]{}
Introduction {#sec:intro}
============
Meson decays provide an important tool for exploring the structures of these simplest bound states in QCD, and for studies on non-perturbative behavior of strong interactions. These studies has become a hot topic in recent years. Flavourless vector mesons play an important role in hadron physics due to their direct coupling to photons and thus provide an invaluable insight into the phenomenology of electromagnetic couplings to hadrons. Thus, a realistic description of vector mesons at the quark level of compositeness would be an important element in our understanding of hadron dynamics and reaction processes at scales where QCD degrees of freedom are relevant. There have been a number of studies[@alkofer02; @ivanov99; @close02; @li08; @cvetic04; @hwang97; @alkofer01; @wang08] on processes involving strong, radiative and leptonic decays of vector mesons. Such studies offer a direct probe of hadron structure and help in revealing some aspects of the underlying quark-gluon dynamics.
In this work we study electromagnetic decays of ground state equal mass vector mesons: $\rho,\omega,\phi,\psi$ and $Y$ (each comprising of equal mass quarks) through the process $V\rightarrow\gamma*\rightarrow e + e-$ which proceeds through the coupling of quark-anti quark loop to the electromagnetic current in the framework of Bethe-Salpeter Equation (BSE), which is a conventional non-perturbative approach in dealing with relativistic bound state problems in QCD and is firmly established in the framework of Field Theory. From the solutions, we obtain useful information about the inner structure of hadrons which is also crucial in high energy hadronic scattering and production processes. Despite its drawback of having to input model dependent kernel, these studies have become an interesting topic in recent years since calculations have satisfactory results as more and more data is being accumulated. We get useful insight about the treatment of various processes using BSE due to the unambiguous definition of the 4D BS wave function which provides exact effective coupling vertex (Hadron-quark vertex) of the hadron with all its constituents (quarks).
We have employed QCD motivated Bethe-Salpeter Equation (BSE) under Covariant Instantaneous Ansatz (CIA)[@mitra01; @mitra99; @bhatnagar91; @mitra92; @bhatnagar06; @bhatnagar11] to calculate this process. CIA is a Lorentz-invariant generalization of Instantaneous Ansatz (IA). What distinguishes CIA from other 3D reductions of BSE is its capacity for a two-way interconnection: an exact 3D BSE reduction for a $q\overline{q}$ system (for calculation of mass spectrum), and an equally exact reconstruction of original 4D BSE (for calculation of transition amplitudes as 4D quark loop integrals). In these studies, the main ingredient is the 4D hadron-quark vertex function $\Gamma$ which plays the role of an exact effective coupling vertex of the hadron with all its constituents (quarks). The complete 4D BS wave function $\Psi(P,q)$ for a hadron of momentum $P$ and internal momentum $q$ comprises of the two quark propagators (corresponding to two constituent quarks) bounding the hadron-quark vertex $\Gamma$. This 4D BS wave function is considered to sum up all the non-perturbative QCD effects in the hadron. Now one of the main ingredients in 4D BS wave function (BSW) is its Dirac structure. The copius Dirac structure of BSW was already studied by C.H.L. Smith[@smith69] much earlier. Recent studies[@alkofer02; @li08; @cvetic04] have revealed that various mesons have many different Dirac structures in their BS wave functions, whose inclusion is necessary to obtain quantitatively accurate observables. It was further noticed that all structures do not contribute equally for calculation of various meson observables[@alkofer02; @alkofer01]. Further, it was amply noted in [@hawes98] that many hadronic processes are particularly sensitive to higher order Dirac structures in BS amplitudes. It was further noted in [@hawes98] that inclusion of higher order Dirac structures is also important to obtain simultaneous agreement with experimental decay widths for a range of processes such as: $V\rightarrow e^{+}e^{-}$, $V\rightarrow \gamma P$, $V\rightarrow
PP$, etc. for a given choice of parameters.
Towards this end, to ensure a systematic procedure of incorporating various Dirac covariants from their complete set in the BSWs of various hadrons (pseudoscalar, vector etc.), we developed a naive power counting rule in ref.[@bhatnagar06], by which we incorporate various Dirac structures in BSW, order-by-order in powers of inverse of meson mass. Using this power counting rule we calculated electromagnetic decay constants of vector mesons ($\rho,\omega,\phi$) using only the leading order (LO) Dirac structures \[$i\gamma\cdot\varepsilon$ and $(\gamma\cdot\varepsilon)(\gamma\cdot P)/M$\]. However in Ref.[@bhatnagar11], we rigorously studied leptonic decays of unequal mass pseudoscalar mesons $\pi,K,D,D_{s},B$ and calculated the leptonic decay constants $f_{P}$ for these mesons employing both the leading order (LO) and the next-to-leading order (NLO) Dirac structures. The contributions of both LO and NLO Dirac structures to $f_{P}$ was worked out. We further studied the relevance of both the LO and the NLO Dirac structures to this calculation. In the present paper, we extended these studies to vector mesons and have employed both LO and NLO Dirac structures identified according to our power counting rule, to calculate $f_{V}$ for ground state vector mesons, $\rho,\omega,\phi,\psi,Y$ and in the process we studied the relevance of various Dirac structures to calculation of decay constants $f_{V}$ for vector mesons in the process $V\rightarrow e^{+}e^{-}$. We found that contributions from NLO Dirac structures are smaller than those of LO Dirac structures for all vector mesons. In what follows, we give a detailed discussion of the fit and calculation up to NLO level after a brief review of our framework.
The paper is organized as follows: In section 2 we discuss the structure of BSW for vector mesons under CIA using the power counting rule we proposed earlier. In section 3 we give the calculation of $f_{V}$ for vector mesons. A detailed presentation of results and the numerical calculation is given in section 4. Section 5 is relegated to Discussion.
BSE under CIA {#sec:theory}
=============
We briefly outline the BSE framework under CIA. For simplicity, lets consider a $\mathrm{q}\overline{\mathrm{q}}$ system comprising of scalar quarks with an effective kernel $K$, 4D wave function $\Phi(P,q)$, and with the 4D BSE, $$i(2\pi)^{4}\Delta_{1}\Delta_{2}\Phi(P,q)=\int
d^{4}qK(q,q^{\prime})\Phi(P,q^{\prime}),\label{eq:2.1}$$ where $\Delta_{1,2}=m_{1,2}^2+p_{1,2}^2$are the inverse propagators, and $m_{1,2}$ are (effective) constituent masses of quarks. The 4-momenta of the quark and anti-quark, $p_{1,2}$, are related to the internal 4-momentum $q_{\mu}$ and total momentum $P_{\mu}$ of hadron of mass $M$ as $p_{1,2}{}_{\mu}=\widehat{m}_{1,2}P_{\mu}\pm q_{\mu},$ where $\widehat{m}_{1,2} =[1\pm(m_1^2-m_2^2)/M^2]/2$ are the Wightman-Garding (WG) definitions of masses of individual quarks. Now it is convenient to express the internal momentum of the hadron $q$ as the sum of two parts, the transverse component, $\hat{q}_{\mu}=q_{\mu}-(q\cdot P)/P^2$ which is orthogonal to total hadron momentum $P$ (ie. $\widehat{q}\cdot P=0$ regardless of whether the individual quarks are on-shell or off-shell), and the longitudinal component, $\sigma P_{\mu} = P_{\mu} (q\cdot P)/P^2$, which is parallel to P. Thus we can decompose $q_{\mu}$ as, $q_\mu=(\widehat{q},Md\sigma)$, where the transverse component, $\widehat{q}$ is an effective 3D vector, while the longitudinal component, $Md\sigma$ plays the role of the fourth component and is like the time component. To obtain the 3D BSE and the Hadron-quark vertex, use an Ansatz on the BS kernel $K$ in Eq. (\[eq:2.1\]) which is assumed to depend on the 3D variables $\hat{q}_{\mu}$, $\hat{q}_{\mu}^{\prime}$ [@mitra92] i.e. $$K(q,q^{\prime})=K(\hat{q},\hat{q}^{\prime}).$$ Hence, the longitudinal component, $\sigma P_{\mu}$ of $q_{\mu}$, does not appear in the form $K(\hat{q},\hat{q}^{\prime})$ of the kernel. Defining $\phi(\hat{q})$ as the 3D wave function, $$\
\phi(\hat{q})=\int\limits_{-\infty}^{+\infty}{Md\sigma\Phi(P,q)},$$ Integrating Eq.(1), and making use of Eqs.(2-3), we obtain the 3D Salpeter equation, $$\
(2\pi)^{3}D(\widehat{q})\phi(\widehat{q})=\int
d^{3}\widehat{q}'K(\widehat{q},\widehat{q}')\phi(\widehat{q}'),$$ where $D(\hat{q})$ is the 3D denominator function defined as [@bhatnagar06; @bhatnagar11; @elias11], $$\begin{aligned}
\nonumber\frac{1}{D(\widehat{q})}&=&\frac{1}{2\pi
i}\int\limits_{-\infty}^{+\infty}\frac{
Md\sigma}{\Delta_1\Delta_2}=\displaystyle\frac{\displaystyle\frac{1}{2\omega_1}
+ \frac{1}{2\omega_2}} {(\omega_1+\omega_2)^2-M^2},\\
\omega_{1,2}^2&=&m_{1,2}^2+\hat{q}^2,\end{aligned}$$ whose value given above is obtained by evaluating the contour integration over inverse quark propagators in the complex $\sigma$-plane by noting their corresponding pole positions (for details see[@bhatnagar06; @bhatnagar11]). which is used for making contact with mass spectra of $q\overline{q}$ mesons.
Further, making use of Eq.(2) and (3) on RHS of Eq.(1), we get, $$\ i(2\pi)^{4}\Delta_{1}\Delta_{2}\Phi(P,q)=\int d^{3}\widehat{q}'K(\widehat{q},\widehat{q}')\phi(\widehat{q}').$$ From equality of RHS of Eq.(4) and (6), we see that an exact interconnection between 3D and 4D BS wave functions is thus brought out. The 4D Hadron-quark vertex function for scalar quarks under CIA can be identified as: $$\Delta_{1}\Delta_{2}\Phi(P,q)=\frac{D(\hat{q})\phi(\hat{q})}{2\pi i}
\equiv\Gamma.\label{eq:2.7}$$
Now for fermionic quarks, the 4D BSE under gluonic which is akin to vector type interaction kernel with a 3D support can be written as: $$\begin{aligned}
\nonumber\ i(2\pi)^{4}\Psi(P,q)&=&S_{F1}(p_{1})S_{F2}(-p_{2})\int d^{4}q'K(\widehat{q},\widehat{q}')\Psi(P,q');\\
K(\widehat{q},\widehat{q}')&=&F_{12}i\gamma_{\mu}^{(1)}\gamma_{\mu}^{(2)}V(\widehat{q},\widehat{q}')\end{aligned}$$
Here, $F_{12}$ is the color factor, $(\bm{\lambda}_1/2)\cdot(\bm{\lambda}_2/2)$ and the potential $V$ involves the scalar structure of the gluon propagator in the perturbative (o.g.e.) as well as the non-perturbative (confinement) regimes. The full structure of $V$ is [@mitra01]:
$$\begin{aligned}
\nonumber K(q,q^{\prime})&=&K(\hat{q},\hat{q}^{\prime})\\
\nonumber K(\widehat{q},\widehat{q'})&=&\left(\frac{1}{2}\bm{\lambda}_1\cdot\frac{1}{2}\bm{\lambda}_2\right)\gamma^{(1)}_{\mu}\gamma^{(2)}_{\mu}V(\widehat{q}-\widehat{q'})\\
\nonumber V(\hat{q}-\hat{q'})&=&V_{OGE}+V_C\\
\nonumber V_{OGE}&=&\frac{4\pi\alpha_{s}(Q^{2})}{(\widehat{q}-\widehat{q}')^{2}}\\
\nonumber V_C&=&\frac{3\omega_{q\bar{q}}^2}{4}\int d^3\bm{r}f(r)
e^{i(\hat{q}-\hat{q'})\cdot\bm{r}}\\
f(r)&=&
\frac{r^2}{(1+4a_0\hat{m_1}\hat{m_2}M^2r^2)^{1/2}}
-\frac{C_0}{\omega_0^2},\end{aligned}$$
which is taken as one-gluon-exchange like as regards color \[$(\bm{\lambda}^{(1)}/2)\cdot(\bm{\lambda}^{(2)}/2)$\] and spin ($
\gamma_\mu^{(1)}\gamma_\mu^{(2)}$) dependence. The scalar function $V(q-q')$ is a sum of one-gluon exchange $V_{OGE}$ and a confining term $V_C$. This confining term simulates an effect of an almost linear confinement ($\sim r$) for heavy quark ($c,b$) sector, while retaining harmonic form ($\sim r^2$) for light quark ($u,d$) sector as is believed to be true for QCD. $$\begin{aligned}
\nonumber\ \omega_{q\overline{q}}^2&=&4\widehat{m}_1\widehat{m}_2M_>
\omega_0^2\alpha_S(M_>^2),\\
M_>&=&\mbox{Max}(M,m_1+m_2).\end{aligned}$$ The values of basic constants are: $C_0=0.29, \omega_0=0.158$ GeV, $\Lambda=0.200$ GeV, $m_{u,d}=0.265$ GeV, $m_s=0.415$ GeV, $m_c=1.530$ GeV and $m_b=4.900$ GeV [@mitra01; @mitra99; @bhatnagar11]. However the form of BSE in Eq.(8) is not convenient to use in practice since Dirac matrices lead to several coupled integral equations. However a considerable simplification is effected by expressing them in Gordon-reduced form which is permissible on the mass shells of quarks (ie. on the surface $P\cdot q=0$). The Gordon reduced BSE form of the fermionic BSE can be written as [@mitra01]: $$\begin{aligned}
\nonumber\ \Delta_{1}\Delta_{2}\Phi(P,q)&=&-i(2\pi)^{4}\int d^{4}q'\widetilde{K}_{12}(\widehat{q},\widehat{q}')\Phi(P,q'),\\
\widetilde{K}_{12}(\widehat{q},\widehat{q}')&=&F_{12}V_{\mu}^{(1)}V_{\mu}^{(2)}V(\widehat{q},\widehat{q}')\end{aligned}$$ where the connection between $\Psi$ and $\Phi$ (whose structure is identical as $\Phi$ in case of scalar quarks in Eq.(1)) is, $$\begin{aligned}
\nonumber\Psi(P,q)&=&(m_1-i\gamma^{(1)}\cdot p_1)(m_{2}+i\gamma^{(2)}\cdot p_2)\Phi(P,q),\\
\nonumber V_{\mu}^{(1,2)}&=&\pm 2m_{1,2}\gamma_{\mu}^{1,2},\\
V_{\mu}^{(1,2)}&=&(p_{1,2}+p'_{1,2})_{\mu}+i\sigma_{\mu\nu}^{(1,2)}(p_{1,2}+p'_{1,2})_{\nu}.\end{aligned}$$
Now to reduce the above BSE to the 3D form, all time-like components $\sigma,\sigma'$ of momenta in $V_{\mu}^{(1)}V_{\mu}^{(2)}$ on RHS of Eq.(9) are replaced by their on-shell values giving us the 3D form, $\bm{V}_1\cdot\bm{V}_2$. Thus $$\begin{aligned}
&&\nonumber\ V_{\mu}^{(1)}V_{\mu}^{(2)}\Rightarrow\bm{V}_1\cdot\bm{V}_2=-4\widehat{m}_{1}\widehat{m}_{2}M^{2}-(\widehat{q}-\widehat{q}')^{2}-2(\widehat{m}_{1}-\widehat{m}_{2})P\cdot(\widehat{q}+\widehat{q}')\\&&
\nonumber\ -i(2\widehat{m}_1P+\widehat{q}+\widehat{q}')_i\sigma_{ij}^{(2)}(\widehat{q}-\widehat{q}')+i(2\widehat{m}_2P-\widehat{q}-\widehat{q}')_i\sigma_{ij}^{(1)}(\widehat{q}-\widehat{q}')_j
+\sigma_{ij}^{(1)}\sigma_{ij}^{(2)}.\end{aligned}$$
The 3D form of BSE then works out as [@mitra99]: $$\begin{aligned}
\nonumber D(\widehat{q})\phi(\widehat{q})&=&\omega_{q\overline{q}}^2\widetilde{D}(\widehat{q})\phi(\widehat{q}),\\
\nonumber \widetilde{D}(\widehat{q})&=&4\widehat{m}_1\widehat{m}_2M^2(\bm{\nabla}^2+C_0/\omega_0^2)+4\widehat{q}^2\bm{\nabla}^2+8\widehat{q}\cdot\bm{\nabla}\\
&& +18-8\bm{J}\cdot\bm{S}+\frac{4C_0}{\omega_0^{2}}\widehat{q}^2.\end{aligned}$$ This is reducible to the equation for a 3D harmonic oscillator with coefficients depending on the mass $M$ and total quantum number $N$. The ground state wave functions [@mitra01; @mitra99; @bhatnagar06] deducible from this equation have gaussian structure and are expressed as: $$\ \phi(\widehat{q})=e^{-\widehat{q}^2/2\beta^2}.$$ and is appropriate for making contact with O(3)-like mass spectrum (for details see [@mitra01]). It is to be noted that this 3D BSE (in Eq.(14)) which is responsible for determination of mass spectra of mesons in CIA is formally equivalent (see [@mitra01; @mitra99; @bhatnagar05]) to the corresponding mass spectral equation deduced earlier using Null-Plane Approximation (NPA)[@mitra90]. Thus the mass spectral predications for $q\bar{q}$ systems in BSE under CIA are identical to the corresponding mass spectral predictions for these systems in BSE under NPA [@mitra90] (see [@mitra01; @mitra99] for details).
We further wish to mention that a similar form for ground state wave function in harmonic oscillator basis using variational arguments has been used in [@arndt99]. In ground state wave function $\phi(\widehat{q})$ in Eq.(15), $\beta$ is the inverse range parameter which incorporates the content of BS dynamics and is dependent on the input kernel $K(q,q^\prime)$ (for details see [@mitra01; @bhatnagar06; @bhatnagar11]) The structure of inverse range parameter $\beta$ in wave functions $\phi(\widehat{q})$ is given as [@mitra01; @bhatnagar06; @bhatnagar11; @elias11]: $$\begin{aligned}
\nonumber\beta^2&=&\left(\frac{2}{\gamma^2}\widehat{m}_1\widehat{m}_2M\omega_{q\bar{q}}^2\right)^{1/2},\\
\nonumber\gamma^2&=&1-\frac{2\omega_{q\bar{q}}^2C_0}{M_>\omega_0^2},\\
\nonumber\omega_{q\overline{q}}^2&=&4\widehat{m}_1\widehat{m}_2M_>
\omega_0^2\alpha_S(M_>^2),\\
\ M_>&=&\mbox{Max}(M,m_1+m_2).\end{aligned}$$
Dirac structure of Hadron-quark vertex function for P-mesons in BSE with power counting scheme
----------------------------------------------------------------------------------------------
Thus, for fermionic quarks, the full 4D BS wave function can be written as
$$\Psi(P,q)=S_{F}(p_1)\Gamma S_{F}(-p_2),$$
where the 4D hadron-quark vertex function is [@mitra01; @bhatnagar06; @bhatnagar06; @bhatnagar11] i.e. $$\Gamma=\frac{1}{2\pi i}(\Omega\cdot\varepsilon) N_{V}D(\widehat{q})\phi(\widehat{q}).$$ The 4D hadron-quark vertex $\Gamma$ in the above equation, satisfies a 4D BSE with a natural off-shell extension over the entire 4D space (due to the positive definiteness of the quantity $\hat{q}^{2}=q^{2}-(q\cdot
P)^{2}/P^{2}$ throughout the entire 4D space) and thus provides a fully Lorentz-invariant basis for evaluation of various transition amplitudes through various quark loop diagrams. $N_{V}$ in the above equation is the 4D BS normalizer.
In the hadron-quark vertex function, $\Gamma$ above, ($\Omega\cdot\varepsilon$) contains the relevant Dirac structures [@smith69] which makes $\Gamma$ a $4\times4$ matrix in the spinor space. In our model, the relevant Dirac structures in $\Gamma$ are incorporated in accordance with our recently proposed power counting rule [@bhatnagar06; @shi07] which identifies the leading order (LO) Dirac structures from the next-to-leading order (NLO) Dirac structures and for a vector meson is expressed as: $$\begin{aligned}
\nonumber
(\Omega\cdot\varepsilon)&=&i(\gamma\cdot\varepsilon)A_{0}+(\gamma\cdot\varepsilon)(\gamma\cdot P)
\frac{A_1}{M}
+[q\cdot\varepsilon-(\gamma\cdot\varepsilon)(\gamma\cdot q)]\frac{A_2}{M}\\&&
\nonumber+i\frac{A_3}{M^2}[(\gamma\cdot\varepsilon)(\gamma\cdot P)(\gamma\cdot q) -
(\gamma\cdot\varepsilon)(\gamma\cdot q)(\gamma\cdot P)+2i(q\cdot\varepsilon)(\gamma\cdot P)]\\&&
+(q\cdot\varepsilon)\frac{A_4}{M}+i(q\cdot\varepsilon)(\gamma\cdot P)\frac{A_5}{M^2}
- i(q\cdot\varepsilon)(\gamma\cdot q)\frac{A_6}{M^2}
+(q\cdot\varepsilon)[(\gamma\cdot P)(\gamma\cdot q)-(\gamma\cdot q)(\gamma\cdot P)]\frac{A_7}{M^3},\end{aligned}$$ where $A_i(i=0,...,7)$ are taken as the eight dimensionless constant coefficients to be determined. But since we take constituent quark masses, where the quark mass, $M$ is approximately half the hadron mass, $M$, we can use the ansatz, $$\ q\ll P\sim M$$ in the rest frame of the hadron. Then each of the eight terms in the above equation receives suppression by different powers of $1/M$. Thus we can arrange these terms as an expansion in powers of $O(1/M)$. We can then see that in the expansion of $\Gamma\cdot\varepsilon$ that the structures associated with the coefficients, $A_{0},A_1$ have magnitudes, $O(1/M^0)$ and are of leading order (LO). Those with $A_2,
A_3, A_4, A_5$ are $O(1/M^1)$ and are next-to-leading order (NLO), while those with $A_6,A_7$ are $O(1/M^2)$ and are NNLO. This naive power counting rule suggests that the maximum contribution to calculation of any vector meson observable should come from Dirac structures $\gamma\cdot\varepsilon$ and $(\gamma\cdot\varepsilon)(\gamma\cdot P)/M$ associated with coefficients, $A_0$ and $A_1$ respectively, followed by the higher order Dirac structures associated with the other four coefficients, $A_2,A_3,A_4,A_5$ and then by Dirac structures associated with coefficients, $A_6,A_7$.
In this work we try to calculate the decay constants using LO and NLO Dirac structures since we wish to calculate the decay constants up to next-to-leading orders. Thus we take the form of the hadron-quark vertex function as, $$\begin{aligned}
\nonumber\ \Gamma&=&\left\{i(\gamma\cdot\varepsilon)A_{0} +
(\gamma\cdot\varepsilon)(\gamma\cdot P)\frac{A_1}{M}\right.
+\left[q\cdot\varepsilon-(\gamma\cdot\varepsilon)(\gamma\cdot q)\right]\frac{A_2}{M}\\&&
\nonumber+i\left[(\gamma\cdot\varepsilon)(\gamma\cdot P)(\gamma\cdot q) -
(\gamma\cdot\varepsilon)(\gamma\cdot q)(\gamma\cdot P)+2i(q\cdot\varepsilon)(\gamma\cdot P)\right]
\frac{A_3}{M^2}\\&&
\left.+(q\cdot\varepsilon)\frac{A_4}{M}+i(q\cdot\varepsilon)(\gamma\cdot P)
\frac{A_5}{M^2}\right\}\frac{1}{2\pi i}N_VD(\widehat{q})\phi(\widehat{q}).\end{aligned}$$ In the above equation, $N_{V}$ is the 4D BS normalizer for ground state vector meson with internal momenta $q$, and is worked out in the framework of Covariant Instantaneous Ansatz (CIA) to give explicit covariance to the full fledged 4D BS wave functions, $\Psi(P,q)$ and hence to the Hadron-quark vertex function, $\Gamma$, employed for calculation of decay constants. In the structure of $\Gamma$ above, $\phi(\widehat{q})$ is the ground state 3D BS wave function for vector meson with internal momenta $q$ and is given in Eq.(15).
Electromagnetic decays of vector mesons through the process $\bf V\rightarrow\gamma*\rightarrow e^++e^-$ {#sec:methodology}
========================================================================================================
Transition amplitude
--------------------
The vector meson decay proceeds through the quark-loop diagram shown below (See Fig.1).
![Quark loop diagram for $V\rightarrow\gamma^*\rightarrow
e^++e^-$ showing the coupling of electromagnetic current to the quark loop.[]{data-label="fig:quarkloop"}](fig1.eps){width="0.8\linewidth" height="0.3\linewidth"}
The coupling of a vector meson of momentum $P$ and polarization $\varepsilon_{\mu}$ to the photon is expressed via dimensionless coupling constant $g_V$ which can be described by the matrix element, $$\ \frac{M^2}{g_V}\varepsilon_{\mu}(P)=\langle0\vert\overline{Q}\widehat{\Theta}\gamma_{\mu}Q\vert V(P)\rangle$$ (where $Q$ is the flavour multiplet of quark field and $\widehat{\Theta}$ is the quark electromagnetic charge operator) which can in turn be expressed as a loop integral, $$\ \frac{M^2}{g_V}\varepsilon_\mu(P)=\sqrt3e_Q\int d^4q Tr[\Psi_V(P,q)i\gamma_\mu].$$ Here $e_Q$ arises from the flavour configuration of individual vector mesons and has values: $e_Q=1/\sqrt{2},1/3,1/\sqrt{18},2/3$ and $1/3$ for $\rho,\phi,\omega,\psi$ and $Y$ respectively, and the polarization vector $\varepsilon_\mu$ of V-meson satisfies, $\varepsilon\cdot P=0$. Defining leptonic decay constant,$f_{V}$ as, $f_{V}=M/(e_{Q}g_{V})$ [@bhatnagar06], we can express $$\ f_{V}\varepsilon_{\mu}(P)=\frac{\sqrt{3}}{M}\int d^{4}q Tr\left[\Psi_{V}(P,q)i\gamma_{\mu}\right]$$ Plugging $\Psi_{V}(P,q)$, which involves the structure of hadron-quark vertex function in Eq. (2.8) flanked by the Dirac propagators of the two quarks as in Eq. (2.3), into the above equation, evaluating trace over the gamma matrices and noting that only the components of terms on the right hand side in the direction of $\varepsilon_\mu$ will contribute to the integral, we multiply both sides of the above integral by $P_\mu/M^2$, we can then express the leptonic decay constant $f_V$ as, $$\ f_V=f_V^0+f_V^1+f_V^2+f_V^3+f_V^4+f_V^5$$ where $f_V^0, f_V^1,...,f_V^5$ are contributions to $f_V$ from the six Dirac structures associated with $A_0,A_1,...,A_5$ in the expression for hadron-quark vertex function $\Gamma(\widehat{q})$ and are expressed analytically in terms of $d\sigma$ integrations over the poles $\Delta_{1,2}$ of the quark propagators as: $$\begin{aligned}
\nonumber\ f_V^0&=&\sqrt{3}N_V\frac{A_0}{M}\int d^3\widehat{q}D(\widehat{q})\phi(\widehat{q})\int \frac{Md\sigma}{2\pi i\Delta_1\Delta_2}4\left[\left(\frac{M^2}{6}+\frac{2}{3}m^2\right)+\frac{\Delta_1+\Delta_2}{6}\right]\\
\nonumber\ f_V^1&=&\sqrt{3}N_V\frac{A_1}{M}\int d^3\widehat{q}D(\widehat{q})\phi(\widehat{q})\int\frac{Md\sigma}{2\pi i\Delta_1\Delta_2}(-4mM)\\
\nonumber\ f_V^2&=&\sqrt{3}N_V\frac{A_2}{M}\int d^3\widehat{q}D(\widehat{q})\phi(\widehat{q})\int\frac{Md\sigma}{2\pi i\Delta_1\Delta_2}\left[-\frac{4}{3}m(\Delta_1-\Delta_2)\right]\\
\nonumber\ f_V^3&=&\sqrt{3}N_V\frac{A_3}{M}\int d^3\widehat{q}D(\widehat{q})\phi(\widehat{q})\int\frac{Md\sigma}{2\pi i \Delta_1\Delta_2}\left[-\frac{8}{3}(\Delta_1+\Delta_2)+\left(\frac{16}{3}m^2-\frac{4}{3}M^2\right)\right]\\
\nonumber\ f_V^4&=&\sqrt{3}N_V\frac{A_4}{M}\int d^3\widehat{q}D(\widehat{q)}\phi(\widehat{q})\int\frac{Md\sigma}{2\pi i \Delta_1\Delta_2}\left[-\frac{2m}{3M}(\Delta_1+\Delta_2)+\left(\frac{4m^3}{3M}-\frac{1}{3}mM\right)\right]\\
\ f_V^5&=&\sqrt{3}N_V\frac{A_5}{M}\int d^3\widehat{q}D(\widehat{q)}\phi(\widehat{q})\int\frac{Md\sigma}{2\pi i \Delta_1\Delta_2}\left[\left(\frac{4}{3}m^2-\frac{2}{3}M^2\right)(\Delta_1-\Delta_2)\right].\end{aligned}$$ In deriving the above expressions, we have made use of the following relation showing the normalization over the polarization vector, $\varepsilon(P)$ for $V$- meson of 4-momentum, $P$ as, $$\ \varepsilon_\mu\varepsilon_\nu=\frac{1}{3}\left(\delta_{\mu\nu}+\frac{P_\mu P_\nu}{M^2}\right),$$ We made use of the above relation to express the quantities involving dot products of $\varepsilon$ with various momenta like, $(p_1\cdot\varepsilon)(p_2\cdot\varepsilon)$, $(p_1\cdot\varepsilon)(q\cdot\varepsilon)$ and $(p_2\cdot\varepsilon)(q\cdot\varepsilon)$ in terms of dot products of momenta as, $$\begin{aligned}
&&\nonumber\ (p_1\cdot\varepsilon)(p_2\cdot\varepsilon)=\frac{1}{3}p_1\cdot p_2-\frac{(p_1\cdot P)(p_2\cdot P)}{3M^2}\\&&
\nonumber\ (p_1\cdot\varepsilon)(q\cdot\varepsilon)= \frac{1}{6}(p_1^2-p_1\cdot p_2)-\frac{(p_1\cdot P)(p_1\cdot P-p_2\cdot P)}{6M^2}\\&&
\ (p_2\cdot\varepsilon)(q\cdot\varepsilon)=\frac{1}{6}(p_1\cdot p_2-p_2^2)-\frac{(p_2\cdot P)(p_1\cdot P-p_2\cdot P)}{6M^2}\end{aligned}$$
These dot products of momenta were in turn expressible in terms of the inverse propagators, $\Delta_{1,2}$ as: $$\begin{aligned}
\nonumber p_1\cdot P&=&\frac{1}{2}(\Delta_1-\Delta_2-M^2)\\
\nonumber p_2\cdot P&=&\frac{1}{2}(-\Delta_1+\Delta_2-M^2)\\
\nonumber p_1\cdot p_2&=&m^2-\frac{1}{2}(\Delta_1+\Delta_2+M^2)\\
\nonumber p_{1,2}^2&=&\Delta_{1,2}-m^2\\
\nonumber p_1\cdot q&=&\frac{3}{4}\Delta_1+\frac{1}{4}(\Delta_2+M^2)-m^2\\
\nonumber p_2\cdot q&=&-\frac{3}{4}\Delta_2-\frac{1}{4}(\Delta_1+M^2)+m^2\\
P\cdot q&=&\frac{1}{2}(\Delta_1-\Delta_2)\end{aligned}$$
Thus all expressions for $f_{V}^{i}$ above were expressible in terms of $\Delta_{1,2}$. Then carrying out integrations over the off-shell variable, $d\sigma$ by the method of contour integrations by noting the pole positions in the complex $\sigma-$ plane: $$\begin{aligned}
\nonumber \Delta_1&=&0\Rightarrow\sigma_1^\pm=\pm \frac{\omega_1}{M}-\widehat{m}_1\mp i\epsilon,\\
\nonumber \Delta_2&=&0\Rightarrow\sigma_2^\mp=\mp\frac{\omega_2}{M}-\widehat{m}_2\pm i\epsilon,\\
\nonumber \omega_1^2&=&\omega_2^2=m^2+\widehat{q}^2\\
\widehat{m}_1&=&\widehat{m}_2=\frac{1}{2}\end{aligned}$$
we get, the following integrals: $$\begin{aligned}
\nonumber\int\frac{Md\sigma}{\Delta_1\Delta_2}(\Delta_1+\Delta_2)&=&D_0(\widehat{q});\\
\int\frac{Md\sigma}{\Delta_1\Delta_2}&=&\frac{1}{D(\widehat{q})},\end{aligned}$$
where, $$\begin{aligned}
\nonumber D(\widehat{q})&=&\omega D_0(\widehat{q}),\\
D_0(\widehat{q})&=&4\omega^2-M^2.\end{aligned}$$ Thus we can express the various components $f_V^i$ $(i=0,...,5)$ of $f_{V}$ in Eq. (3.5) as,
$$\begin{aligned}
&&\nonumber\ f_V^0=\sqrt{3}N_V\frac{A_0}{M}\int d^3\widehat{q}\phi(\widehat{q})4\left[\frac{M^2}{6}+\frac{2}{3}m^2+\frac{D_0(\widehat{q})}6\right]\\&&
\nonumber\ f_V^1=\sqrt{3}N_V\frac{A_1}{M}\int d^3\widehat{q}\phi(\widehat{q})(-4mM)\\&&
\nonumber\ f_V^2=0\\&&
\nonumber\ f_V^3=\sqrt{3}N_V\frac{A_3}{M}\int d^3\widehat{q}\phi(\widehat{q})\left[-\frac{8}{3}D_{0}(\widehat{q})+\frac{16}{3}m^2-\frac{4}{3}M^2\right]\\&&
\nonumber\ f_V^4=\sqrt{3}N_V\frac{A_4}{M}\int d^3\widehat{q}\phi(\widehat{q})\left[-\frac{2m}{3M}D_0(\widehat{q})+\frac{4m^3}{3M}-\frac{1}{3}mM\right]\\&&
\ f_V^5=0\end{aligned}$$
Note that the components, $f_V^2$ and $f_V^5$ are $0$ on account of equal mass kinematics.
Note that each of the $f_V^i$ involves the BS normalizer $N_V$. This is evaluated using the current conservation condition:[@bhatnagar06; @bhatnagar11; @elias11], $$2iP_\mu=(2\pi)^4\int d^4q \mbox{Tr}\left\{\overline{\Psi}(P,q)\left[\frac{\partial}{\partial P_{\mu}}S_{F}^{-1}(p_1)\right]\Psi(P,q)S_{F}^{-1}(-p_2)\right\} + (1\rightleftharpoons2)$$ Putting BS wave function $\Psi(P,q)$ for a vector meson in Eqs. (2.3) and (2.8) in the above equation, carrying out derivatives of inverse quark propagators of constituent quarks with respect to total hadron momentum $P_\mu$, evaluating trace over products of gamma matrices, following usual steps, and multiplying both sides of the equation by $P_\mu/(-M^2)$ to extract out the normalizer $N_V$ from the above equation, we then express the above equation in terms of the integration variables $\widehat{q}$ and $\sigma$. Noting that the 4D volume element $d^4q =d^3\widehat{q}Md\sigma$, we then perform the contour integration in the complex $\sigma$- plane by making use of the corresponding pole positions. For details of these mathematical steps involved in the calculations of BS normalizers for vector and pseudoscalar mesons see [@bhatnagar06; @bhatnagar11], where in the present calculation, we take both, the leading order (LO) as well as the next-to-leading order (NLO) Dirac structures for vector mesons in their respective 4D BS wave functions $\Psi(P,q)$. Then integration over the variable $\widehat{q}$ is finally performed to extract out the numerical results for $N_V$ for different vector mesons. The calculation of $N_V$ is quite complex due to the $6$ Dirac structures involved in the calculation. The structure of the normalizer is of the form, $$N_V^{-2}=i(2\pi)^2\int d^3\widehat{q}D^2(\widehat{q})\phi^2(\widehat{q})
\displaystyle\sum\limits_{ij}A_i A_jI_{ij}(m,M,\widehat{q},S).
\label{eq:Nvm2}$$ Here, $A=(A_0,A_1,A_2,A_3,A_4,A_5)$ are functions of parameters $A_i$, $S=(R,D_1,D_2,D_{11},$ $D_{12},D_{22})$, where the $S_i$ are the results of integrations over off-shell parameter $\sigma$ whose results are given in Eq. (\[eq:theSi\]). The $I_{ij}(m,M,\widehat{q},R)$ are involved functions of $m,M,\widehat{q}$ and $R$. Explicit expressions are listed below in Eq. (\[eq:I20\]): $$\begin{aligned}
I_{00}&=&\frac{1}{3}\left[\frac{1}{M^2}\left(1-12\frac{m^2}{M^2}\right)D_{11}+\frac{3}{M^2}\left(-1+\frac{m^2}{M^2}\right)D_{12}+\frac{1}{M^2}(7m^4-3m^2M^2)R\right.\\
&\phantom{=}&\left.+\frac{1}{M^2}\left(M^2+\frac{13m^4}{M^2}-5m^2\right)D_1-\frac{13m^4D_2}{M^4}+\frac{9m^2D_{22}}{M^4}\right]\\
I_{11}&=&\frac{2}{3}\left[-\frac{4D_{11}}{M^2}+\frac{3D_{12}}{M^2}-\frac{4D_{22}}{M^2}-\frac{8M^2D_1}{M^2}+\frac{(8m^2+M^2)D_2}{M^2}+2(4m^2+M^2)R\right]\\
I_{22}&=&\frac{2}{3}\left[\frac{D_{11}}{M^2}+\frac{3D_{12}}{M^2}-\frac{4D_{22}}{M^2}-\frac{8m^2D_1}{M^2}+\frac{(8m^2+M^2)D_2}{M^2}+2(4m^2+M^2)R\right]\\
I_{33}&=&\frac{1}{3}\left[\frac{D_{11}}{M^2}\left(\frac{-6m^4}{M^4}+\frac{9m^2}{M^2}-\frac{22m^3}{M^3}-\frac{55m^2}{2M^2}+13\right)\right.\\
&\phantom{=}& +\frac{D_{12}}{M^2}\left(\frac{18m^4}{M^4}-\frac{7m^2}{M^2}+\frac{70m^2}{M^2}+\frac{12m^3}{M^3}-4\right)\\
&\phantom{=}& +\frac{D_{22}}{M^2}\left(\frac{9m^4}{M^4}-\frac{67m^2}{M^2}+\frac{10m^3}{M^3}+35-\frac{9m^2}{M^2}\right)\\
&\phantom{=}& +\frac{D_1}{M^2}\left(\frac{29}{2}M^2-13m^2+\frac{12m^4}{M^2}-\frac{3m^3}{M}-\frac{48m^4}{M^2}+\frac{16m^6}{M^4}-\frac{36m^5}{M^3}\right)\\
&\phantom{=}& +\frac{D_2}{M^2}\left(22M^2-7m^2-\frac{12m^4}{M^2}+\frac{3m^3}{M}+\frac{48m^4}{M^2}-\frac{16m^6}{M^4}+\frac{4m^5}{M^3}\right)\\
&\phantom{=}&\left.+\frac{R}{M^2}\left(-m^2M^2-14m^3M+76m^4+16mM^3+6M^4+\frac{5}{2}m^4\right)\right]\\
I_{44}&=&\frac{1}{3}\left[\frac{D_{11}}{M^2}\left(\frac{4m^2}{M^2}+1\right)+\frac{7D_{12}}{2M^2}+\frac{D_{22}}{M^2}\left(\frac{1}{6}-\frac{4m^2}{M^2}\right)\right.\\
&\phantom{=}& +\left(2-\frac{6m^4}{M^4}-\frac{m^5}{M^5}-\frac{6m^2}{M^2}-\frac{m^3}{M^3}\right)D_1+\left(\frac{1}{6}+\frac{m^5}{M^5}+\frac{6m^4}{M^4}-\frac{8m^2}{M^2}\right)D_2\\
&\phantom{=}&\left. +\left(-4m^2+\frac{8m^4}{M^2}+\frac{m^5}{M^3}+\frac{M^2}{3}\right)R\right]\\
I_{55}&=&\frac{1}{3}\left[-\frac{3}{2M^2}\left(\frac{m^4}{3M^4}+\frac{11m^2}{3M^2}+\frac{3}{2}\right)D_{12}+\frac{D_1}{M^2}\left(-M^2+\frac{3m^4}{2M^2}+\frac{7m^5}{2M^3}+\frac{17m^2}{4}\right)\right.\\
&\phantom{=}&\left.+\frac{D_2}{M^2}\left(-\frac{M^2}{4}-\frac{3m^5}{2M^3}-\frac{7m^4}{2M^2}+\frac{15m^2}{4}\right)+\frac{R}{M^2}\left(-\frac{3M^4}{2}+\frac{9m^2M^2}{4}+\frac{7m^4}{2}\right)R\right]\\
I_{01}&=&\left[-\frac{4mD_{11}}{M^3}-\frac{8mD_{22}}{M^3}+\frac{4mD_{12}}{M^3}-\frac{4D_1}{3M}\left(\frac{4m^3}{M^2}+m\right)\right.\\&&
\left.+\frac{4D_2}{M}\left(\frac{4m^3}{3M^2}-m\right)+\left(\frac{16m^3}{3M}-\frac{4mM}{3}R\right)\right]\\
I_{02}&=&\left[\frac{8m}{M}(D_1-D_2)-8mMR\right];\\
I_{03}&=&\left[-\frac{2m}{M^3}D_{11}+\frac{4m}{3M^3}D_{12}+\frac{2m}{M^3}D_{22}+\frac{8D_1}{3M}(\frac{m^3}{M^2}-m)+\frac{3m^3}{3M^3}D_2\right.\\&&
\left.+(\frac{8m^3}{3M}-\frac{2mM}{3})R\right]\\
I_{04}&=&\left[-\frac{7m^2D_{11}}{3M^4}+\frac{D_{22}}{3M^2}\left(-\frac{2m^4}{M^4}-\frac{7m^2}{M^2}+8\right)+\frac{D_{12}}{3M^2}\left(\frac{2m^2}{M^2}-8\right)\right.\\&&
\left.+\frac{D_1}{M^2}\left(-M^2-\frac{2m^4}{3M^2}+\frac{8m^2}{3}\right)+\frac{D_2}{M^2}\left(\frac{4M^2}{3}+\frac{2m^4}{3M^2}-\frac{8m^2}{3})+\frac{R}{M^2}(-m^2M^2-2m^4\right)\right]\\
I_{05}&=&\left[\frac{D_{11}}{M^2}\left(\frac{6m^2}{M^2}-\frac{34}{3}\right)+\frac{D_{22}}{3M^2}\left(\frac{26m^2}{M^2}-14\right)-\frac{D_{12}}{3M^2}\left(\frac{44m^2}{M^2}+20\right)\right.\\&&
+\frac{D_1}{M^2}\left(-6M^2+\frac{8m^4}{3M^2}-12m^2\right)D_1+\frac{D_2}{M^2}\left(-\frac{22M^2}{3}-\frac{8m^4}{M^2}+\frac{44m^2}{3}\right)\\&&
+\frac{R}{M^2}\left(14m^2M^2-\frac{8m^4}{3}-\frac{10M^4}{3}\right)\}\\
I_{12}&=&\left[\frac{D_{11}}{M^2}\left(\frac{m^2}{M^2}+\frac{2}{3}\right)+\frac{D_{22}}{M^2}\left(\frac{m^2}{M^2}-4\right)+\frac{D_{12}}{M^2}\left(-\frac{2m^2}{M^2}+\frac{10}{3}\right)\right.\\&&
\left.+\frac{D_1}{M^2}\left(\frac{8M^2}{3}-\frac{2m^2}{3}\right)+\frac{D_2}{M^2}\left(-2M^2+\frac{2m^2}{3}\right)+\frac{5m^2R}{3}\right]\\
I_{13}&=&\left[\frac{D_{11}}{3M^3}\left(-\frac{4m^3}{M^2}-34m\right)+\frac{D_{22}}{3M^3}\left(-\frac{4m^3}{M^2}-26m\right)+\left(\frac{8m^2}{M^4}+\frac{20}{M^2}\right)D_{12}\right.\\&&
+\frac{D_{22}}{M^3}\left(-\frac{16m^5}{3M^4}+\frac{8m^3}{3M^2}+28m\right)+\frac{D_2}{M}\left(\frac{16m^5}{3M^4}-\frac{8m^3}{3M^2}-\frac{46m}{3}\right)\\&&
\left.+\left(-\frac{28m^4}{3M^2}-\frac{8m^3}{3M}-2mM\right)R\right]\\
I_{14}&=&\left[\frac{D_{11}}{M^2}\left(-\frac{m^4}{M^4}+\frac{4}{3}\right)+\frac{D_{22}}{M^2}\left(-\frac{m^2}{M^2}+2\right)+\frac{D_{12}}{M^2}\left(\frac{2m^2}{M^2}-\frac{10}{3}\right)\right.\\&&
\left.+\frac{D_1}{M^2}\frac{13m^2}{3}-\frac{D_2}{M^2}\frac{5m^2}{3}+\left(-\frac{8m^2}{3}+\frac{2M^2}{3}\right)R\right]\\
I_{15}&=&\left[\frac{5m}{3M^3}D_{11}-\frac{m}{3M^3}D_{22}-\frac{4m}{3M^3}D_{12}+\frac{D_2}{M}\left(-\frac{2m^3}{M^2}+\frac{2m}{3}\right)D_2+\left(-\frac{2m^3}{M}+mM\right)R\right]\\
I_{23}&=&\left[\left(\frac{7m^3}{3M^5}+\frac{3m^2}{4M^4}+\frac{5m}{12M^3}\right)D_{11}+\left(\frac{7m^3}{3M^5}-\frac{3m^2}{6M^4}\right)D_{22}\right.\\&&
+\left(-\frac{14m^3}{3M^5}+\frac{m}{3M^3}-\frac{3m^2}{4M^4}\right)D_{12}\\&&
+\left(-\frac{7m^3}{6M^3}+\frac{m}{6M}-\frac{3m^4}{8M^4}+\frac{3m^2}{8M^2}\right)D_1+D_2\left(\frac{7m^3}{6M^3}-\frac{m}{M}+\frac{3m^4}{8M^4}\right)\\&&
\left.+\left(\frac{5m^3}{6M}-\frac{mM}{4}+\frac{3m^4}{16M^2}\right)R\right];\\
I_{24}&=&\left[\frac{13m^3}{M}R+\frac{D_{11}}{M^3}\left(-\frac{3m^3}{2M^2}+\frac{7m}{3}\right)-\frac{D_{22}}{M^3}\left(\frac{m^3}{3M^2}+6m\right)+\frac{D_{12}}{M^3}\left(\frac{3m^3}{M^2}+\frac{11m}{3}\right)\right.\\&&
\left.+\frac{D_1}{3M}\left(-\frac{25m^3}{M^2}-5m\right)+\frac{D_2}{3M}\left(\frac{25m^3}{M^2}-19m\right)\right]\\
I_{25}&=&\left[\frac{D_{11}}{M^2}\left(-\frac{3m^2}{2M^2}+\frac{4m^4}{M^4}+1\right)+\frac{D_{22}}{M^2}\left(-\frac{7m^2}{2M^2}-\frac{4m^4}{3M^4}+\frac{10}{3}\right)\right.\\&&
+\frac{D_1}{M^2}\left(\frac{7M^2}{6}+\frac{m^4}{M^2}-10\frac{4m^2}{3}-\frac{8m^6}{3M^4}\right)+\frac{D_2}{M^2}\left(\frac{8M^2}{3}-\frac{m^4}{3M^2}+\frac{8m^6}{3M^4}-8m^2\right)\\&&
\left.+\frac{D_1D_2}{M^2}\left(\frac{5m^2}{M^2}-\frac{8m^4}{3M^4}+2\right)+\left(-\frac{23m^2}{6}+\frac{19m^4}{3M^2}+\frac{8m^6}{3M^4}+\frac{2M^2}{3}\right)R\right]\\
I_{34}&=&\left[-\frac{m^3}{M^5}\left(D_{11}+D_{22}\right)+\frac{4m^3}{M^5}D_{12}\right]\\
I_{35}&=&\left[\frac{D_{11}}{M^2}\left(\frac{m^4}{M^4}+\frac{28m^2}{M^2}-\frac{16m^4}{M^4}-\frac{14}{3}\right)+\frac{D_{22}}{M^2}\left(\frac{m^2}{M^2}-\frac{8m^4}{3M^4}+\frac{16m^3}{M^3}-\frac{2}{3}\right)\right.\\&&
+\frac{D_1}{M^2}\left(\frac{2M^2}{3}-\frac{10m^3}{M}-\frac{2m^4}{3M^2}+\frac{16m^6}{3M^4}+\frac{22m^2}{3}\right)+\frac{D_2}{M^2}\left(2M^2-\frac{12m^2}{3}+\frac{26m^4}{3M^2}-\frac{2m^4}{3M^2}-\frac{16m^6}{3M^4}\right)\\&&
\left.+\frac{D_{12}}{M^2}\left(-\frac{2m^4}{3M^4}+\frac{50m^2}{3M^2}+\frac{4}{3}\right)+\frac{R}{M^2}\left(-2m^4-\frac{8M^4}{3}-\frac{14m^6}{3M^2}-\frac{10m^2M^2}{3}\right)\right]\\
I_{45}&=&\left[\frac{D_{11}}{M^3}\left(-\frac{8m^3}{M^2}-\frac{8m}{3}\right)+\frac{D_{22}}{M^3}\left(-\frac{8m^3}{M^2}+\frac{16m}{3}\right)+\frac{D_{12}}{M^3}\left(\frac{16m^3}{M^2}-\frac{8m}{3}\right)\right.\\&&
\left.\frac{D_1}{M^2}\left(\frac{8m^3}{M^2}-\frac{4m}{3}\right)+\frac{D_2}{M}\left(-\frac{8m^3}{M^2}+\frac{4m}{3}\right)\right]
\label{eq:I20}\end{aligned}$$ $$\phantom{1=1}$$
Here $R,D_1,D_2,D_{11},D_{12},D_{22}$ are the analytic results of contour integrations over the off-shell parameter $d\sigma$ in the complex $\sigma$-plane. The results of these integrals are given as: $$\begin{aligned}
\nonumber D_1&=&\int \frac{Md\sigma}{\Delta_1^2\Delta_2}\Delta_1=2\pi i\;\frac{1}{D(\widehat{q})}\\
\nonumber D_2&=&\int \frac{Md\sigma}{\Delta_1^2\Delta_2}\Delta_2=2\pi i\;\frac{2}{(2\omega)^3}\\
\nonumber D_{12}&=&\int \frac{Md\sigma}{\Delta_1^2\Delta_2}\Delta_1\Delta_2=2\pi i\;\frac{1}{2\omega}\\
\nonumber D_{11}&=&\int \frac{Md\sigma}{\Delta_1^2\Delta_2}\Delta_1^2=2\pi i\;\frac{1}{2\omega}\\
\nonumber D_{22}&=&\int \frac{Md\sigma}{\Delta_1^2\Delta_2}\Delta_2^2=2\pi i\;\frac{\displaystyle\omega^2-\frac{M^2}{2}}{\omega^3}\\
R&=&\int \frac{Md\sigma}{\Delta_1^2\Delta_2}=2\pi i\;\frac{M^2-12\omega^2}{4\omega^3(M^2-4\omega^2)^2}
\label{eq:theSi}\end{aligned}$$ Final result for the BS normalizer has the form, $$\begin{aligned}
N_V^{-2} &=&\frac{\pi^{5/2}}{72M^7\beta^3}e^{m^2/(2\beta^2)}\left\{
[G_{13}(m,M,A) + G_{11}(m,M,A)\beta^2 +
G_9(m,M,A)\beta^4]K_0\left(\frac{m^2}{2\beta^2}\right)\right.\nonumber\\
&&+[G_{13}(m,M,A) + G_{11}(m,M,A)\beta^2 +
G_9(m,M,A)\beta^4 + G_7(m,M,A)\beta^6]K_1\left(\frac{m^2}{2\beta^2}\right)\nonumber\\
&&+\left[H_5(m,M)\beta^6U\left(\frac{1}{2},-3,\frac{m^2}{\beta^2}\right)\right.\nonumber\\
&&+H_7(m,M)\beta^4U\left(\frac{1}{2},-2,\frac{m^2}{\beta^2}\right)\nonumber\\
&&\left.\left.+H_{11}(m,M)U\left(\frac{1}{2},0,\frac{m^2}{\beta^2}\right)\right]\beta^2A_3A_5\right\}
\label{eq:formNVm2}.\end{aligned}$$ Here, $K_n(x)$ is the second class modified Bessel function, $U(a,b,x)$ is the confluent hypergeometric function, $G_n$ and $H_n$ are polynomials of $n$-th degree in $m$ and $M$, and the $G_n$ are quadratic functions of the $A_i$ coefficients.
In these expressions, $\phi(\hat{q})$ is a decaying function of $\hat{q}^2$. Thus despite the fact that the integrands contain growing factors like $\hat{q}^2$, the overall integrals converge, and can be analytically integrated. Then, the $f_V^{i}$ $(i=0,...,5)$ can be expressed in the following analytic form, $$\begin{aligned}
\nonumber f_V^0&=&\frac{A_0}{M}N_V16\sqrt{\frac{2}{3}}\pi^{3/4}\beta^{3/2}(2m^2+3\beta^2)\\
\nonumber f_V^1&=&-A_1N_V8\sqrt{6}\pi^{3/4}\beta^{3/2}m\\
\nonumber f_V^2&=&0\\
\nonumber f_V^3&=&\frac{A_3}{M^2}N_V\frac{2\pi^{1/4}}{\sqrt{3}}\sqrt{\beta}
\left\{3M\sqrt{2\pi}\beta(-4m^2+M^2-28\beta^2)\phantom{\frac{2}{4}}\right.\\&&
\nonumber \left.+2e^{m^2/(4\beta^2)}m^2\left[2m^2K_0\left(\frac{m^2}{4\beta^2}\right)
\ +(2m^2-M^2+16\beta^2)K_1\left(\frac{m^2}{4\beta^2}\right)
\right]\right\}\\
\nonumber f_V^4&=&-\frac{A_4}{M^3}N_V\frac{4\pi^{1/4}}{\sqrt{3}}\sqrt{\beta}
\left\{m^2M^2
e^{m^2/(4\beta^2)}K_1\left(\frac{m^2}{4\beta^2}\right)
+2\sqrt{\pi}\beta^3\left[3\sqrt{2}M-8\beta U\left(-\frac{3}{2},-2,\frac{m^2}{2\beta^2}\right)
\right]\right\}\\
f_V^5&=&0
\label{eq:formfv}\end{aligned}$$ with $N_{V}$ given in the previous equation.
Results {#sec:results}
=======
Numerical Calculation {#subsec:numerical}
---------------------
Eq. (3.18) which expresses decay constants $f_V$ of vector mesons in terms of the parameters $A\equiv($$A_0$, $A_1$, $A_2$, $A_3$, $A_4$, $A_5)$ is a non linear function of the $A_i$’s. Then numerical methods must be applied to solve the problem of determining the best values of the $A_i$’s.
We used a simple ‘Mathematica’ procedure for searching for accurate values of the $A_i$ ($i=0,...5$). We defined the following auxiliary function $W(A)$, which is positive definite, as, $$W(A)=\sum\limits_V\left[f_V(A)-f_V^{EXP}\right]^2.
\label{eq:3.14}$$ The summation in the above equation runs over the five ground state vector mesons $\rho$, $\omega$, $\phi$, $\psi$ and $Y$ studied in this work. $f_V^{EXP}$ are the central values of experimental data on decay constants used (indicated in Table 2) which are calculable from the data on the partial widths, $\Gamma(e^++e^-)$ for the five studied mesons in [@beringen12]. Their results are the following: $\rho$(770): 7.04$\pm$0.06 keV, $\omega$(782): 0.60$\pm$0.02 keV, $\phi$(1020): 1.27$\pm$0.04 keV, $\psi$(1S): 5.5$\pm$0.1 keV, $Y$(1S): 1.34$\pm$0.02 keV. They are related to our decay constants $f_V$ by formula, $$\Gamma= \frac{4\pi\alpha^{2}e_Q^{2}|f_{V}|^{2}}{3M}.
\label{eq:width2decay}$$ Here, $M$ is the meson mass, $\alpha$ is the QED coupling constant ([*i.e.*]{} the fine structure constant), $e_Q$ plays the role of effective electric charge of the meson and has values as listed after Eq.(3.2)for different vector mesons.
From formula (\[eq:width2decay\]) were obtained the data cited in Table 1. We see that the error bars of experimental data on decay widths given by PDG tables for the $\rho$, $\omega$, $\phi$, $\psi$ and $Y$ mesons represent, 0.8%, 3.3%, 3.1%, 1.8%, 1.5%, respectively. We can say that average error bars are 2.1% or 0.05 keV. The average error in decay constants derived from experimental data of the five mesons studied is 1%. As a general rule, while we are averaging data coming from different measurements, we should give larger weight to more precise data and lesser weight to less precise data. For instance, $\rho$’s experimental error bar of $f_V$ is 0.4%, while for $\omega$ meson error bar of $f_V$ is 1.6%; [*i.e.*]{} $\rho$ has a higher precision than $\omega$ meson. However, by using the $W(A)$ function, we have fit the data for all mesons from $\rho$ to $Y$ at the same level. This means that we can not expect a fitting with an average error significantly less than $\sim1.6$%.
From the numerical point of view, the problem reduces to finding values of $A_i$’s such that $W(A)$ has a minimum, and that at such a minimum it takes the value zero. We used the Mathematica package which has some useful functions for minimizing. It is clear that the 6-dimensional hypersurface $W(A)$ has many minima, but the only acceptable minima are those minima for which $W(A)$ is very close to zero. The experimental values of the $f_V$’s are 0.220, 0.195, 0.228, 0.410 and 0.708 GeV. Then, we adopt as criterium of “sufficient closeness” to zero the value of $W(A)$ which is less than $0.02^2=0.0004$ GeV$^2$.
Besides, it is important to ensure a degree of robustness of the solution. It means, that if we change each of the experimental values $f_V^{EXP}$ according to their error bars, then the values of the $A_i$ which minimize $W(A)$ remain near the corresponding values which minimize at the central values $f_V^{EXP}$. This is the main criterion of stability which must be satisfied for the model to be physically acceptable. In other words, when point $(A)$ is within a box determined by the averages and the error bars of the $A_i$, the function $W(A)$ has a minimum very near to zero for $f_V$ in a box determined by the experimental data of the $f_V$ and their error bars.
An additional check is done by evaluating the percent average of the absolute values of the differences between the predicted $f_V$ values from the experimental value $f_V^{EXP}$.
Using this method, we found that the values of coefficients should respectively be: $A_0=1$, $A_1=0.006773$, $A_2=1.24011$, $A_3=-0.414747$, $A_4=0.013611$, $A_5=-1.84191$ to predict the decay constant values, $f_\rho=0.207440$ GeV, $f_\omega=0.206914$ GeV, $f_\phi=0.230219$ GeV, $f_\psi=0.415707$ GeV, and $f_Y=0.758994$ GeV. These decay constant values have an average error with respect to the experimental data of 4%.
The robustness of the model can be quantified in the following way. The set of $A_i$ which, when replaced in $f_V(A)$ give a value “near” the experimental value of the decay constant for each of the studied vector mesons. We found that point $(A)$ is located within certain “box” of sides $2\Delta A_i=0.008$. The center of the box is at the values given in the last paragraph. All those results were obtained by randomly choosing 22 sets of the $f_V$ within their experimental error bars and finding in each case the point $A$ where $W(A)$ has a minimum near to zero. Fig. 2 represents the $f_V$ amplitudes for all five studied mesons, sets of dots were obtained by varying the $A_i$ coefficients around their average values, procedure was done to show stability of results.
The normalization factors $N_V$ were found to be: $N_\rho=0.18707795$ GeV$^{-3}$, $N_\omega=0.18624605$ GeV$^{-3}$, $N_\phi=0.13413963$ GeV$^{-3}$, $N_\psi=0.03270152$ GeV$^{-3}$, and $N_Y=0.00495780$ GeV$^{-3}$. Values of $f_V$ along with the contributions from various covariants and experimental results are listed in Table 2. Comparison of our results with those of other models and data is presented in Table 3.
The results {#subsec:results}
-----------
Formulas found in section \[sec:theory\] express decay constants $f_V$ of vector mesons in terms of the constant parameters $A_0$, $A_1$, $A_2$, $A_3$, $A_4$, $A_5$. Our model should be capable of predicting the values of those parameters if one uses the known experimental values of the decay constants for the $\rho$, $\omega$, $\phi$, $\psi$, $\psi'$ and $Y$ mesons. Expression for $f_V$ in Eq. (3.12) and Eq. (3.18) is a linear function of the $A_i$’s (for i=0,1,2,3,4,5). However $f_{V}$ involves the BS normalizer $N_V$, which is evaluated by integrating with respect to $\hat{q}$, is a highly non-linear function of the $A_i$’s. Analytical expressions for the $A_i$’s as functions of quark masses and other parameters corresponding to each of the vector mesons can not be obtained. However, numerical methods give acceptable solutions of the problem.
Expression for normalizer (\[eq:formNVm2\]) has the form, $$\frac{1}{N_V^2} = \sum\limits_{i=0}^5\sum\limits_{j=i}^5I_{ij}A_iA_j,
\label{eq:quadraticNv}$$ where matrix elements $I_{ij}$ were given in Eq. (\[eq:I20\]).
Contributions $f_V^i$ to the decay constants, given by Eq. ($\ref{eq:formfv}$), by definition are proportional to $A_i$, $$f_V = \sum\limits_{i=0}^5f_V^i \equiv \sum\limits_{i=0}^5\frac{f_{V_i}}{N_V}A_i.$$
Our idea is that the $A_i$ can be obtained by fitting formulas to experimental results, $$\frac{\sum\limits_{i=0}^5f_{V_i}A_i}{\sqrt{\sum\limits_{i=0}^5
\sum\limits_{j=i}^5I_{ij}A_iA_j}} = f_V^{EXP}.
\label{eq:equationAi}$$
(see Eq. (\[eq:Nvm2\]) and Eq. (\[eq:I20\])). We must notice that \[eq:equationAi\] is a homogenous function of the $A_i$, fact that precludes finding solutions for the $A_0,... A_5$ by using available data of the five considered mesons. However, a nonhomogenous system of equations can be constructed by putting $A_0=1$ and leaving as unknowns $A_1,... A_5$. With this procedure it is sufficient to consider the five mesons $\rho$, $\omega$, $\phi$, $\psi$, and $Y$. System of five equations is obtained by introducing in Eq. (\[eq:equationAi\]) the appropriate parameters of the five mesons.
For the $f_V^{EXP}$ we used the results from PDG tables [@beringen12]. We see that the “error bars” on data for $f_{V}$ for $\rho,\omega,\phi,\psi$ and $Y$ mesons are 0.4%, 1.7%, 1.6%, 0.9% and 0.8% respectively, whose average is about 1.1%.
System of five equations has many solutions, several of them complex. Complex solutions appear in conjugate pairs. Available algorithms allow finding all solutions with very high precision. In this way the percent average of the absolute values of the difference between the calculated $f_V$ and the $f_V^{EXP.}$ experimentally found, is much lower than the error bars of experiments. However the values of the $A_i$ found were complex, which lead to complex $f_V$’s. We made different checks and selected the parameter set giving the results shown in Table 2 which predicts the experimental values of all the $f_V$’s with less precision. The $f_V$’s predicted for the considered mesons match approximately with the experimental values. Regarding meson $\psi'$, our theoretical value is 0.1798 GeV, while value calculated from experiments is 0.2200 GeV, with error 18%.
Meson $m$ (GeV) $M$ (GeV) $e_Q$ $\beta$ $\Gamma^{EXP.}$ (keV) $f_V^{EXP.}$ (GeV)
--------------- ----------- --------------------- --------------- ---------- ----------------------- --------------------
$\rho(770)$ 0.265 0.7755$\pm$0.0003 $1/\sqrt{2}$ 0.265294 7.04$\pm$0.06 0.2201$\pm0.0009$
$\omega(782)$ 0.265 0.7827$\pm$0.0001 $1/\sqrt{18}$ 0.2661 0.60$\pm$0.02 0.195$\pm0.003$
$\phi(1020)$ 0.415 1.01946$\pm$0.00002 $1/3$ 0.293347 1.27$\pm$0.04 0.228$\pm0.003$
$J/\psi(1S)$ 1.532 3.09692$\pm$0.00001 $2/3$ 0.444214 5.5$\pm$0.1 0.410$\pm0.003$
$Y(1S)$ 4.9 9.4603$\pm$0.0003 $1/3$ 0.721345 1.34$\pm$0.02 0.708$\pm0.005$
$f_V^0$ $f_V^1$ $f_V^3$ $f_V^4$ $f_V^{LO}$ $f_V^{NLO}$ $f_V^{LO}$(%) $f_V^{NLO}$(%) $\bm{f_V^{TH.}}$ $\bm{f_V^{EXP}}$
---------- --------- ----------- --------- ----------- ------------ ------------- --------------- ---------------- ------------------ ------------------
$\rho$ 0.1156 -0.00068 0.09336 -0.00026 0.1149 0.0931 55% 45% **0.2080** **0.2200**
$\omega$ 0.1155 -0.000689 0.093 -0.000258 0.1148 0.0927 56% 44% **0.2075** **0.1952**
$\phi$ 0.1461 -0.00104 0.086 -0.000288 0.1450 0.0859 63% 37% **0.2302** **0.2285**
$\psi$ 0.352 -0.00321 0.06257 -0.000254 0.3487 0.06232 84.8% 15.2% **0.411** **0.4104**
$Y$ 0.6617 -0.00628 0.05268 -0.000224 0.6555 0.0525 92.6% 7.4% **0.7079** **0.7080**
$\Gamma^{TH.}$ (keV) $\Gamma^{EXP.}$ (keV)
------------------ ---------------------- -----------------------
$\rho(770)$ 8.952 7.04$\pm$0.06
$\omega(782)$ 0.642 0.60$\pm$0.02
$\phi(1020)$ 1.294 1.27$\pm$0.04
$\psi(1S)(3096)$ 5.414 5.5$\pm$0.1
$Y(1S)(9460)$ 1.345 1.34$\pm$0.02
$f_\rho$ $f_\omega$ $f_\phi$ $f_\psi$ $f_Y$
------------ ------------------- ----------------- ----------------- ----------------- -----------------
$BSE-CIA$ 0.208 0.207 0.230 0.411 0.7079
$BSE [4]$ 0.215 0.224
$SDE [2]$ 0.163
$SDE [7]$ 0.207 0.259
$BSE [8]$ 0.459 0.498
$Exp.[22]$ 0.2201$\pm$0.0009 0.195$\pm$0.003 0.228$\pm$0.003 0.410$\pm$0.003 0.708$\pm$0.005
{height="0.5\linewidth" width="0.8\linewidth"}
{height="0.4\linewidth" width="0.5\linewidth"}{height="0.4\linewidth" width="0.5\linewidth"} {height="0.4\linewidth" width="0.5\linewidth"}{height="0.4\linewidth" width="0.5\linewidth"} {height="0.4\linewidth" width="0.5\linewidth"}
Discussion {#sec:disc}
==========
In this paper we have calculated the decay constants $f_V$ of vector mesons $\rho$, $\omega$, $\phi$, $\psi$ and $Y$ in BSE under Covariant Instantaneous Ansatz (CIA) using Hadron-quark vertex function which incorporates various Dirac covariants order-by-order in powers of inverse of meson mass within its structure in accordance with a recently proposed power counting rule from their complete set. This power counting rule suggests that the maximum contribution to any meson observable should come from Dirac structures associated with Leading Order (LO) terms alone, followed by Dirac structures associated with Next-to-Leading Order (NLO) terms in the vertex function. Incorporation of all these covariants is found to bring calculated $f_V$ values much closer to results of experimental data [@beringen12] and some recent calculations [@alkofer02; @ivanov99; @close02; @li08; @cvetic04; @hwang97; @alkofer01; @wang08] for $\rho$, $\omega$, $\phi$, $\psi$ and $Y$ mesons. The $f_V$ predicted are within the error bars of experimental data for each one of these five mesons.
The results for $\rho$, $\omega$, $\phi$, $\psi$ and $Y$ mesons with parameter set: $A_0 = 1$, $A_1 = 0.007\pm0.001$, $A_2 = 1.240\pm0.001$, $A_3 = -0.415\pm0.001$, $A_4 = 0.014\pm0.001$, $A_5 = -1.842\pm0.001$ (giving $f_V$ values with average error with respect to experimental data of 4%) are presented in Table 2. Comparison with experimental data and other models is shown in Table 2.
In Fig. 3 we are plotting vs $\hat{q}$ the integrands of $f^{LO}_V$, $f^{NLO}_V$ and $f_V$ for each of studied mesons. Those plots show that the contribution to $f_V$ from NLO covariants is smaller than the contribution from LO covariants for $\rho$, $\omega$, $\phi$, $\psi$ and $Y$ mesons. And for $\psi$ and $Y$ mesons, NLO contribution is negligible in comparison to LO contribution. Then, it is concluded from Table 2 that as far as the various contributions to decay constants $f_V$ are concerned, for $\rho$ and $\omega$ mesons, the LO terms contribute only 55%, while NLO terms 45%. However as one goes to $\phi$ meson, the LO contribution increases to 62.8%, while NLO contribution is 37.2%. But as one goes to heavy ($c\overline{c}$ and $b\overline{b}$) mesons, for $\psi$ meson, LO contribution is 84.8%, while NLO contribution is 15.2%, and for $Y$ meson the LO contribution is 92.6%, while NLO contribution reduces to just 7.4%. Thus the drop in contribution to decay constants from NLO covariants vis-a-vis LO covariants is more pronounced for heavy mesons $\psi$ and $Y$. And among the two LO covariants, it can be seen that the most leading covariant $i\gamma_{\mu}$ contributes the maximum for all vector mesons from $\rho$ to $Y$. These results on decay constants $f_{V}$ for vector mesons are completely in conformity with the corresponding results on decay constants $f_{P}$ for pseudoscalar mesons, $K,D,D_{S}$ and $B$ done recently[@bhatnagar11] where it was also noticed that the NLO contribution is much smaller than the LO contribution for hevier mesons like $D,D_{S}$ and $B$, where the contribution drops from 10% to 4%, and the most leading covariant was found to be $\gamma_{5}$.
This is in conformity with the power counting rule according to which the leading order covariants, $\gamma_5$ and $i\gamma_5(\gamma\cdot
P)(1/M)$ (associated with coefficients $A_0$ and $A_1$) should contribute maximum to decay constants followed by the next-to-leading order covariants, $-i\gamma_5(\gamma\cdot q)(1/M)$ and $-\gamma_5[(\gamma\cdot P)(\gamma\cdot q)-(\gamma\cdot q)(\gamma\cdot
P)](1/M^2)$ (associated with coefficients $A_2 $ and $A_3$) in the BS wave function.
We observe in Fig. 2 that though LO and NLO Dirac covariants are sufficient to correctly predict amplitudes for $\phi$ and $\psi_0$ vector mesons, but only LO covariants are sufficient for $Y$ meson. But for $\rho$ and $\omega$ mesons, the LO and NLO Dirac covariants are not sufficient to predict accurately their amplitudes and is thus necessary to include even higher order NNLO Dirac covariants in their hadron-quark vertex functions. This can also be seen from Fig. 3.
However, the numerical results for $f_V$ for equal mass vector mesons, obtained in our framework with use of leading order Dirac covariants along with the next to leading order Dirac covariants, along with a similar calculation for $f_{P}$ done recently [@bhatnagar11] for pseudoscalar mesons demonstrates the validity of our power counting rule, which also provides a practical means of incorporating various Dirac covariants in the BS wave function of a hadron. By this rule, we also get to understand the relative importance of various covariants to calculation of meson observables. This would in turn help in obtaining a better understanding of the hadron structure.
Acknowledgments {#acknowledgments .unnumbered}
===============
One of the authors (SB) is thankful to Prof. S-Y. Li, Shandong University, China for discussions. JM thanks the support from Programa de Sostenibilidad University of Antioquia.
[99]{}
R.Alkofer, P.Watson, H.Weigel, Phys. Rev. D[**65**]{}, 094026 (2002). M.A.Ivanov, Yu A. Kalinovski, C.D. Roberts, Phys. Rev. D[**60**]{}, 034018 (1999). F.E.Close, A.Donnachi, Yu S. Kalashnikova, Phys. Rev. D[**65**]{}, 092003 (2002). H-M.Li, S-L.Wan, Chin. Phys. Lett. [**25**]{}, 1239 (2008). G.Cvetic et al., Phys. Lett. B[**596**]{}, 84 (2004). D.S.Hwang, G.H.Kim, Phys. Rev. D[**55**]{}, 6944 (1997). R.Alkofer, L.W.Smekel, Phys. Rep. [**353**]{}, 281 (2001). G.L.Wang, Intl. J. Mod. Phys. A[**23**]{}, 3263 (2008). A. N. Mitra, B. M. Sodermark, Nucl. Phys. A[**695**]{}, 328 (2001) and references therein. A.N.Mitra, PINSA [**65**]{}, 527 (1999). S. Bhatnagar, D. S. Kulshreshtha, A. N. Mitra, Phys. Lett. B[**263**]{}, 485 (1991). A. N. Mitra, S. Bhatnagar, Intl. J. Mod. Phys. A[**7**]{}, 121 (1992). S.Bhatnagar, S-Y. Li, J. Phys. G[**32**]{}, 949 (2006) S. Bhatnagar, S-Y.Li, J. Mahecha, Intl. J. Mod. Phys. E[**20**]{},1437 (2011). C.H.L.Smith, Ann. Phys. [**53**]{}, 521 (1969). F.T.Hawes, M.A.Pichowsky, arxiv:nucl-th/9806025, 1998. E.Mengesha, S.Bhatnagar, Intl. J. Mod. Phys. E[**20**]{}, 2521 (2011). S.Bhatnagar, Intl. J. Mod. Phys. E14, 909 (2005). K.K.Gupta et al., Phys. Rev. D42, 1604 (1990). S.Bhatnagar and S.Y.Li, *In the Proceedings of 9th Workshop on Non-Perturbative Quantum Chromodynamics, Paris, France, 4-8 Jun 2007, pp 12*. D.Arndt, C-R.Ji, *Light-Cone Quark Model Analysis of Radially Excited Pseudoscalar and Vector Mesons*. arXiv:hep-ph/9905360v1, 1999. J.Beringen et al., (Particle Data Group), Phys. Rev. D[**86**]{}, 010001 (2012).
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
author:
- |
G. Iakovidis$^{a,b}$[^1] \
National Technical University of Athens\
Zografou Campus, GR15773, Athens, Greece\
Brookhaven National Laboratory\
Upton, NY 11973, USA\
E-mail:
title: The Micromegas Project for the ATLAS Upgrade
---
=1
Introduction {#sec:intro}
============
The Large Hadron Collider (LHC) at CERN, after the scheduled shutdown of 2017-2018, will resume its operation with a luminosity increase of five times its original design luminosity of $\mathcal{L}=10^{34}\,\mathrm{cm^{-2}s^{-1}}$. For the ATLAS detector (ref. [@cite:atlastdr]), such a luminosity increase means higher particle rates. While in most of the ATLAS muon system the detectors have enough safety margin to handle these rates, the first forward station of the muon spectrometer, called the Small Wheel, will exceed its design capabilities. At pseudorapidity $\eta = \pm2.7$, rates up to $15\,\mathrm{kHz/cm^{2}}$ are expected, far higher than what the currently installed detectors can handle. Furthermore, the upgraded Small Wheel is expected to take part in the Level 1 trigger decision, something that the present system was not designed for. The physics objective is to sharpen the trigger threshold turn-on as well as discriminate against background while maintaining the low transverse momentum ($p_{\mathrm{T}}$) threshold for single leptons ($e$ and $\mu$) and keeping the Level-1 rate at a manageable level.
The Muon ATLAS MicroMegas Activity (MAMMA) R&D explored the potential of the micromegas technology for its use in LHC detectors and finally proposed to equip the New Small Wheel (NSW) with micromegas detectors, combining trigger and precision tracking functionality in a single device. In middle 2013 the ATLAS Collaboration endorsed the proposal (ref. [@cite:tdr]). In total, eight planes of micromegas (ref. [@cite:Giomataris]) detectors covering the full NSW will be installed, corresponding to a total detector area of 1200$\mathrm{m^2}$. In addition to the micromegas, the NSW should also be equipped with eight planes of thin-gap multiwire detectors, called sTGC, such as to create a fully redundant system, both for trigger and tracking.
Conclusions {#sec:conclusions}
===========
The extensive MAMMA R&D program has transformed micromegas detectors spark resistant allowing them to operate in high energy experiments like ATLAS. Construction techniques developed, allowed us to build large area detectors of 2m$^2$ while involving industry. The first generation of the VMM ASIC unveiled the capabilities of micromegas as a trigger and tracking detector. All the above achievements brought micromegas detectors ahead of the R&D phase. A full sector identical to a detector that can be installed to ATLAS will be build in 2014 featuring the next generation of the VMM ASIC providing trigger and tracking information. Following the ATLAS schedule, 1200m$^2$ of micromegas detectors will be constructed and assembled on 2015–2016 while the installation and commission of the full system will follow on 2017–2018.
[^1]: On behalf of the MAMMA collaboration.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'We investigate departures from LTE in the line formation of Fe for a number of well-studied late-type stars in different evolutionary stages. A new model of Fe atom was constructed from the most up-to-date theoretical and experimental atomic data available so far. Non-local thermodynamic equilibrium (NLTE) line formation calculations for Fe were performed using 1D hydrostatic MARCS and MAFAGS-OS model atmospheres, as well as the spatial and temporal average stratifications from full 3D hydrodynamical simulations of stellar convection computed using the Stagger code. It is shown that the Fe I/Fe II ionization balance can be well established with the 1D and mean 3D models under NLTE including calibrated inelastic collisions with H I calculated from the Drawin’s (1969) formulae. Strong low-excitation Fe I lines are very sensitive to the atmospheric structure; classical 1D models fail to provide consistent excitation balance, particularly so for cool metal-poor stars. A better agreement between Fe I lines spanning a range of excitation potentials is obtained with the mean 3D models. Mean NLTE metallicities determined for the standard stars using the 1D and mean 3D models are fully consistent. Also, the NLTE spectroscopic effective temperatures and gravities from ionization balance agree with that determined by other methods, e.g., infrared flux method and parallaxes, if one of the stellar parameters is constrained independently.'
author:
- |
Maria Bergemann$^1$[^1], K. Lind$^1$, R. Collet$^{2,3,1}$, Z. Magic$^1$, M. Asplund$^{4,1}$\
$^1$ Max Planck Institute for Astrophysics, Karl-Schwarzschild Str. 1, 85741, Garching, Germany\
$^2$ Natural History Museum of Denmark, Centre for Star and Planet Formation, [Ø]{}ster Voldgade 5-7, DK–1350 Copenhagen, Denmark\
$^3$ Astronomical Observatory / Niels Bohr Institute, Juliane Maries Vej 30, DK–2100 Copenhagen, Denmark\
$^4$ Australian National University, Ellery Crescent Acton ACT 2601, Australia
bibliography:
- 'references.bib'
date: 'Accepted Date. Received Date; in original Date'
title: 'NLTE line formation of Fe for late-type stars. I. Standard stars with 1D and $\td$ model atmospheres'
---
\[firstpage\]
Atomic data – Line: profiles – Line: formation – Stars: abundances
Introduction
============
Accurate determination of basic stellar parameters is fundamental for calculations of chemical composition, ages, and evolutionary stages of stars. One of the most commonly used methods to determine effective temperature $\Teff$, surface gravity $\logg$ and metallicity \[Fe/H\], is to exploit excitation-ionization equilibria of various chemical elements in stellar atmospheres. Iron, with its partly filled $3d$ subshell, has, by far, the largest number of lines all over the spectrum of a typical late-type star. This atomic property coupled to a relatively large abundance makes it a reference element for spectroscopic estimates of stellar parameters.
The goal of this work is to study systematic uncertainties in this method, which are related to 1) the assumption of local thermodynamic equilibrium (LTE) in spectral line formation calculations, and 2) the use of theoretical 1D hydrostatic model atmospheres. These two approximations are inherent in most of the line formation codes utilized in spectroscopic studies, since they strongly reduce the complexity of the problem permitting the analysis of very large stellar samples in short timescales. Yet, in conditions when the breakdown of LTE/1D modelling occurs the inferred stellar parameters may suffer from large systematic biases (e.g., Asplund 2005). To assess the extent of the latter, more physically realistic modeling is necessary.
Non-local thermodynamic equilibrium (NLTE) effects on the / level populations for FGK stars have been extensively discussed in the literature (@1972ApJ...176..809A; @1980ApJ...241..374C; @1990SvAL...16...91B; @1999ApJ...521..753T; ; ; Collet et al. 2005; Mashonkina et al. 2011). These and other studies showed that NLTE effects in the ionization balance of / are large for giants and metal-poor stars. The effect on solar-metallicity stars is smaller, but it is must be taken into account if one aims at the accuracy of few percent, as is the case for the Sun. Despite major efforts aimed at understanding how non-equilibrium thermodynamics affects the line formation of Fe, there have been only few attempts to quantify these deviations in a systematic manner across the Hertzsprung-Russell diagram. @2011mast.conf..314M provided a small grid of NLTE corrections to five lines for the solar-metallicity stars with $\Teff > 6500$ K and $\logg >3$. @1999ApJ...521..753T explored a larger range of stellar parameters including FGK stars down to \[Fe/H\] $\approx
-3$, however, they did not present a regular grid of NLTE corrections.
Furthermore, it has only recently become possible to perform full time-dependent, 3D, hydrodynamical simulations of radiative and convective energy transport in stellar atmospheres. Such simulations (Nordlund & Drawins 1990, Asplund et al. 1999, Collet et al. 2006, Ludwig et al. 2006, @2009LRSP....6....2N, Freytag et al. 2010) have evidenced important shortcomings of 1D, stationary, hydrostatic models. Especially at low metallicity, it has been realized that 1D models, by necessarily enforcing radiative equilibrium, overestimate the average temperatures of shallow atmospheric layers, with profound implications for the spectral line formation ().
In this work, we perform NLTE line formation calculations for and using 1D hydrostatic and mean 3D model atmospheres obtained from temporal and spatial averaging of 3D hydrodynamical simulations (hereafter, $\td$). A new model of Fe atom is constructed from the most up-to-date theoretical and experimental atomic data available so far. Non-local thermodynamic equilibrium Fe abundances, effective temperatures, and surface gravities are derived for the Sun, Procyon, and four metal-poor stars. The efficiency of thermalization caused by inelastic collisions is calibrated as to satisfy ionization equilibrium by scaling the classical Drawin (1968, 1969) formulae. In the next paper in the series [@Paper2 hereafter Paper II], we discuss NLTE effects on and over a wide range of stellar parameters. That paper also presents a large grid of 1D NLTE abundance corrections for a wealth of lines in metal-rich and metal-poor dwarf and giant spectra.
Before proceeding with the description of the methods, we shall point out a few important aspects of our study. First, due to a comparative nature of the analysis (1D LTE vs. 1D NLTE and $\td$ NLTE) no attempt is made to fine-tune various parameters in order to achieve full agreement with other results in the literature. Second, although by the use of the averaged 3D models we roughly account for hydrodynamic cooling associated with convective overshoot in the simulations, the effect of horizontal inhomogeneities is not addressed because such calculations with our new realistic extensive model atom are beyond current computational capabilities. We note, however, that our results are expected to closely resemble any future full 3D NLTE calculations once these become feasible. The reason is that in NLTE the line formation is largely dictated by the radiation field (as explained in detail below) originating in deep atmospheric layers, where the significance of the atmospheric inhomogeneities is greatly reduced. Detailed tests of the $\td$ models, which involve comparison with other observable quantities, will be presented elsewhere (Collet et al., in prep.). Full 3D NLTE calculations with a simpler model atom of Fe will be a subject of a forthcoming publication.
The paper is structured as follows. Section \[sec:methods\] gives a brief description of the input model atmospheres, model atom, and programs used to compute NLTE populations and line formation. The results of statistical equilibrium calculations are presented in Sec. \[sec:statec\]. The analysis of the solar spectrum along with the re-determination of the Fe abundances for the Sun and Procyon is described in detail in Sec. \[sec:solspec\]. Section \[sec:abstars\] presents and discusses metallicities, temperatures, and gravities obtained for the metal-poor stars. Finally, a comparison with stellar evolution calculations is given in Sec. \[sec:evolution\]. A short summary of the work and conclusions are given in Sec. \[sec:conclusions\].
Methods
=======
[\[sec:methods\]]{}
Model atmospheres
-----------------
[\[sec:atmos\]]{} In the present study, we employ both <span style="font-variant:small-caps;">mafags-os</span> (, , ) and <span style="font-variant:small-caps;">marcs</span> 1D LTE hydrostatic model atmospheres, as well as 3D hydrodynamical models computed with the Stagger code (Galsgaard & Nordlund 1995, Collet et al. 2011).
The <span style="font-variant:small-caps;">mafags-os</span> models for all investigated stars were kindly provided by F. Grupp. The models are plane-parallel with convective energy transport based on the turbulent convection model of @1991ApJ...370..295C. Compressible turbulence is accounted for using a mixing length, $\alpha_{\rm cm} = l/\rm{H_p}$[^2]$ = 0.82$. The reasons for this choice are discussed in . We note that this $\alpha_{\rm cm}$ is also consistent with the results of @1999ASPC..173..225F, who attempted to calibrate the mixing length parameter using the 2D radiation hydrodynamics simulations of stellar convection. Line absorption is computed by the method of opacity sampling including $\sim 86\,000$ wavelength points from extreme UV to far IR. Extensive line lists were extracted from Kurucz and VALD databases. The reference solar abundances were compiled from various literature sources, giving preference to NLTE determinations by the Munich group . In particular, the solar Fe abundance is set to $\absun = 7.50$.
The <span style="font-variant:small-caps;">marcs</span> models are based on opacity sampling with $\sim 10^4 - 10^5$ wavelength points using Kurucz and VALD linelists. Convection is included in the @1965ApJ...142..841H version of the mixing length theory with the mixing length parameter set to $\alpha_{\rm MLT} = 1.5$. A detailed description of the models can be found on the MARCS web-site[^3]. For the majority of the elements, the reference solar abundances are that of Asplund et al. (2005).
Three-dimensional radiation-hydrodynamic simulations of convection at the surface of the Sun and reference stars were computed using the Copenhagen Stagger code (@nordlund95). The physical domains of the simulations cover a representative portion of the stellar surface. They include the whole photosphere as well as the upper part of the convective layers, typically encompassing $12$ to $15$ pressure scale heights vertically. Horizontally, they extend over an area sufficiently large to host about ten granules at the surface. The simulations use realistic input physics, including state-of-the-art equation-of-state, opacities, and treatment of non-grey radiative transfer. The adopted reference solar chemical composition for the simulations is that of . For a more detailed description of the simulations’ setup, we refer to @2011JPhCS.328a2003C.
For the purpose of the current study, we computed spatial and temporal averages of these simulations over surfaces of equal optical depth $\opd$ at the chosen reference wavelength at $5000$ Å. The independent thermodynamic variables, gas density $\rho$ and internal energy per unit mass $\varepsilon$, as well as the temperature $T$ are first interpolated with cubic splines for all columns in the full 3D structure to the reference optical depth scale. The reference scale was constructed to cover the relevant range for line-formation calculations (${-5}{\la}~{\opd}~{\la}{2}$) evenly in $\opd$. For density and internal energy, a logarithmic interpolation is used. Other physical quantities, namely gas pressure $P_{\rm gas}$ and electron number density $n_{\rm el}$, are looked up from the simulations’ equation-of-state tables as a function of density and internal energy. Finally, mean $\td$ stratifications are constructed by averaging the various physical quantities pertinent to line-formation calculations, $\ln{\rho}$, $T$, $\ln{P_{\rm gas}}$, and $\ln{n_{\rm el}}$, on surfaces of equal optical depth $\opd$ and over time (i.e., over all simulation snapshots). We emphasize that no hydrostatic equilibrium was enforced after the averaging process. In the present application of mean $\td$ stratifications to line-formation calculations, averages of velocity fields were not considered, and line broadening associated with bulk gas flows and turbulence was accounted for by means of a classical, depth-independent micro-turbulence parameter (Sect. \[sec:methods\]) as in 1D models.
The temperature and pressure stratifications from the 1D <span style="font-variant:small-caps;">marcs</span> model atmospheres and the mean $\td$ hydrodynamical models of the Sun and three metal-poor standard stars are shown in Fig. \[atm\_abs\] as a function of continuum optical depth at $5000$ [Å]{}.The input parameters are the ones listed in Table \[tab:init\_param\]. The <span style="font-variant:small-caps;">mafags-os</span> models are not included in the plots, because they adopt slightly different stellar parameters. Due to the similar equation-of-state, <span style="font-variant:small-caps;">marcs</span> and <span style="font-variant:small-caps;">mafags-os</span> model atmospheres are essentially identical in the outer layers. The differences between 1D and $\td$ models are more pronounced, especially in the case of metal-poor stellar atmospheres. At low metallicities, 3D stellar surface convection simulations predict cooler upper photospheric stratifications (, ), than corresponding classical, hydrostatic, stationary 1D model atmospheres generated for the same stellar parameters. The temperature in the outer layers of time-dependent, 3D, hydrodynamical simulations is mainly regulated by two mechanisms: radiative heating due to reabsorption of continuum radiation by spectral lines and adiabatic cooling associated with expanding gas above granules. At low metallicities, the coupling between radiation and matter is weakened with respect to the solar-metallicity case because of the decreased line opacities; the adiabatic cooling term therefore prevails, causing the thermal balance to shift toward on average lower temperatures. Stationary, 1D, hydrostatic models do not account for this cooling term associated with diverging gas flows, and the thermal balance in the upper photosphere is controlled by radiative heating and cooling only, ultimately leading, at low metallicities, to higher equilibrium temperatures than predicted by 3D models (Fig. \[atm\_abs\], left panels). The lower average temperatures in the outer layers of metal-poor 3D models also imply smaller values of the pressure scale height and, consequently, steeper pressure stratifications on a geometrical scale with respect to corresponding 1D models. However, the lower temperatures also result in lower opacities; therefore, on an optical depth scale, the average gas pressure in the optically thin layers of the metal-poor 3D models appears typically higher at a given optical depth than in 1D models (Fig. \[atm\_abs\], right panels).
Statistical equilibrium codes
-----------------------------
[\[sec:nlte\_codes\]]{}
The NLTE level populations of and were computed with an updated version of the <span style="font-variant:small-caps;">detail</span> code [@Butler85] and <span style="font-variant:small-caps;">multi</span> [@1986UppOR..33.....C; @1992ASPC...26..499C]. In both codes, the solution of the coupled statistical equilibrium and radiative transfer equations is obtained using an approximate lambda iteration method. In <span style="font-variant:small-caps;">detail</span>, the latter is implemented following a $\Psi$-operator approach of the kind described by , which allows for self-consistent treatment of overlapping transitions and continua. <span style="font-variant:small-caps;">multi</span> is based on the method described by @1985JCoPh..59...56S with the local operator by @1986JQSRT..35..431O. Scattering in the bound-bound (hereafter, b-b) transitions included in the model atom follows complete frequency redistribution. Upper boundary conditions differ in that <span style="font-variant:small-caps;">detail</span> assumes no incoming radiation at the top, whereas <span style="font-variant:small-caps;">multi</span> estimates the minor contribution from the optically thin gas and uses a second order Taylor expansion of the Feautrier variables. Ng convergence acceleration [@1974JChPh..61.2680N; @1987nrt..book..101A] is implemented in both codes.
In both codes, the background line opacity is computed with a Planckian source function. However, there is one difference. In <span style="font-variant:small-caps;">detail</span>, line opacity is consistently added at the frequencies of all b-b and b-f transitions in the NLTE atom. In <span style="font-variant:small-caps;">multi</span>, metal line opacities were added to the continuous opacities for the calculation of photoionization rates only, as described by , while blends were neglected in the calculation of the bound-bound radiative rates. This approximation is well justified and saves computational time. Firstly, Fe itself dominates line opacity in the UV. Secondly, according to our tests, the upper limit to the differences in the NLTE equivalent widths computed with and without blends is $\sim 0.2$ percent for the Sun and $5$ percent for HD 122563.
When solving for statistical equilibrium, Fe line profiles were computed with a Gaussian in <span style="font-variant:small-caps;">detail</span> and with Voigt profiles in <span style="font-variant:small-caps;">multi</span>. Although the latter is also possible with <span style="font-variant:small-caps;">detail</span>, it is un-necessary. Our tests showed that NLTE effects in are insensitive to the adopted profile functions. Once the level populations were converged with <span style="font-variant:small-caps;">detail</span>, the synthetic line profiles were computed with SIU [@reetz].
Finally, we remark on the handling of the equation-of-state in the NLTE codes. Whereas <span style="font-variant:small-caps;">multi</span> includes a subroutine to compute ionization fractions and molecular equilibria internally, <span style="font-variant:small-caps;">detail</span> and <span style="font-variant:small-caps;">siu</span> require partial pressures of all atoms and important molecules to be supplied with a model atmosphere. These are included in the <span style="font-variant:small-caps;">mafags-os</span> models. For the <span style="font-variant:small-caps;">marcs</span> and $\td$ models, we computed the atomic partial pressures using the equation-of-state package built in <span style="font-variant:small-caps;">multi</span>.
Model atom of Fe
----------------
[\[sec:modelatom\]]{}
### Levels and radiative transitions
The number of levels and discrete radiative transitions in Fe is enormous. Recent calculations of Kurucz[^4] predict $\sim 37\,000$ theoretical energy levels below and above the first ionization threshold of , as well as $6\,025\,000$ radiatively-permitted transitions between them. Such atomic models are not tractable with our NLTE codes; thus, we combined atomic levels and transitions into super-levels and super-lines (see below). In our final model, which also includes all experimental data from the NIST atomic database [@2012AAS...21944301R], the number of energy levels is $296$ for and $112$ for , with uppermost excited levels located at $0.03$ eV and $2.72$ eV below the respective ionization limits, $7.9$ eV and $16.19$ eV. The model is closed by the ground state. The total number of radiatively-allowed transitions is $16\,207$ ($13\,888$ and $2\,316$ ). Fine structure was neglected for all levels, but the ground states of (configuration $1s^2 2s^2 2p^6 3s^2 3p^6 3d^6 4s^2$) and . Excitation energy of each LS term is a weighted mean of statistical weights and excitation energies of fine structure levels. The mean wavelength for a multiplet is computed from the weighted energy levels.
All predicted energy levels of with the same parity above $\Elow
\geq 5.1$ eV were grouped into superlevels. For the upper levels above $54000$ cm$^-1$ ($6.7$ eV), we combined all levels within $1000$ cm$^{-1}$ ($0.12$ eV)[^5], whereas below this energy limit only levels deviating by less than $10$ cm$^{-1}$ ($0.001$ eV) were combined. Thus, not only predicted, but also some experimental levels above $5.1$ eV were combined to superlevels. None of the Fe lines selected for the subsequent abundance analysis has either the lower or the upper level combined.
Transitions between the components of the superlevels were also grouped preserving the parity and angular momentum conservation rules. The total transition probability of a superline is a weighted average of $\log gf$’s of individual transitions, and is computed in analogy to a $gf$-value for a multiplet [@1988atps.book.....M]: $$f_{\rm mean} = \frac{1}{\bar{\lambda}_{\rm mean} \sum_{l} g_l}
\sum_{l,u} g_l \times \lambda_{\rm lu} \times f_{\rm lu}$$ where $l$ and $u$ are indices of lower and upper levels of the un-grouped transition, $\bar{\lambda}_{\rm mean}$ is the mean Ritz wavelength of a superline. Grotrian diagram of the Fe model atom is shown in Fig. \[fe\_grotrian\]. Although the accuracy of individual $gf$-values for the multitude of theoretically-computed transitions is hard to quantify, we expect our model to be a good representation of the global atomic properties of the / and provide physically valid description of statistical equilibrium of the atom in the atmospheres of late-type stars.
Accurate radiative bound-free cross-sections for computed using the close-coupling method were kindly provided by M. Bautista (private communication 2011, ). These data are computed on a more accurate energy mesh and provide better resolution of photoionization resonances compared to the older data, e.g. provided in the TOPbase. Thus, $136$ levels of , including different multiplet systems, from singlets to septets, are represented by quantum-mechanical data. Hydrogenic approximation was used for the other levels.
### Collisional transitions
The rates of transitions induced by inelastic collisions with free electrons (e$^-$) and atoms were computed using different recipes. For states coupled by allowed b-b and b-f transitions, we used the formulas of @1962ApJ...136..906V and @1962amp..conf..375S, respectively, in the case of e$^-$ collisions and that of , in the case of collisions. Also, all states are coupled by forbidden transitions induced by $e^-$ collisions using the formula of @1973asqu.book.....A and by collisions with the formula of @1994PASJ...46...53T.
We constrain the efficiency of collisions with empirically (Sect. \[sec:solspec\], \[sec:abstars\]). We apply a scaling factor $\SH$ to the cross-sections, which is iteratively derived from the constraint of / ionization equilibrium for the standard stars with independently-fixed stellar parameters. As pointed out by , such approach is an over-simplification of the problem. However, at present there are no useful alternatives to this classical recipe. Evidence for the necessity of including inelastic collisions with in statistical equilibrium calculations for Fe and other elements has been demonstrated in many studies , which is also confirmed by us (Sect. \[sec:solspec\], \[sec:abstars\]).
Statistical equilibrium of Fe
=============================
[\[sec:statec\]]{} In the following section, we present the results of statistical equilibrium calculations for Fe obtained with <span style="font-variant:small-caps;">detail</span> and <span style="font-variant:small-caps;">multi</span>. We briefly describe the physical processes responsible for deviations from LTE level populations in the and atoms for atmospheric conditions typical of FGK stars. A comprehensive description of NLTE effects as a function of stellar parameters is deferred to the second paper of this series [@Paper2]. The main aim here is to understand the differences in terms of NLTE line formation with 1D and $\td$ model atmospheres, which is necessary in order to explain the large differences between the spectroscopic values of $\Teff$, $\log g$, and \[Fe/H\] obtained for the reference stars in the two cases (Sect. \[sec:abstars\]).
is a minority ion in the atmospheres of late-type stars, which increases its sensitivity to NLTE over-ionization. The overall effect is that, compared to LTE, the statistical equilibrium of Fe favors lower number densities of the neutral ion, , although the number densities of relevant levels remain nearly thermalized. In general, this leads to a weakening of lines compared to LTE that, in turn, requires larger Fe abundance to fit a given observed spectral line. The actual magnitude of departures and NLTE abundance corrections depends on stellar parameters (Sect. \[sec:abstars\]).
Figure \[dep\] presents the departure coefficients[^6] of selected and levels computed for the <span style="font-variant:small-caps;">marcs</span> and $\td$ model atmospheres with <span style="font-variant:small-caps;">multi</span>. The results are shown for the Sun, the metal-poor subdwarf HD 84937, the metal-poor sub-giant HD 140283 and the metal-poor giant HD 122563. Only a few, selected levels, typical for the dependence of their associated departure coefficients with depth, are included in the plot: (ground state of ), ($2.4$ eV), ($5.4$ eV), , and two levels, which belong to the ion: (ground state) and ($5.5$ eV). In the optically thin atmospheric layers these and majority of other levels are underpopulated compared to LTE, $b_i < 1$. For all stars, but HD 122563, the number densities remain close to LTE values throughout the atmosphere, $b_i \approx 1$, and a minor overpopulation of the ground state develops only close to the outer atmospheric boundary.
Deviations from LTE in the distribution of atomic level populations arise because the mean radiation field, $J_\nu$, at different depths and frequencies is not equal to the Planck function, $B_\nu [T_{\rm e}(\tau)]$. For , excess of the mean intensity over Planck function in the UV continua leads to over-ionization, which sets in at $\opd \approx 0.2$, i.e., as soon as the optical depth in the photoionization continuum of low-excitation levels with $E \approx 2$ eV falls below unity (Fig. \[dep\]a).
In the layers with $\opd < -1$, the dominant mechanism is over-ionization from the levels with excitation energies at $2-5$ eV. The lower-lying levels, including the ground state of , maintain underpopulation due to radiative and collisional coupling with the former. Excitation balance of is also mildly affected by radiative imbalances in line transitions. These include radiative pumping by the non-local UV radiative field, as well as photon suction driven by photon losses in large-probability transitions between highly-excited levels. These processes leave a characteristic imprint on the behavior of $b_i$-factors in the outer atmospheric layers, $\opd < -2$. In the infrared continuum, $J_\nu < B_\nu$ leads to over-recombination, which is very efficient for our atomic model with only $0.03$ eV energy gap of the upper levels from the ground state. This is illustrated in Fig. \[rates\], which shows net radiative brackets[^7] for all levels in the model atom at the depths $\opd = 0.12$, $-1.2$, and $-3.86$. All levels with excitation energy between $2$ and $6$ eV experience net over-ionization, and the loop is closed by net over-recombination to the upper levels. Note that only radiative rates are plotted. We also compared absolute radiative rates from <span style="font-variant:small-caps;">multi</span> and <span style="font-variant:small-caps;">detail</span> and found that they agree very well with each other.
![Mean radiation field compared to the local Planck function in the UV continuum computed with the 1D and $\td$ model atmospheres of the metal-poor sub-giant HD 140283.[]{data-label="jnu"}](fig16.ps)
A comparison of the departure coefficients computed with 1D and $\td$ model atmospheres reveals important differences (Fig. \[dep\]), which depend on stellar parameters, although their qualitative behavior is the same. Generally, metal-poor $\td$ models show more thermalization close to continuum-forming layers, while in the outer layers atoms experience a larger degree of over-ionization. This can be primarily understood based on the temperature gradients, which in the $\td$ stratifications are shallower in the inner layers but steeper in the outer layers compared to 1D models. We briefly consider the case of the metal-poor subgiant HD 140283, for which the 1D and $\td$ model atmospheres are compared in Fig. \[atm\_abs\]. The $\td$ model leads to smaller deviations from LTE at $-2 \lesssim \opd \lesssim 0$. That is, for the majority of the underpopulated levels, $b_i$(1D) $<$ $b_i$($\td$). Inspection of the respective T$(\tau)$ relations in Fig. \[atm\_abs\] shows that the $\td$ model is cooler than the <span style="font-variant:small-caps;">marcs</span> model at $0 \lesssim \opd \lesssim 1$, where the UV continua form, but it is $\sim
100$ K hotter higher up. Figure \[jnu\] illustrates mean angle-averaged intensities at $2200$ Å, i.e., at the wavelength sampled by the ionization edges of the important levels. These are shown for the 1D and $\td$ models, and are compared with the local Planck functions $B_{\nu}$ at each optical depth. At $-1.5 \lesssim \opd \lesssim 0$, the $J_{\nu} -
B_{\nu}$ imbalance is smaller in the $\td$ model compared to the 1D model, and, thus, over-ionization is less efficient. In this range of optical depths, densities in the $\td$ stratification are also slightly larger than in the 1D model (Fig. \[atm\_abs\]). At $\opd \lesssim -2$, the situation is reverse. The $\td$ model is nearly $1000$ K cooler and less dense than the 1D model. The $J_{\nu} -
B_{\nu}$ imbalance becomes very large causing significant underpopulation of the lower levels, while uppermost levels and the ground state and excited levels develop appreciable overpopulation compared to their LTE occupation numbers. In contrast, 1D models predict nearly thermalized level populations, over the whole optical depth scale.
Figure \[total\_number\] shows the ratios of / number densities for the four cases: <span style="font-variant:small-caps;">marcs</span> and $\td$ model atmospheres, NLTE and LTE. In the LTE case, about $\sim 5 - 10$ percent of the element is in the form of above $\opd \sim 0$. In the $\td$ models, due to their steeper T($\tau$) relation, LTE recombination of to is even more efficient, raising / to $20 - 50$ percent. Under NLTE, this effect is suppressed since ionization balance is primarily set by radiative ionization and recombination. For the metal-poor stars, occupies less then $1$ percent of the total element abundance at $\opd < -2$ notwithstanding much cooler outer layers of the $\td$ models. We note, however, that the huge differences in the ionization balance of / in the outer layers of the models are not important for our NLTE abundance determinations. Most of the lines observed in spectra of the selected metal-poor stars form close to the continuum forming regions, $-2 <
\opd < 0$.
Extensive tests demonstrated that the behavior of departure coefficients is robust against the differences in terms of model atmospheres and NLTE codes used for the statistical equilibrium calculations. Perhaps, the only systematic effect is that <span style="font-variant:small-caps;">multi</span> predicts slightly stronger NLTE effects than <span style="font-variant:small-caps;">detail</span> mainly due to the differences in the background opacity. Also, the comparison of the LTE and NLTE equivalent widths obtained with <span style="font-variant:small-caps;">multi</span> and <span style="font-variant:small-caps;">detail</span>/<span style="font-variant:small-caps;">siu</span> show agreement to within $1 - 5$ percent for the lines with $\EW > 10 ~\mA$; NLTE $\EW$’s are somewhat divergent for the weakest lines, with the relative $(\textsc{multi}-\textsc{detail})$ differences being about $10$ percent. These differences shall be taken as a formal error, which arises as a consequence of different implementations of the same basic physics in the various codes. Still, this is not of a big concern in our work because the calibration of the models and grid calculations are consistently performed with same code (<span style="font-variant:small-caps;">multi</span>). Furthermore, the magnitude of the NLTE effects is initially effectively controlled by a free parameter, the scaling factor $\SH$ for the efficiency of the collisions.
Spectroscopic stellar parameters
================================
[\[sec:solspec\]]{}
Method
------
[\[sec:method\]]{} Spectroscopic stellar parameters, including metallicity, effective temperature, gravity, and microturbulence, were determined for each model atmosphere (<span style="font-variant:small-caps;">marcs</span>, <span style="font-variant:small-caps;">mafags</span>, $\td$) in LTE and NLTE using the following approach.
In a first step, we performed full spectrum synthesis for the reference stars with the <span style="font-variant:small-caps;">siu</span> line formation code using the 1D <span style="font-variant:small-caps;">mafags-os</span> model atmospheres computed with parameters described in Sect. \[sec:obs\]. The <span style="font-variant:small-caps;">mafags-os</span> models were used at this stage because they include partial pressures for molecules, which are important contributors to the background opacity of cooler stars. The profiles of all diagnostic Fe lines computed in LTE and NLTE were visually fitted to the observed spectra. The lines were computed with depth-dependent Voigt profiles taking into account various external broadening mechanisms. The equivalent widths of the Fe lines were then obtained from the best-fitting NLTE profiles, excluding contribution of blends. The $\EW$’s are given in Tables \[ew1\] and \[ew2\] of the Appendix and are accurate to within $1-2$ percent.
In a second step, the equivalent widths $\EW$’s were applied to determine spectroscopic parameters using the grids of LTE and NLTE equivalent widths computed with <span style="font-variant:small-caps;">multi</span> for <span style="font-variant:small-caps;">marcs</span> and $\td$ model atmospheres. For each star, we constructed a local model atmosphere grid with two effective temperature points (the IRFM value and a value $200$K lower), two values of microturbulence ($1$ and $2$ km s$^{-1}$), and a range of plausible metallicities ($\Delta$\[Fe/H\] up to $\pm 1.5$ dex) with a step-size of $0.25$dex. Equivalent widths were computed for the grid models in LTE as well as in NLTE for two values of the collision efficiency parameter, $\SH
=0.1$ and $\SH =1$. In order to save computational time in the construction of the local model grids in the $\td$ case, the $\td$ models were simply adjusted by multiplying the temperature and electron pressure at a given optical depth with the ratios of those quantities obtained from corresponding 1D models. Based on our experience with scaling 1D and 3D model stellar atmospheres to different stellar parameters, this procedure appears accurate to first-order level. The grid was thus constructed for <span style="font-variant:small-caps;">marcs</span> and $\td$ models.
<span style="font-variant:small-caps;">mafags-os</span> model parameters were adjusted by assuming the equivalent widths to have the same sensitivity to stellar parameters as for <span style="font-variant:small-caps;">marcs</span> model, which is reasonable considering the similarity between the two codes.
An iterative procedure was then applied to determine spectroscopic parameters from and lines. First of all the microturbulent velocity was determined by flattening the slope of line abundances with reduced equivalent width, thereby circumventing differential NLTE effects with line strength of lines. Only for the most metal-poor star G 64-12, the microturbulence is poorly defined due to the lack of strong lines. We then followed two different approaches, keeping either $\log g$ or $\Teff$ fixed, and optimizing the other parameter to establish ionization balance between the and lines. In order to avoid biased results, we have not attempted to adjust the oscillator strengths of the Fe lines, giving preference to the experimental data from the laboratory measurements, or their weighted means if few measurements were available (Sect. \[sec:line\_par\]). Hence, all abundance results are absolute values. Furthermore, we did not perform a differential stellar abundance analysis with respect to the Sun. First, that would conflict with the use of the reference stars as a calibration sample. Second, a differential approach would introduce yet an another source of error related to the usual problems of fitting the strong Fe lines in the solar spectrum, in particular the ambiguity between the effect of van der Waals damping and abundance on a line profile.
The standard technique to infer effective temperatures by flattening the slope of line abundances with excitation potential is particularly sensitive to the choice of transition probabilities. We found, however, that it is only weakly sensitive to $\SH$, which is the main parameter we seek to constrain. Initial attempts to vary all parameters simultaneously, i.e. microturbulence from line strengths, effective temperature by excitation balance and surface gravity by ionization balance, did not give a conclusive answer in terms of the best choice of $\SH$. This apparent weakness is likely related to the problem of multidimensionality in the parameter space, and, as such, shall be inherent to the method itself. In addition, we do not exclude remaining systematic uncertainties in the atomic and atmospheric models. It is possible that a differential approach between similar stars would be more successful in establishing temperatures based on excitation balance. We therefore reduced the dimensionality of the problem by not enforcing strict excitation equilibrium and focusing only on the ionization balance, as described above. As the results will show, an adequately flat trend with excitation potential is anyway naturally achieved for the majority of stars, HD 122563 being a notable exception.
The analysis described above was applied to each star in the sample using the <span style="font-variant:small-caps;">marcs</span>, <span style="font-variant:small-caps;">mafags</span>, and $\td$ model atmospheres, thus yielding four desired quantities: $\Teff$, $\log g$, \[Fe/H\], and $\Vmic$. The results are discussed in Sect. \[sec:abstars\]. Table \[tab:fin\_param\] gives the mean metallicities determined using the reference $\Teff$ and $\log
g$, and abundances averaged over the measured and lines together with their standard deviations are presented in Table \[tab:ion\_abu\].
Observations and reference stellar parameters
---------------------------------------------
[\[sec:obs\]]{} Our reference sample consists of six late-type stars (Table \[tab:init\_param\]), and includes two solar-metallicity stars (the Sun and Procyon), two metal-poor dwarfs (HD 84937 and G 64-12), a metal-poor subgiant (HD 140283), and a very bright metal-poor giant (HD 122563). The following observed spectra were adopted here. For the Sun, we used the KPNO flux spectrum (Kurucz et al. 1984). The UVES observations of Procyon (HD 61421), HD 84937, HD 140283, and HD 122563 were taken from the UVES-POP survey [@2003Msngr.114...10B]. These spectra have a slit-determined resolution of $\lambda/\Delta\lambda \sim 80\,000$ and a signal-to-noise ratio $S/N \sim 300$ near $5000$ Å. The UVES spectrum for G 64-12 was taken from ESO/ST-ECF Science Archive Facility (67.D-0554(A)). The number of Fe lines suitable for the spectrum synthesis at this spectral quality, is $40$ (Procyon) to $10$ (G 64-12). For comparison, we also used the Keck/HIRES spectra of G 64-12 with $\lambda/\Delta\lambda \sim 100\,000$ and $S/N \sim 500$ kindly provided by J. Melendez (private communication).
A crucial step in our analysis is the choice of the reference $\Teff$ and $\log
g$ values for the selected stars, which are used as a benchmark for testing the new NLTE model atom and $\td$ model atmospheres. These parameters were taken from the literature, giving preference to the least model-dependent methods, such as interferometry, infra-red flux method (IRFM), and parallaxes. The adopted values and their uncertainties are given in the Table \[tab:init\_param\]. A brief description of these data is given below.
For the four metal-poor stars, we adopted the IRFM effective temperatures by . The mean internal uncertainty of the data is about $70$ K, which includes the uncertainty on the zero point of the $\Teff$ scale, reddening and photometric errors.
The IRFM value for the metal-poor giant HD 122563, $\Teff = 4665 \pm 80$ K, was kindly provided by L. Casagrande (private communication). This estimate incorporates a correction due to reddening, $E(B-V) = 0.005$, which was determined from the interstellar Na D lines detected in the UVES spectrum. Our equivalent widths for the $5889.95$ and $5895.92$ Å lines are $22.3$ and $12.4$ , respectively.
For Procyon, an interferometric estimate of angular diameter is available [@2005ApJ...633..424A], which gives $\Teff = 6545 \pm 83$ K and combined with a very accurate parallax $\pi = 284.56 \pm 1.26$ milli-arcsec, surface gravity can be estimated, $\log g = 3.99 \pm 0.02$.
For the other stars, we determined surface gravities from the [Hipparcos]{} parallaxes using the masses estimated from the tracks of @2000ApJ...532..430V by and apparent bolometric magnitudes from . The uncertainties were computed by mapping the uncertainty in mass[^8] (0.2 M$_\odot$ for HD 122563 and G 64-12, and $0.1$ M$_\odot$ for HD 84937 and HD 140283), temperature and parallax into the range of possible gravities ($\Delta \logg \leq 0.16$ dex). The only exception is G 64-12, for which the parallax is too uncertain, $\pi = 0.57 \pm 2.83$ milli-arcsec. From comparison with metal-poor evolutionary tracks for reasonable masses (Fig. \[tracks\]), one may also conclude that within the errors of the given $\Teff$ ($6464$ K) surface gravity of G 64-12 is in the range $\log g =
3.8 \ldots 4.6$ dex. We, thus adopt the value derived from the upper limit on the parallax, $\log g = 4.3$ and assign a nominal error of $0.3$ dex. In comparison, and estimate $\log g = 4.26$.
Line selection
--------------
[\[sec:line\_par\]]{}
The Fe lines for the abundance calculations were selected on a star by star basis, i.e., carefully inspecting the observed stellar spectra. We rejected lines affected by blends, strong damping wings or are located in the spectral windows where continuum placement is uncertain. Our solar line list includes $59$ lines of and $24$ lines of in the wavelength range $4400 - 8500$ Å. In spectra of the metal-poor stars most of these lines are very weak. Thus, $12$ strong lines of were added to the analysis of the metal-poor stars. All line parameters are given in Tables \[sun2601\], \[sun2602\], and \[stars26\].
The $gf$-values adopted in this work are weighted averages of different experimental values. An elaborate discussion of their accuracy can be found in . The largest weights are typically assigned to the $\log gf$ values measured at Oxford (@1979MNRAS.186..633B; @1979MNRAS.186..657B; @1979MNRAS.186..673B; @1980MNRAS.191..445B; @1982MNRAS.199...33B; @1982MNRAS.199...43B; @1982MNRAS.201..595B; @1986MNRAS.220..549B). According to the NIST database, the uncertainties of the data are $3 - 10\%$. Smaller weights are assigned to the $gf$-values of the Hannover group (, , uncertainties typically $10 - 25\%$) and @1991JOSAB...8.1185O [uncertainties $25 -50
\%$]. The $gf$-values for the transitions are taken from , who renormalized the branching fractions from and other theoretical sources to the experimental lifetimes (). To avoid biased results, we excluded any of their ’solar’ values, which were obtained by a 1D LTE spectroscopic analysis of the solar flux spectrum. That is, for the transitions $\lambda\lambda\ 5284.1,
6239.95, 6247.56, 6456.38\ \AA$ we kept the NIST-recommended values.
Our adopted transition probabilites for the and lines are compared to the NIST-recommended values in Fig. \[gf\]. There are small differences for the weak transitions, which reflect the discrepancies between the NIST data based essentially on the study and the adopted values of . Note that in the re-analysis of the solar Fe abundance discuss the possibility of underestimated $\log gf$’s from for the optical lines, although they do not find the same problem for the UV lines.
Line broadening due to elastic collisions with atoms was computed from the tables of @2000yCat..41420467B and . For the four lines ($4445.47$, $5600.22$, $5661.35$, $8293.51$ Å), broadening cross-sections $\sigma$ and their velocity dependence $\alpha$ were kindly provided by P. Barklem (2011, private communication). For seven lines, the @1955QB461.U55...... values increased by a factor of $1.5$ were adopted. In Tables \[sun2601\], \[sun2602\], and and \[stars26\] these values are given in terms of commonly-used van der Waals damping constants $\log
C_6$.
Sun and Procyon
---------------
[\[sec:sun\_procyon\]]{} Owing to the superior quality of the solar spectrum, the initial spectrum synthesis for the Sun was performed taking into account solar rotation with $V_{\rm rot , \odot} = 1.8$ and a radial-tangential macroturbulence velocity $\Vmac$, which was adjusted for each line separately. The typical values of $\Vmac$ required to match the observed spectrum are $2.5 \ldots 4$ for the profile fitting with the 1D models, and $1.6 \ldots 3 $ for that with the $\td$ model. We note that in full 3D line formation calculations found that no macroturbulence was necessary to fit the line profiles due to the Doppler shifts from the convective motions, which are not explicitly taken into account in the $\td$ models used here.
Selected and solar line profiles computed under LTE and NLTE with the $\td$ models are compared with the observations in Fig. \[profiles\]. In general, NLTE line profiles are weaker compared to LTE, which is driven by the NLTE effect on the line opacity. For the strongest lines, forming very far out in the atmospheres, deviation of the line source function from the Planck function, $S_{\rm ij} <
B_{\nu}(\Te)$, leads to line core darkening. However, their wings, which dominate the total line strength, are formed at the depths where $b_i < 1$ due to over-ionization, and the NLTE absorption coefficient is smaller. Therefore the NLTE equivalent widths are smaller compared to LTE for a given abundance. As discussed in Sect. \[sec:statec\], NLTE effects on the lines are negligible. We also inspected the influence of inelastic collisions on the line profiles finding that the profile shapes are rather insensitive to the $\SH$ value. Any Fe line could be fitted by slightly adjusting the abundance or the damping parameters within the error bars (typically $10$ percent).
The mean Fe abundances computed with the <span style="font-variant:small-caps;">mafags-os</span>, <span style="font-variant:small-caps;">marcs</span>, and the $\td$ model atmospheres are given in the Table \[tab:fin\_param\] (see also Fig. \[st12\]). The total errors of the mean are only shown for the NLTE results obtained with the $\td$ models. The total error was computed as:
$$\sigma_{\rm tot} = \left((\sigma_{\rm tot, Fe I}^2 + \sigma_{\rm tot, Fe
II}^2)/2 \right)^{1/2}$$ where, $$\sigma_{\rm tot,~Fe} = (\sigma_{\rm Fe}^2 + \sigma_{\log gf,
\rm{Fe}}^2 + \sigma_{\rm \log g, Fe}^2 + \sigma_{\Teff, \rm{Fe}}^2)^{1/2}$$ The total error includes the following uncertainties: observational errors given by the standard deviation, errors stemming from the reference values of the surface gravity and effective temperatures (Table \[tab:init\_param\]), and in the $\log gf$ values. For the latter, we adopted the uncertainties given by NIST and assumed they are uncorrelated, since the $gf$-values come from different sources in the literature[^9]. That is, the total uncertainty $\sigma_{\log gf}$ is decreased by a factor $\sqrt N$, where $N$ is the number of lines, and it takes into account the relative number of the and lines analyzed per star.
We find that the excitation and ionization equilibrium of Fe in the Sun and Procyon is well satisfied under NLTE with both 1D and $\td$ models (Fig. \[st12\], Table \[tab:ion\_abu\]). The statistical uncertainty of the mean abundance, which corresponds to the standard error of the line sample, is not larger than $0.01$ dex. Some lines with $\EW > 100\ \mA$ seem to indicate slightly lower abundances, by $\sim 0.03$ dex, compared to the rest of the sample both in 1D and $\td$ modelling. We do not assign much weight to these lines. First, it is almost impossible to discern the influence of abundance, damping, and weak blends in the line wings. Second, their asymmetric profiles are clearly shaped by convective velocity fields, which are not accounted for in our models. As a consequence, it is hardly possible to assign a unique ’best-fitting’ abundance to such lines at all, that, in addition to a systematic error caused by the neglect of 3D effects, introduces subjectivity in abundance estimates.
The mean NLTE abundances of Fe in the Sun and Procyon, determined using the $\td$ model atmospheres, are $7.46 \pm 0.02$ respectively $7.43 \pm 0.07$ dex. 1D models (<span style="font-variant:small-caps;">marcs</span> or <span style="font-variant:small-caps;">mafags-os</span>) yield slightly lower values, which are, however, all consistent within the total error. For the Sun, the latter only reflects the statistical uncertainty of the $gf$-values, as described above. We also find that adopting LTE or scaling down the cross-sections for inelastic collisions with by an order of magnitude, $\SH =0.1$ does not yield any significant effect on the mean abundance. The difference between the two extreme cases $\Delta$(NLTE-LTE) is $\approx 0.02$ dex. Thus, based on the analysis of the solar-metallicity stars only it is not possible to single out the optimum value of $\SH$. It turns out, however, that this problem can be solved using the metal-poor stars with reliable stellar parameters (Sect. \[sec:abstars\]).
Comparing our results with the published values of the Fe abundance in the Sun and Procyon, we find full agreement with all data obtained with similar techniques and model atmospheres. We list only few most recent examples. using a $\td$ solar model very similar to that adopted in this work and LTE approach obtain $\logeFetwos\ = 7.45$ dex. The NLTE 1D estimate by obtained using the <span style="font-variant:small-caps;">mafags-odf</span> (note the difference with the <span style="font-variant:small-caps;">mafags-os</span> version used here) models is $\logeFesun\ = 7.48 ... 7.51$ dex. In the follow-up NLTE study with the <span style="font-variant:small-caps;">mafags-os</span> models, find $\logeFesun\ = 7.41 \ldots 7.56$ depending on the source of $gf$-values used for the lines. For Procyon, claim discrepant abundances from the NLTE analysis of the and lines, $\logeFeone\ = 7.46 \pm 0.07$ respectively $\logeFetwo\ = 7.52 \pm 0.05$ dex (their Table 3, solution $\SH = 0.1$) suggesting that either increasing $\Teff$ by $80$ K or decreasing $\log g$ by $0.15$ dex may solve the problem. Their results for Procyon with $\SH = 1$, which is consistent with our study, are discrepant, $\logeFeone\ = 7.41 \pm 0.07$ respectively $\logeFetwo\ = 7.53 \pm 0.05$ dex, which is most likely due to the different sources of $\log gf$ values for the lines.
We emphasize that further, more detailed, comparison to these and other studies is not meaningful because of various methodical differences, such as atomic data, line selection, and microturbulence, to name just a few.
Analysis of metal-poor stars {#sec:abstars}
----------------------------
We start with a discussion of how NLTE affects the Fe abundances obtained with 1D and $\td$ models, delineating the key differences. In Sect. \[sec:teff\], we then illustrate how these differences propagate in the determination of effective temperatures and surface gravities.
### Metallicities
[\[sec:meta\]]{}
The most important test of our new models is whether they are able to recover ionization equilibrium of and for the stars with parameters determined by independent techniques, given a unique $\SH$. As described in Sect. \[sec:obs\], a small set of metal-poor stars in different evolutionary stages was selected for this purpose with $\Teff$ known from the interferometry and/or IRFM, and gravities well constrained by parallaxes. Metallicities and microturbulence parameters were then determined keeping the $\Teff$ and $\logg$ fixed, and varying the value of $\SH$ in the NLTE calculations.
We find that the solution providing optimum ionization balance of Fe for all metal-poor stars is achieved with $\SH = 1$, i.e, unscaled Drawin’s collisions (see also discussion in Sect. \[sec:teff\]). The Fe abundances determined in this way, along with the optimized $\td$ NLTE microturbulence $\Vmic$, are given in the Table \[tab:fin\_param\]. These are mean abundances from the and lines, and the errors were computed as described in Sect. \[sec:sun\_procyon\]. For all stars, but the Sun and G 64-12 with its very uncertain parallax, the errors are dominated by the uncertainties of the reference $\Teff$. The abundances are shown in Figs. \[st12\], \[st34\], and \[st56\] as a function of line equivalent widths and lower level excitation potentials (in eV). The error bars in the figures indicate the size of NLTE abundance corrections for the low-excitation ($< 2.5$ eV) and high-excitation lines. They are levelled out at the mean NLTE abundance obtained from the and lines with $\SH = 1$. The bar’s upper and lower ends correspond to the mean NLTE abundance obtained with $\SH = 0.1$ and the LTE abundance, respectively. The abundances averaged over the measured respectively lines are presented in Table \[tab:ion\_abu\] along with their one sigma errors. Note that these are not total propagated errors as described in Sec. 4.4, but standard deviations of the mean abundance for each ionization stage. These quantities mainly serve to show the difference in the line scatter between the 1D (<span style="font-variant:small-caps;">marcs</span>) and $\td$ models.
The major differences between 1D and $\td$, as well as LTE and NLTE, are apparent from inspection of the Fe abundances in the Table \[tab:fin\_param\] and the NLTE abundance corrections[^10] shown in Fig. \[delta\]. For the Sun and Procyon, the NLTE effects are rather subtle. The difference between the LTE and NLTE abundances determined from the lines is even slightly smaller in the $\td$ case compared to 1D (Fig. \[delta\], top panels). The basic reason is that the temperature gradient in the solar-metallicity <span style="font-variant:small-caps;">marcs</span> models is slightly steeper than in the $\td$ models (Fig. \[atm\_abs\]). At $\opd > 0$, where the UV-blue continuum is formed, $\left( d T/d \tau \right)_{\rm 1D} >
\left( d T/d \tau \right)_{\rm 3D}$, whereas in the outer layers the gradients are very similar. Over-ionization in by radiation field emerging from the hot deep layers is, thus, somewhat stronger in the <span style="font-variant:small-caps;">marcs</span> models.
Over-ionization is more important at low metallicities (see Sect. \[sec:statec\] of this paper, and a more extensive discussion in Paper II). This causes non-negligible differences between the NLTE and LTE abundances inferred from the lines. The differences are in the range $0.05 -
0.15$ dex for 1D models and weak lines, but increase to $0.2 - 0.5$ dex for the strong low-excitation lines with the $\td$ models (Fig. \[delta\]). The larger NLTE effects of metal-poor $\td$ models compared to 1D models are due to the larger decoupling between the radiation temperature and the local temperature, owing to the dramatically cooler outer layers of the $\td$ models compared to 1D (see Sect. \[sec:statec\] and Fig. \[jnu\]). NLTE corrections to the lines are negative and are not larger than $-0.03$ dex. As a result, for the metal-poor stars, the difference between the mean LTE and NLTE Fe abundances is, at least, twice as large in the $\td$ case compared to that obtained with <span style="font-variant:small-caps;">marcs</span> and <span style="font-variant:small-caps;">mafags-os</span> models. On the other side, the same phenomenon, which leads to amplified NLTE effects in the $\td$ calculations, i.e., very steep $\left( d T/d \tau \right)$ above $\opd
\approx 0$, is also responsible for the LTE strengthening of the lines in $\td$ compared to 1D. The lines are only slightly weaker in $\td$ calculations. Our LTE 1D metallicities are, thus, systematically larger than the $\td$ results, in agreement with other studies . We note, however, that direct comparison of our results with that of, e.g., , is not meaningful since they performed LTE line formation calculations with full 3D hydrodynamic model atmospheres.
The fact that the $\td$ models also predict somewhat lower LTE Fe abundances, yet larger NLTE abundance corrections than 1D models, has the important consequence that the mean NLTE metallicities in 1D and $\td$ turn out to be in agreement for the whole range of stellar parameters ($\Delta
\log\varepsilon_{\rm Fe} \leq 0.1$).
However, even though the mean metallicities are similar, individual line abundances are still somewhat discrepant, and show systematic trends with line equivalent width and excitation potential of the lower level of a transition, $E_{\rm low}$ (Figs. \[st34\], \[st56\]). NLTE abundances inferred from the low-excitation lines of with 1D models are too large compared to the high-excitation lines and lines of . A very similar picture is obtained with the LTE approximation, and it reverses sign in LTE $\td$. A combination of $\td$ and NLTE alleviates this discrepancy, so that and abundances become more consistent. The improvement comes from the different sensitivity of the NLTE effects to $E_{\rm low}$ in the 1D and $\td$ case. This is clearly seen in Fig. \[delta\]. A least-square fit to the NLTE abundance corrections is also shown, which is a simple albeit very illustrative measure of the mean NLTE effect on the excitation balance of . In the 1D case, the lines have rather similar NLTE corrections irrespective of their excitation potential and equivalent width. In the $\td$ case, however, the low-excitation transitions, $E_{\rm low} \leq 2$ eV, tend to experience significantly larger departures from LTE compared to the higher-excitation transitions. The reason is that, owing to the dramatically cooler surfaces of the metal-poor $\td$ models, low-excitation lines become stronger and more sensitive to the physical conditions in the outer atmospheric layers, where the influence of non-local radiation field is extreme, not only decreasing the line opacity but also pushing the line source functions to super-thermal values. Thus, the NLTE $\td$ profiles of these lines are weaker compared to LTE. The magnitude of this effect depends on stellar parameters (see also the discussion in Sect. \[sec:statec\]). As a consequence, one would also expect that excitation balance achieved under LTE with the $\td$ models will strongly overestimate $\Teff$, and the error shall increase with $\Teff$ and decreasing metallicity as indicated by the increasing slope of $\Delta ({\rm NLTE-LTE})$ vs $E_{\rm low}$ for more metal-poor and hotter stars.
A corollary is that accurate metallicities for late-type stars can be obtained with standard 1D model atmospheres if NLTE effects in are taken into account. However, this is true only if the following condition is satisfied: a sufficiently large number of and lines of different types are included in an analysis, so that individual line-to-line abundance discrepancies cancel out. This approach, however, does not eliminate residual trends of abundance with line equivalent width and excitation potential. Although these trends can be, at least in part, corrected for by adjusting microturbulence, a better approach is to restrict an analysis with 1D hydrostatic models to high-excitation lines only, as also advocated in the literature [e.g., @1983PASP...95..101G].
### Effective temperatures and surface gravities
[\[sec:teff\]]{}
 
The spectroscopic effective temperatures and surface gravities determined according to the procedure described in Sect. \[sec:method\] are given in the Table \[tab:opt\_param\].
Fig. \[teff\_logg\] (top panel) plots the effective temperatures obtained for the reference stars adopting a fixed surface gravity inferred from stellar parallaxes. The results are shown for the three cases: LTE, NLTE with the standard model atom ($\SH = 1$), and NLTE with Drawin’s collision cross-sections scaled by a factor of $\SH = 0.1$. The reference temperature scale, T$_{\rm eff, REF}$, refers to the IRFM results (see Sect. \[sec:obs\]) for all stars but Procyon with the interferometric angular diameters; their uncertainties are shown with blue filled areas. The uncertainties of our spectroscopic $\Teff$’s were estimated by mapping the errors in atomic transition probabilities, observational errors, and errors in the input surface gravity into a range in temperatures. In order not to overload the figures, these errors are shown for the NLTE $\SH = 1$ solution only. We emphasize that the $\Teff$’s were obtained by enforcing ionization balance, but not requiring null slopes with excitation potential for the reasons discussed above.
The overall best result in terms of consistency between the spectroscopic effective temperatures, T$_{\rm eff, Fe}$, and effective temperatures obtained by independent methods, T$_{\rm eff, REF}$, is achieved under NLTE using the reference model atom with $\SH = 1$, i.e. standard Drawin’s collisions. The 1D and $\td$ models perform similarly, predicting only minor offsets of T$_{\rm eff, Fe}$ from the reference IRFM $\Teff$ scale, with an rms error of the order of $\sim 50$ K. Assuming LTE or decreasing the efficiency of inelastic collisions in the NLTE statistical equilibrium calculations by an order of magnitude, $\SH = 0.1$, leads to larger discrepancies. The latter result is very important. At low metallicity NLTE effects in are extremely sensitive to the magnitude of collisional thermalization by . Our analysis, thus, not only confirms that it is possible to constrain the absolute magnitude of atomic collision cross-sections using metal-poor stars, but also demonstrates that a single scaling factor (unity in our case) to the classical Drawin’s formula is a reasonable approximation. What concerns the LTE $\Teff$ values, they are still within the errors of T$_{\rm
eff, REF}$ in the 1D case. We caution, however, that the NLTE effects for our stars are small in 1D, and this statement may not apply for stellar parameters where the NLTE effects are more significant, i.e. for stars with higher $\Teff$ and lower $\log g$. The discrepancy between the LTE spectroscopic and the IRFM scale increases to $200$ K for the metal-poor stars when $\td$ models are employed, which is consistent with the discussion in the previous section.
Our second approach, in which the $\Teff$ values are kept fixed and surface gravities are obtained by forcing Fe ionization equilibria, returns similar results. Fig. \[teff\_logg\] (bottom panel) compares the spectroscopic $\log
g$’s with the values determined from parallaxes, the shaded regions indicate the uncertainty of the latter (see Sect. \[sec:obs\]) and the error bars of the former account for the uncertainties in $\log gf$, IRFM effective temperatures, and observational uncertainties. As seen from this figure, only the NLTE approach with $\SH = 1$ gives $\log g$ values on average consistent with that inferred from parallaxes. In $\td$, LTE gravities are systematically too low. In 1D, the discrepancy is smaller, but appears to increase with $\Teff$. The assumption of $\SH = 0.1$ brings the abundances so high, that much larger gravities are generally needed to match with .
### Comparison with other studies
Korn et al. (2003) performed a study similar to ours aiming to constrain the efficiency of inelastic collisions from the NLTE 1D Fe calculations for a small sample of late-type stars. Using the <span style="font-variant:small-caps;">mafags-os</span> model atmospheres and the Drawin’s formulae for collision cross-sections, they determined $\SH = 3$ as a scaling factor to the latter yielding satisfactory ionization balance of and . For the stars in common, HD 140283, HD 84937, and Procyon, their results are consistent with our values for $\SH = 1$. This study was, however, later superseded by that of Mashonkina et al. (2011), in which the Fe model atom of , also used by Korn et al. (2003), was updated by more recent atomic data. This work, in fact, showed that a lower efficiency of inelastic collisions ($\SH = 0.1$) is needed to satisfy ionization balance of Fe. Their results for HD 84937 and HD 122563 are in agreement with our values within the error bars: $-2.00 \pm 0.07$ (Fe I) respectively $-2.08
\pm 0.04$ (Fe II) for HD 84937, and $-2.61 \pm 0.09$ (Fe I) respectively $-2.56
\pm 0.07$ (Fe II) for HD 122563. They also find that the NLTE -based abundances are correlated with line excitation potential, with a slope of $0.013$ dex eV$^{-1}$ for HD 84937 and $-0.054$ dex eV$^{-1}$ for HD 122563. From Fig. \[st34\], we can not confirm their slope for HD 84937 with 1D models, however, we would get exactly same results, adopting Mashonkina et al. linelists. Their positive slope for HD 84937 would likely had vanished had they included few more subordinate near-UV lines with very accurate $gf$-values in their analysis. For HD 122563, however, we also find a small discrepancy between the high-excited and lines, in addition to residual slopes of the NLTE Fe abundances with $\Elow$: $-0.11$ dex eV$^{-1}$ for the <span style="font-variant:small-caps;">mafags-os</span>, $-0.09$ dex eV$^{-1}$ for <span style="font-variant:small-caps;">marcs</span>, and $-0.05$ dex eV$^{-1}$ for the $\td$ model. One of the important differences between our and their study, which also explains our more extreme slopes in 1D, is the way we handle microturbulence. We use lines, whereas Mashonkina et al. (2011) rely on the lines. Nevertheless, the fact that our results obtained with 1D model atmospheres support the findings of Mashonkina et al. (2011) is reassuring.
### HD 122563
[\[sec:hd122563\]]{}
HD 122563 is one of the most well-studied metal-poor giants. Nevertheless, its parameters are badly constrained. For example, Fulbright (2000) gives $\Teff =
4425$ K, $\log g = 0.6$, and \[Fe/H\] $=-2.6$, whereas according to Mashonkina et al. (2011) the star is better described by a model with $\Teff = 4600$ K, $\log
g = 1.6$, and \[Fe/H\] $\approx -2.58$.
At $\SH = 1$, no combination of parameters can achieve a fully satisfactory spectroscopic solution for this star, consistently achieving adequate excitation and ionization balance. This is evident from the residual offset between the high-excitation and lines in Fig. \[st56\]. In NLTE with $\SH = 1$, the spectroscopic $\log g$ is $0.3$ dex lower than the value derived from the Hipparcos parallax. The discrepancy can be minimized by adopting less efficient hydrogen collisions, which, however, contradicts the spectroscopic results of the other metal-poor stars.
Based on HD122563 alone, the evidence for a lower $\SH$ is thus limited. Similar investigation of larger samples of metal-poor giants and dwarfs with well-defined parameters may help to better clarify the situation.
Comparison with evolutionary tracks
-----------------------------------
[\[sec:evolution\]]{}
In an attempt to assess the trustworthiness of the spectroscopic results versus those obtained from independent measurements, we compared them to evolutionary tracks. The position of the four metal-poor stars in the $\Teff$-$\log g$ plane is shown in Fig. \[tracks\]. The black cross with error bars indicated is the result with the IRFM $\Teff$ and astrometric surface gravity, and the blue square and red diamond correspond to our spectroscopic $\td$ NLTE $\Teff$ and $\log g$, respectively. For comparison, evolutionary tracks with close parameters computed with the <span style="font-variant:small-caps;">garstec</span> code are shown. These were kindly made available to us by A. Serenelli (2011, private communication). In order not to overload the figures, we show only the NLTE $\td$ spectroscopic results, while briefly discussing other solutions from the Table \[tab:opt\_param\] below.
For all metal-poor stars, a combination of astrometric gravity and spectroscopic $\Teff$/\[Fe/H\], obtained with NLTE $\SH = 1$ or LTE with 1D and $\td$ models, appears to be most consistent with stellar evolution calculations. The error bars of our data are omitted in the plots for clarity (they are given in the Table \[tab:opt\_param\]). 1D LTE $\Teff$’s are somewhat larger than the 1D NLTE results, that makes the stars a bit more massive. It is interesting that even with $\td$ LTE effective temperatures, which are $\sim 50 - 300$ K larger compared to NLTE results, the stars still match the tracks with adequate mass and metallicity, which, in fact, also implies more realistic ages.
A combination of IRFM $\Teff$ and the spectroscopic $\td$ LTE $\logg$/\[Fe/H\] makes the agreement with stellar evolution less satisfactory. In particular, HD 140283 is now in the ’forbidden’ zone of the HRD. Situation is even worse for the NLTE $\SH = 0.1$ results, obtained with 1D and with $\td$ model atmospheres. In this case, all four metal-poor stars are some $\sim 2 - 3 \sigma$ in $\Teff$ and $\logg$ away from any reasonable track.
HD 122563 is a somewhat outstanding object. Assuming the tracks to be realistic, it also becomes clear that the IRFM $\Teff$ and/or luminosity of the star inferred from its relatively accurate parallax are too low. Other attempts, such as variation of the mass or decreasing the $\alpha$-enhancement, either do not have any effect or make the discrepancy with masses and ages even larger. In this respect, it is encouraging that our spectroscopic $\td$ NLTE values of $\Teff$ and $\log g$ bring the star into a much better agreement with stellar evolution, although these value appear to be shifted away from the presumably more reliable IRFM temperature and astrometric gravity.
Summary and conclusions
=======================
[\[sec:conclusions\]]{} One important problem in spectroscopy of late-type stars is to quantify systematic errrors in the determination of $\Teff$, $\log g$, and \[Fe/H\] from the / excitation-ionization equilibrium due to the assumptions of LTE and hydrostatic equilibrium in the models.
We have constructed a new extensive NLTE Fe model atom using the most up-to-date theoretical and experimental atomic data available so far. The model has been applied to the analysis of the Sun and a number of well-studied late-type stars with parameters determined by other independent methods. We have used standard 1D LTE <span style="font-variant:small-caps;">marcs</span> and <span style="font-variant:small-caps;">mafags-os</span> model atmospheres, as well as average stratifications computed from full time-dependent, 3D, hydrodynamical simulations of stellar surface convection. In addition, we compared the commonly-used NLTE line formation program packages <span style="font-variant:small-caps;">detail</span>/<span style="font-variant:small-caps;">siu</span> and <span style="font-variant:small-caps;">multi</span>. Despite obvious differences in their numerical scheme and input physical data, we find that the final results are consistent both in terms of the NLTE statistical equilibria and absolute LTE and NLTE abundances.
Our LTE and NLTE results for the 1D models are similar to most of the previous findings in the literature. Statistical equilibrium of Fe, which is a minority ion at the typical conditions of these cool and dense atmospheres, favors lower number densities of compared to LTE. The number densities of are hardly affected by NLTE. In general, this leads to a weakening of lines compared to LTE, which, in turn, requires larger Fe abundance to fit a given observed spectral line. The magnitude of departures from LTE depends on stellar parameters. With 1D hydrostatic model atmospheres the NLTE corrections on lines do not exceed $0.1$ dex for stars with \[Fe/H\] $> -3.5$, and they are negligible for the lines. The situation changes dramatically for the mean 3D ($\td$) models with their cooler surfaces, and, thus, steeper T$(\tau)$ relations. The NLTE abundance corrections can be as large as $\sim 0.5$ dex for the resonance lines formed in very metal-poor atmospheres. In contrast to LTE, NLTE strengths of the lines predicted by the $\td$ and 1D models are rather similar, because the line formation is largely dictated by the radiation field forming around the optical surface, $\opd \sim 0$, where the thermal structures of $\td$ and 1D models are similar. This suggests that the full 3D NLTE results, once they become feasible, will be even closer to our $\td$ NLTE calculations.
The solar Fe abundance obtained under NLTE using the $\td$ model atmospheres is $7.46 \pm 0.06$ dex, and the uncertainty is fully dominated by the errors of the experimental transition probabilities. Adopting LTE or 1D hydrostatic model atmospheres lowers the solar Fe abundance by $\sim 0.02$ dex. For Procyon we infer slightly sub-solar metallicity, \[Fe/H\] $= -0.03$ dex.
We find that the / ionization balance can be well established for all reference stars with the 1D and $\td$ model atmospheres and the NLTE model atom with Drawin’s collision cross-sections. Strong resonance and subordinate lines are very sensitive to the atmospheric structure, thus, classical 1D models fail to provide consistent excitation balance. A better agreement among lines spanning a range of excitation potential is obtained with the mean 3D models, although the optimum solution, which fully eliminates the correlation of the abundance and line excitation potential, may necessitate self-consistent NLTE modelling with full 3D hydrodynamical model atmospheres. The assumption of LTE in combination with $\td$ models leads to large errors in $\Teff$ and $\log g$ inferred from the Fe ionization balance, yet we find satisfactory results for certain combinations of stellar parameters using standard hydrostatic model atmospheres.
The mean NLTE metallicities determined with 1D and $\td$ models are in agreement in the range of stellar parameters investigated here, although there are marked residual slopes of abundance with line excitation potential and strength. We can, thus, conclude that accurate, albeit not precise, metallicities for late-type stars can be obtained with classical hydrostatic 1D model atmospheres if NLTE effects in are taken into account and a sufficiently large number of and lines with different excitation potentials is used, so that individual line-to-line abundance discrepancies cancel out. An alternative solution, which avoids additional biases introduced by adjusting microturbulence parameter, is to use high-excitation lines only. Thus, the results obtained with classical 1D models depend on the choice of the line list and, in particular, on the balance of the number of lines of different types. A combination of $\td$ and NLTE alleviates this problem.
Our NLTE effective temperatures and gravities are consistent with the parameters determined by other less model-dependent methods, in particular with IRFM $\Teff$’s and $\log g$’s inferred from parallaxes. The results for the metal- poor dwarfs and subgiant are also consistent with stellar evolution predictions. For HD 122563, the results are inconclusive. In addition to a residual slope of -based abundances with line excitation potential, the NLTE $\td$ spectroscopic gravity is roughly $0.3$ dex lower than the astrometric result. On the one side, the latter is somewhat uncertain due to the difficulty of determining the mass of this metal-poor giant. Furthermore, comparisons with evolutionary tracks favor higher $\Teff$ and lower $\log g$ which would be more consistent with our spectroscopic estimates. On the other side, the discrepancies might be indicative of remaining systematic uncertainties in the stellar atmosphere models, such as neglect of convective inhomogeneities and atmospheric extension, which become increasingly important with increasing $\Teff$ and decreasing $\log g$. The importance of the latter is also amplified at higher metallicities.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work is based on observations collected at the European Southern Observatory, Chile, 67.D-0086A and ESO DDT Program ID 266.D-5655 (PI Christlieb). We made extensive use of the NIST and Kurucz’s atomic databases. We acknowledge valuable discussions with Lyudmila Mashonkina, and are indebted to Thomas Gehren for providing the codes for calculations performed in this work. We thank Jorge Melendez for providing the observed spectra for some of the studied stars. We are grateful to Manuel Bautista for providing the photoionization cross-sections for Fe and to Aldo Serenelli for providing the <span style="font-variant:small-caps;">garstec</span> evolutionary tracks.
\[lastpage\]
[^1]: E-mail: [email protected]
[^2]: $H_p$ is the pressure scale height and $l$ the so-called ’mixing length’
[^3]: http://marcs.astro.uu.se/index.php
[^4]: http://kurucz.harvard.edu/atoms.html
[^5]: For a typical F-type star ($\Teff = 6000$ K), the thermal energy is $kT \sim 0.4$ eV.
[^6]: We follow the definition of [@1972SoPh...23..265W], in which $b_i$ is a ratio of NLTE to LTE atomic level populations, $b_i = n_i^{\rm NLTE}/n_i^{\rm LTE}$.
[^7]: Net radiative brackets $\rho$ are defined as $\rho R_{ci}n_{c}
= R_{ci}n_{c} - R_{ic}n_{i}$, where $R_{ci}$ are rate coefficients for radiative transitions from the continuum $c$ to a bound level $i$ of an atom and $n_{c}$ are atomic level populations .
[^8]: Note that the uncertainty in mass is nominal and was determined by comparing positions of the stars on the HRD with theoretical isochrones.
[^9]: Note that it is not possible to assess systematic errors in $\log gf$’s.
[^10]: NLTE abundance correction, $\Delta_{\rm Fe} = \log \rm{A (Fe)}_{\rm NLTE} - \log \rm{A (Fe)}_{\rm LTE}$, is the difference between the NLTE and LTE abundances required to fit a line with a given equivalent width
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'Beam-helicity asymmetries have been measured at the MAMI accelerator in Mainz in the three isospin channels $\vec{\gamma}p\rightarrow \pi^{+}\pi^0n$, $\vec{\gamma}p\rightarrow \pi^{0}\pi^0p$ and $\vec{\gamma}p\rightarrow \pi^{+}\pi^{-}p$ . The circularly polarized photons, produced from bremsstrahlung of longitudinally polarized electrons, were tagged with the Glasgow magnetic spectrometer. Charged pions and the decay photons of $\pi^0$ mesons were detected in a $~4\pi$ electromagnetic calorimeter which combined the Crystal Ball detector with the TAPS detector. The precisely measured asymmetries are very sensitive to details of the production processes and are thus key observables in the modeling of the reaction dynamics.'
author:
- |
D. Krambrich$^1$, F. Zehr$^2$, A. Fix$^{18}$, L. Roca$^{19}$, P. Aguar$^{1}$, J. Ahrens$^{1}$, J.R.M. Annand$^{3}$, H.J. Arends$^{1}$, R. Beck$^{1,4}$, V. Bekrenev$^{5}$, B. Boillat$^{2}$, A. Braghieri$^{6}$, D. Branford$^{7}$, W.J. Briscoe$^{8}$, J. Brudvik$^{9}$, S. Cherepnya$^{10}$, R. Codling$^{3}$, E.J. Downie$^{3}$, P. Dexler$^{11}$, D.I. Glazier$^{7}$, P. Grabmayr$^{12}$, R. Gregor$^{11}$, E. Heid$^{1}$, D. Hornidge$^{13}$, O. Jahn$^{1}$, V.L. Kashevarov$^{10}$, A. Knezevic$^{14}$, R. Kondratiev$^{15}$, M. Korolija$^{14}$, M. Kotulla$^{2}$, B. Krusche$^{2}$, A. Kulbardis$^{5}$, M. Lang$^{1,4}$, V. Lisin$^{15}$, K. Livingston$^{3}$, S. Lugert$^{11}$, I.J.D. MacGregor$^{3}$, D.M. Manley$^{16}$, M. Martinez$^{1}$, J.C. McGeorge$^{3}$, D. Mekterovic$^{14}$, V. Metag$^{11}$, B.M.K. Nefkens$^{9}$, A. Nikolaev$^{1,4}$, P. Pedroni$^{6}$, F. Pheron$^{2}$, A. Polonski$^{15}$, S.N. Prakhov$^{9}$, J.W. Price$^{9}$, G. Rosner$^{3}$, M. Rost$^{1}$, T. Rostomyan$^{6}$, S. Schumann$^{1,4}$, D. Sober$^{17}$, A. Starostin$^{9}$, I. Supek$^{14}$, C.M. Tarbert$^{7}$, A. Thomas$^{1}$, M. Unverzagt$^{1,4}$, Th. Walcher$^{1}$, D.P. Watts$^{7}$\
(The Crystal Ball at MAMI, TAPS, and A2 Collaborations)
title: 'Beam-Helicity Asymmetries in Double Pion Photoproduction off the Proton'
---
Double pion photoproduction allows the study of sequential decays of nucleon resonances via intermediate excited states, as well as the coupling of nucleon resonances to $N\rho$ and $N\sigma$. It has therefore become an attractive tool for the study of the excitation spectrum of the nucleon, which is intimately connected to the properties of QCD in the non-perturbative range. Its contribution to the total electromagnetic response of the nucleon is substantial. In the second resonance region, comprising the $P_{11}$(1440), $S_{11}$(1535), and $D_{13}$(1520) resonances, roughly 50% of the total photoabsorption cross section originates from it. In this energy region total cross sections and invariant mass distributions of the $\pi\pi$- and the $\pi N$-pairs have been measured with the DAPHNE and TAPS detectors at the MAMI accelerator in Mainz [@Braghieri_95; @Haerter_97; @Zabrodin_97; @Zabrodin_99; @Wolf_00; @Kleber_00; @Langgaertner_01; @Kotulla_04], at GRAAL in Grenoble (also linearly polarized beam asymmetry) [@Assafiri_03; @Ajaka_07], with the CLAS detector at JLab (electroproduction) [@Ripani_03], and at ELSA in Bonn [@Thoma_07; @Sarantsev_07]. More recently, also polarization observables have been measured at the MAMI accelerator [@Ahrens_03; @Ahrens_05; @Ahrens_07] and at the CLAS facility at Jlab [@Strauch_05].
In spite of all these efforts, even the interpretation of the data in the energy region, where only few resonances can contribute, is still surprisingly controversial [@Krusche_03] since the available data do not sufficiently constrain the model analyses. It is thus evident that the search for missing resonances at higher energy requires a better understanding of the reaction mechanisms. The controversy has far reaching consequences not only for the N$^{\star}$ excitation spectrum itself, but as discussed below also in the field of the much discussed hadron in-medium properties.
There is agreement that the $\pi^+\pi^-$ final state is dominated by background terms in particular of the $\Delta$-Kroll-Rudermann type, while $\pi^0\pi^0$ has only small background contributions and thus is particularly suited for the study of sequential resonance decays. However, even for the latter the results of different reaction models are contradictory. Calculations by the Valencia group [@Gomez_96; @Nacher_01; @Nacher_02] emphasize a large contribution from the $D_{13}\to \Delta\pi^0\to p\pi^0\pi^0$ decay. Laget and coworkers [@Assafiri_03], instead find a dominant contribution from the $P_{11}$(1440)$\rightarrow N\sigma$ decay and a recent analysis by the Bonn-Gatchina group [@Thoma_07; @Sarantsev_07] reports a strong contribution from the $D_{33}$(1700) resonance, which is not seen in other models. Modifications of the invariant mass distributions of the $\pi^0$ pairs for photoproduction off heavy nuclei have been discussed in view of the predicted $\sigma$ in-medium modification resulting from partial chiral symmetry restoration [@Messchendorp_02; @Bloch_07], however, a better understanding of the elementary production processes is obligatory. Similarly, for the mixed charge channel $n\pi^+\pi^0$ all early model calculations (see e.g. [@Gomez_96]) failed already in the reproduction of the total cross section. Only the introduction of a strong contribution from the $\rho$ meson [@Nacher_01; @Nacher_02; @Fix_05], motivated by the shape of the measured invariant mass distributions [@Zabrodin_99; @Langgaertner_01], improved the situation. Again, a close connection to a different problem, namely the still unexplained strong suppression of the second resonance bump in photoproduction off nuclei (see e.g. [@Bianchi_94]) is involved, where a possible explanation might arise from the in-medium modification of the $D_{13}(1520)\rightarrow N\rho$ decay [@Langgaertner_01].
Recently, model predictions [@Nacher_02; @Roberts_05; @Fix_05; @Roca_05], which indicated that polarization observables are extremely sensitive for the disentanglement of the reaction mechanisms, have triggered wide-spread experimental activities. The advent of accelerators with highly polarized electron beams has provided a new tool for this field: meson photoproduction using circularly polarized photons. They are produced by the bremsstrahlung of longitudinally polarized electrons in an amorphous radiator. The polarization transfer obeys a simple formula given by Olsen and Maximon [@Olsen_59]. The beam helicity asymmetry can then be measured by comparing the event rates for the two helicity states of the beam. Parity conservation precludes any sensitivity of the cross section in a two-body reaction to beam helicity alone, but in a reaction with three or more particles in the final state, circularly polarized photons can lead to asymmetries even for an unpolarized target. Until recently there was little effort to study these effects until two experimental programs at JLab observed strong signals. In hyperon photoproduction, the decay of the final state $\Lambda$ or $\Sigma$ hyperon has an angular dependence on the hyperon polarization, and a recent experiment [@Bradford_07] has shown that the polarization transfer along the photon momentum axis is nearly 100%. In an analysis of charged double-pion production $\gamma p\to p\pi^+\pi^-$ measured with CLAS, Strauch [*et al.*]{} [@Strauch_05] found a large helicity asymmetry in the distribution of $\Phi$, the angle between the two-pion plane and the $\gamma p$ reaction plane (see Fig. \[fig:asym\]).
The Crystal-Ball-TAPS collaboration at the Mainz microtron MAMI [@Walcher_90] has recently taken data on the photoproduction of the three $\pi\pi N$ final states accessible with a proton target: $\gamma p\to p\pi^+\pi^-$, $\gamma p\to p\pi^0\pi^0$, and $\gamma p\to n\pi^+\pi^0$, using circularly polarized photons. This Letter presents the beam-helicity asymmetries in a form similar to that of Strauch [*et al.*]{} The data were taken with tagged photons incident on a 4.8 cm long liquid hydrogen target (surface density 0.201 nuclei/barn). Contributions from the target windows (2$\times$60 $\mu$m Kapton) were determined with empty target measurements and subtracted. The photons of up to 820 MeV, were produced by the bremsstrahlung of 883 MeV longitudinally polarized electrons. The energy of the photons was determined by the Glasgow photon tagger [@Anthony_91; @Hall_96] with a resolution of approximately 2 MeV full width. The target was located inside the Crystal Ball (CB) [@Starostin_01], consisting of 672 NaI crystals that covered the full azimuthal range for polar angles between 20$^{\circ}$ - 160$^{\circ}$. The angular region from 20$^{\circ}$ down to 1$^{\circ}$ was covered by the TAPS detector [@Novotny_91; @Gabler_94] with 510 BaF$_2$ crystals arranged as a hexagonal wall. The target was surrounded by a [**P**]{}article [**I**]{}dentification [**D**]{}etector (PID) [@Watts_04] and two cylindrical multiple wire proportional chambers (MWPC) [@Audit_91]. Protons and charged pions hitting the CB were identified by an $E-\Delta E$ analysis, using the energy information of the CB and the PID. For TAPS, the separation of photons, neutrons, protons, and charged pions can be achieved in principal as discussed in [@Bloch_07]. However, here these methods were only used for a clean identification of photons. Protons and charged pions in TAPS were not included in the analysis, since their separation was less clean than the $E-\Delta E$ analysis by PID and CB.
\
In the first step of the reaction identification for the $p\pi^0\pi^0$ final state, events with four photons and one or no proton candidate were selected. Similarly, for the $n\pi^+\pi^0$ final state two photons, a $\pi^+$, and one or no neutron were required. The $\pi^0$ mesons where then identified by a standard invariant mass analysis. Further identification of the reactions was based on missing mass analyses for the recoil nucleons in a manner similar to that described in [@Wolf_00; @Kleber_00; @Langgaertner_01; @Kotulla_04]. It was used to remove small residual background from $\eta\to 3\pi$ decays, which however, was much less important than in previous experiments, since due to the large solid angle coverage, in most cases the third pion was also seen. For both reactions, the recoil nucleons were treated as missing particles, no matter whether a candidate was found or not, so that the results are independent of the detector acceptance and efficiency for recoil nucleons. This event selection guaranteed full solid angle coverage for the $\pi^0\pi^0$ channel. For the $\pi^+\pi^0$ reaction only events with the $\pi^+$ at laboratory polar angles smaller than 20$^{\circ}$ (and larger than 160$^{\circ}$) were excluded, which has a negligible effect on the measured asymmetries. Since identification of the double charged channel is missing the powerful tool of invariant mass analysis, in this case detection of all three charged particles was required in order to achieve an equally good background suppression as for the other channels (residual background from $\gamma p\rightarrow p\pi^+\pi^-\pi^0$ was again removed with a missing mass analysis). This selection means that for the double charged channel only events with all three particles at laboratory polar angles larger than 20$^{\circ}$ (and smaller than 160$^{\circ}$) were accepted. This limitation was accounted for in the model calculations.
The missing mass spectra for all three reactions were extremely clean and very well reproduced by Monte Carlo simulations. Residual background was estimated at maximum at the few per cent level (certainly well below 5 % for all channels) and is thus not relevant for any results presented here. Details of the analysis will be discussed in an upcoming paper about total cross sections and invariant mass distributions.
In a reaction produced by circularly polarized photons on an unpolarized target the beam-helicity asymmetry $I^{\odot}$ is defined by: $$I^{\odot}(\Phi)=\frac{1}{P_{\gamma}}
\frac{d\sigma^{+}-d\sigma^{-}}{d\sigma^{+}+d\sigma^{-}}
=\frac{1}{P_{\gamma}}
\frac{N^{+}-N^{-}}{N^{+}+N^{-}}
\label{eq:circ}$$ where $d\sigma^{\pm}$ is the differential cross section for each of the two photon helicity states, and $P_{\gamma}$ is the degree of circular polarization of the photons. The latter is calculated as product of the polarization degree of the longitudinally polarized electrons (82$\pm$5)% and the photon-energy-dependent polarization transfer factor [@Olsen_59]. In the energy range of interest, $P_{\gamma}$ was between 60% and 80%. Possible differences in the number of incident photons for the two helicity states have been determined to be at the 5$\times$10$^{-4}$ level, i.e. they are negligible. The angle $\Phi$ between reaction and production plane is calculated as defined in the work of Roca [@Roca_05] from the three-momenta of the particles (the same construction was used for the analysis of the CLAS-data [@Strauch_05]). For $\pi^+\pi^0$ production the two pions are ordered as shown in Fig. \[fig:asym\]. For double $\pi^0$ production and for the double charged state their assignment is randomized since the experiment cannot distinguish positively and negatively charged pions. This means that for the latter two $I^{\odot}(\Phi)=I^{\odot}(\Phi +\pi)$.
For the extraction of the asymmetry $I^{\odot}(\Phi,\Theta_{\pi_1},\Theta_{\pi_2},...)$ in a limited region of kinematics, the differential cross sections $d\sigma^{\pm}$ can be replaced by the respective count rates $N^{\pm}$ (right hand side of Eq. \[eq:circ\]), since all normalization factors cancel in the ratio. In principle efficiency weighted count rates ought to be used to obtain the angle integrated asymmetries. For the two final states $\pi^0\pi^0$ and $\pi^+\pi^0$, for which also precise total cross sections and invariant mass distributions will be published elsewhere, the detection efficiency was modeled with Monte Carlo simulations. However, since the efficiencies are rather flat functions of the pion polar angles, the effect of the efficiency corrections on the asymmetries was negligible. As in the CLAS experiment [@Strauch_05] only the raw asymmetries are given for $\pi^+\pi^-$ production.
The measured asymmetries are summarized in Figs. \[fig:pi\_char\], \[fig:pi\_neu\] as functions of $\Phi$. Parity conservation enforces $I^{\odot}(\Phi)=-I^{\odot}(2\pi-\Phi)$. This condition was not used as a constraint in the analysis but is very well respected, demonstrating the excellent quality of the data.
The asymmetries are compared to the results from the model of Fix and Arenhövel [@Fix_05] and the Valencia model [@Roca_05], which were calculated taking into account the acceptance limitations for $\pi^+\pi^-$ and the fact that $\pi^-$ could not be distinguished from $\pi^+$ in the detectors. For this channel also the prediction of the Valencia model for full 4$\pi$ acceptance is shown. At least in the framework of the model, the effect from the acceptance limitation is small. A similar result as in the CLAS experiment [@Strauch_05] is found. The two models make similar predictions, but agree with the measurements only in the energy range around 715 MeV. For $n\pi^0\pi^+$, the model results are similar above 700 MeV, but are nowhere in agreement with the data. For the Valencia model [@Roca_05] also the solution without the $D_{13}\to N\rho$ contribution is shown. It was introduced into the model [@Nacher_01; @Nacher_02] in order to reproduce the previously non-understood total cross section and pion invariant mass distributions [@Zabrodin_99; @Langgaertner_01]. However, in the $D_{13}$ range, inclusion of this contribution does not at all improve the agreement with the asymmetries. Finally, for $p\pi^0\pi^0$ Fix’s model and the Bonn-Gatchina analysis (not available for the other iso-spin channels) [@Thoma_07; @Sarantsev_07] are in fairly good agreement with the data, while the Valencia model is out of phase.
Due to its symmetry $I^{\odot}(\Phi)$ can be expanded into a sine-series (odd coefficients vanish for identical pions): $$I^{\odot}(\Phi)=\sum_{n=1}^{\infty}A_{n}\mbox{sin}(n\Phi)
\label{eq:series}$$ The data have have been fitted to Eq. \[eq:series\] for $n\leq 4$ (higher orders were not significant), and the results are summarized in Fig. \[fig:coeff\]. For the $p\pi^0\pi^0$ and $p\pi^+\pi^-$ final states the results for the odd terms $A_1$ and $A_3$ are consistent with zero, which is additional evidence that no false asymmetries have been generated in the experiment. For comparison the CLAS results [@Strauch_05] for $p\pi^+\pi^-$ have also been fitted. Since in the CLAS experiment negatively and positively charged pions were distinguished, the odd terms can also contribute, but the even terms $A_2$ and $A_4$ can be compared to the present results. One must, however, keep in mind that the acceptance was not identical (CLAS covered polar angles down to 8$^{\circ}$, this experiment down to 20$^{\circ}$). For $A_2$ the energy dependence is similar, although the present values are somewhat larger. No significant contribution from $n=4$ was found for the CLAS experiment, but in the present measurement it contributes up to 2%. The comparison of the results for the three final states highlights the different reaction mechanisms in the three isospin channels.
In summary, precise measurements of the beam-helicity asymmetry for double pion photoproduction on the proton have been presented for all three isospin channels. The comparison with model predictions highlights, both the challenges and potential rewards for the extraction of resonance properties. On the one hand, the progress in experimental techniques allows precise measurements of this observable, and the predictions for it are very sensitive to the internal mechanisms of the models. On the other hand, the general lack of agreement between experiment and theory signals that significant improvements in the models are needed. The present data can provide rigorous tests for future developments on the way to an eventual reliable extraction of resonance contributions from double pion photoproduction. The very precise results for the total cross sections, including sensitive measurements of the threshold behavior in view of the predictions of chiral perturbation theory, and the invariant mass distributions of pion - pion and pion - nucleon pairs, which have been extracted with a precision far superior to any previous measurements, will be presented in an upcoming paper.
We wish to acknowledge the outstanding support of the accelerator group and operators of MAMI. This work was supported by Deutsche Forschungsgemeinschaft (SFB 443, SFB/TR 16), DFG-RFBR (Grant No. 05-02-04014), Schweizerischer Nationalfonds, UK Engineering and Physical Sciences Research Council, European Community-Research Infrastructure Activity (FP6), the US DOE, US NSF and NSERC (Canada). We thank the undergraduate students of Mount Allison and George Washington Universities for their assistence.
A. Braghieri [*et al.*]{}, Phys. Lett. B [**363**]{}, 46. (1995). F. Härter [*et al.*]{}, Phys. Lett. B [**401**]{}, 229 (1997). A. Zabrodin [*et al.*]{}, Phys. Rev. C [**55**]{}, R1617 (1997). A. Zabrodin [*et al.*]{}, Phys. Rev. C [**60**]{}, 055201 (1999). M. Wolf [*et al.*]{}, Eur. Phys. J. A [**9**]{}, 5 (2000). V. Kleber [*et al.*]{}, Eur. Phys. J. A[**9**]{}, 1 (2000). W. Langgärtner [*et al.*]{}, Phys. Rev. Lett. [**87**]{}, 052001 (2001). M. Kotulla [*et al.*]{}, Phys. Lett. B [**578**]{}, 63 (2004). Y. Assafiri [*et al.*]{}, Phys. Rev. Lett. [**90**]{}, 222001 (2003). J. Ajaka [*et al.*]{}, Phys. Lett. B [**651**]{}, 108 (2007). M. Ripani [*et al.*]{}, Phys. Rev. Lett. [**91**]{}, 022002 (2003). U. Thoma [*et al.*]{}, Phys. Lett. B [**659**]{}, 87 (2008). A.V. Sarantsev [*et al.*]{}, Phys. Lett. B [**659**]{}, 94 (2008). J. Ahrens [*et al.*]{}, Phys. Lett. B [**551**]{}, 49 (2003). J. Ahrens [*et al.*]{}, Phys. Lett. B [**624**]{} 173, (2005). J. Ahrens [*et al.*]{}, Eur. Phys. J. A [**34**]{} 11, (2007). S. Strauch [*et al.*]{}, Phys. Rev. Lett. [**95**]{}, 162003 (2005). B. Krusche and S. Schadmand, Prog. Part. Nucl. Phys. [**51**]{}, 399 (2003). J.A. Gomez-Tejedor and E. Oset, Nucl. Phys. A [**600**]{}, 413 (1996). J.C. Nacher [*et al.*]{}, Nucl. Phys. A [**695**]{}, 295 (2001). J.C. Nacher and E. Oset, Nucl. Phys. A [**697**]{}, 372 (2002). J.G. Messchendorp et al., Phys. Rev. Lett. [**89**]{} 222302 (2002). F. Bloch [*et al.*]{}, Eur. Phys. J. A [**32**]{}, 219 (2007). A. Fix and H. Ahrenhövel, Eur. Phys. J. A [**25**]{}, 115 (2005). N. Bianchi et al., Phys. Lett. B 3[**25**]{}, 333 (1994). W. Roberts and T. Oed, Phys. Rev. C [**71**]{}, 055201 (2005). L. Roca, Nucl. Phys. A [**748**]{}, 192 (2005). H. Olsen and L.C. Maximon, Phys. Rev. [**114**]{} 887, (1959). R.K. Bradford [*et al.*]{}, Phys. Rev. C [**75**]{}, 235205 (2007). Th. Walcher, Prog. Part. Nucl. Phys. [**24**]{}, 189 (1990). I. Anthony [*et al.*]{}, Nucl. Inst. and Meth. A [**301**]{}, 230 (1991). S.J. Hall, G.J. Miller, R. Beck, P.Jennewein, Nucl. Inst. Meth. A [**368**]{}, 698 (1996). A. Starostin [*et al.*]{}, Phys. Rev. C [**64**]{}, 055205 (2001). R. Novotny, IEEE Trans. Nucl. Sci. [**38**]{}, 379 (1991). A.R. Gabler [*et al.*]{}, Nucl. Instr. and Meth. A [**346**]{} 168 (1994). D. Watts (2004), 11. Int. Conf. on Calorimetry in High Energy Physics, Perugia, Italy. G. Audit et al., Nucl. Instrum. Methods A [**301**]{}, 473 (1991).
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
address: |
High Energy Physics Division\
Argonne National Laboratory\
Argonne IL 60439\
[E-mail: [email protected]]{}
author:
- MAURY GOODMAN
title: ' SEARCHING FOR THE NEUTRINO MIXING ANGLE ${\theta_{13}}$ AT REACTORS [^1] '
---
Introduction
============
Since 2003, there has been a worldwide effort to plan new reactor neutrino experiments using two or more detectors to measure or further limit the only unmeasured neutrino mixing angle, ${\theta_{13}}$. The best current limit on ${\theta_{13}}$ comes from the reactor experiment CHOOZ,[@bib:chooz] which ran in the 1990’s along with Palo Verde[@bib:palo] to determine if the atmospheric neutrino anomaly could be explained with $\theta_{12}$. Here I will describe some features and the current status of reactor neutrino experiments, which I expect to be the first to improve our knowledge of ${\theta_{13}}$ further. The reader whose only interest is in the status of current projects can skip to Section 3.2.
In a sense, reactor neutrino ${\bar{\nu}_e}$ disappearance experiments are complementary to the new off-axis accelerator $\nu_e$ appearance experiments, T2K & NO$\nu$A[@bib:t2k; @bib:nova], whose goal is also to study ${\theta_{13}}$. The magnitude of a ${\theta_{13}}$ signal at a new reactor neutrino experiment is affected only by the uncertainty of the value of ${\Delta m^2_{32}}$, which is currently bounded by $2.48 < {\Delta m^2_{32}}< 3.18 \times 10^{-3}eV^2/c^4$.[@bib:minos; @bib:trish] On the other hand, the ability of an accelerator experiment to measure ${\theta_{13}}$ is also affected by the uncertainty in $\theta_{23}$, $ 0.36 < \sin^2(\theta_{23}) < 0.63$[@bib:pdg], and the uncertainty in the CP violating phase $\delta$, $0 < \delta < 2\pi$. Thus a precise measurement of ${\theta_{13}}$ by both reactor and accelerator experiments could be used to constrain $\theta_{23}$ and/or $\delta$. On the other hand, a failure to find a value for ${\theta_{13}}$ by the reactor experiments will have a rather negative effect on the expected physics capabilities for the accelerator experiments which are much more expensive. For example, a limit of $\sin^2 (2 {\theta_{13}}) < 0.02$, which reactor experiments could achieve before T2K or NO$\nu$A start running, would mean that the accelerator experiments, even with increases in beam power, could not measure evidence for matter effects or CP violation and would only have a narrow window to find evidence for a non-zero ${\theta_{13}}$
The current limit on ${\theta_{13}}$ from the Chooz experiment is shown in Figure \[fig:chooz\]. The analysis was actually done for $\theta_{12}$, but given the value of $\Delta m^2_{21}$, it serves as a valid analysis for ${\theta_{13}}$. The curve shows the 90% CL allowed and prohibited values of ${\sin^2(2\theta_{13})}$ as a function of ${\Delta m^2_{32}}$. In order to compare experiments, it is common to quote a single number as the ${\theta_{13}}$ limit, but this requires a few assumptions. As a consequence, a large variety of numbers are quoted as the CHOOZ limit, such as ${\sin^2(2\theta_{13})}< 0.10, 0.11, 0.14, 0.15, 0.20$. One cause for this is the time-dependence of the ${\Delta m^2_{32}}$ measurement of Super-K and now MINOS. There is also no unique method of picking the value of ${\Delta m^2_{32}}$ to use. (The union of two CL curves is not a Confidence Level.) While the “best fit" ${\Delta m^2_{32}}$ value is often chosen, the PDG has elected to use the one sigma low value of ${\Delta m^2_{32}}$ where the larger value of ${\theta_{13}}$ is achieved, and they obtain ${\sin^2(2\theta_{13})}< 0.19$. Another more mundane issue which requires care is that, depending on the application, ${\theta_{13}}$ is often expressed in degrees, in radians, as $\sin({\theta_{13}})$, $\sin^2({\theta_{13}})$, $\sin^2(\theta_{\mu e})$, ${\sin^2(2\theta_{13})}$ and $U_{e3}$. The relationships between these expressions are simple, but factors of two errors are common. Finally, since comparison with the sensitivity of future accelerator experiments is often made, note that the accelerator experiments have additional ambiguities and degeneracies in interpreting a ${\theta_{13}}$ limit from $\theta_{23}$, $\delta$ and the mass hierarchy.[@bib:lindneramb]
The planning of a new generation of reactor neutrino experiments
================================================================
The ingredients for the design of a new reactor ${\theta_{13}}$ experiment have been laid out in detail in the white paper, “A new reactor neutrino experiment to measure ${\theta_{13}}$" prepared by an International Working Group comprised of 125 authors from 40 institutions in 9 countries[@bib:white]. The optimum location for the far detector depends on ${\Delta m^2_{32}}$, and also on the experiment’s ultimate exposure. Sites from 1.1 to 2.0 km have been chosen. The near detector needs to be located close to the core to measure the unoscillated spectrum. Local factors, such as reactor access and topological features modify where detectors will be placed. Important general features of reactor experiments are the effects of luminosity on the sensitivity, detector design, scintillator stability, calibration, backgrounds and systematic errors.
The neutrino oscillation sensitivity for a reactor neutrino experiment comes from measuring a smaller number of neutrinos than would be expected if ${\theta_{13}}=0$, and measuring an energy distribution consistent with ${\bar{\nu}_e}$ disappearance due to oscillations. These can be called the “rate" test and the “shape" test, but every experiment will use all available information. The effective “luminosity" for a reactor experiment can be expressed in GW-ton-years, or the product of the reactor’s thermal power times the size of the detector times the length of time the detectors operate. An example of how the sensitivity of an experiment varies with luminosity is given in Figure \[fig:lindner\]. Two extreme examples, which represent straight lines on this log-log plot, are for no systematic error, and for infinite systematic error in normalization and energy calibration. In the latter case, an oscillation signature is recognized by the appropriate wiggles in an energy distribution. Such a signal would be affected by bin-to-bin systematic errors, but not by the same systematic errors which limit the “rate" test. Two other curves are drawn with possibly realistic estimates of systematic error for the next round of experiments. Vertical lines are drawn at 12 GW-ton-years, corresponding to CHOOZ, 400 GW-ton-years, which could quickly and dramatically increase the world’s sensitivity, and a more ambitious project with 8000 GW-ton-years.
CHOOZ had a volume of Gd-loaded liquid scintillator, optically connected and surrounded by liquid scintillator without Gadolinium. New detector designs involve the addition of a third volume of mineral oil without scintillator, as shown in the Double Chooz design of Figure \[fig:choozdet\]. An inner volume of Gd loaded scintillator serves as a well-defined fiducial volume for neutrino interactions, with a very high neutron capture cross section. A second layer of scintillator, called the “$\gamma$-catcher", measures the energy of any photons from positron annihilation or neutron capture which escape the fiducial volume, and a third volume, or “buffer", shields the active volume from backgrounds originating in the rock or phototubes.
The scintillator in CHOOZ showed a degradation of its transparency over time, which resulted in a decrease of the light yield. Such a degradation would be unacceptable in a new experiment, particularly if it differed between two detectors. Suspicions concentrate on possible contaminants which may have caused the Gd to come out of solution, so that clean and robust liquid handling systems will be required to maintain good optical qualities. Newly developed scintillator formulations from groups at Heidelberg[@bib:dc] and Brookhaven[@bib:db] have shown that it is possible to satisfy the stability requirements for the long time periods needed.
Precise calibration will be necessary to ensure that the response of two or more detectors is identical. This will be accomplished by the introduction of radioactive sources that emit gammas, electrons, positrons and neutrons. Light flasher systems and lasers will be used to test the stability of photo-detectors. Cosmic ray muons will also be used, and particular cosmogenic nuclei, such as $^12$B also can be used to provide a calibration. An important consideration is that identical calibration systems be used for all detectors. One possibility that has been proposed is to use multiple and movable detectors, in order to increase the available information regarding cross-calibration.
The neutrino signature is a coincidence between a prompt positron annihilation and a delayed neutron capture with a mean life of 30 $\mu$s. There are two kinds of backgrounds: accidental ones where the two signals have different causes, and correlated backgrounds. Two important correlated backgrounds are fast neutrons, which can cause two signals separated by a typical neutron capture time, and $^9$Li, which can be created by spallation when a muon passes through the scintillator. The danger of $^9$Li is that it has a long decay time ($\sim$ 130 ms), and the decay leads to both a neutron and an electron, creating a signal much like a reactor neutrino. The long decay time makes it unrealistic to veto every throughgoing muon which might have been the cosmogenic source. While $^9$Li production has been measured, its dependence on the muon energy is poorly known, so predictions of the rates at a particular depth may not be accurate. All correlated backgrounds can be reduced by putting the detectors deep enough underground so that there are large overburdens, though this has a cost in civil construction.
Finally, it is necessary to reduce the systematic errors for counting reactor neutrino events below those that were achieved by CHOOZ and Palo Verde. The use of a second detector and the definition of the fiducial volume at the target/$\gamma$-catcher interface provide large reductions in systematic errors, and the experiments have to be careful that other effects do not limit them for their planned-for statistics. A comparison of the systematic error goals for Double Chooz and Daya Bay has been tabulated by Mention et al.[@bib:mention] and is presented in Table \[tab:syst\].
------------------------------ ---------------- -------------- ---------------- -------------- -------------- --------------
Error Description CHOOZ Double Chooz Daya Bay
No R&D R&D
[Absolute]{} [Absolute]{} [Relative]{} [Absolute]{} [Relative]{} [Relative]{}
Reactor
Production $\sigma$ 1.90 % 1.90 % 1.90 %
Core powers 0.70 % 2.00 % 2.00 %
Energy/fission 0.60 % 0.50 % 0.50 %
Solid angle 0.07 % 0.08 % 0.08 %
Detector
Detection $\sigma$ 0.30 % 0.10 % 0.10 %
Target Mass 0.30 % 0.20 % 0.20 % 0.20 % 0.20 % 0.02 %
Fiducial volume 0.20 %
Target % H 0.80 % 0.50 % ? 0.20 % 0.10 %
Dead time 0.25 %
Analysis
$e^+$ escape 0.10 %
$e^+$ identification 0.80 % 0.10 % 0.10 %
n escape 0.10 %
n capture % Gd 0.85 % 0.30 % 0.30 % 0.10 % 0.10 % 0.10 %
n identification 0.40 % 0.20 % 0.20 % 0.20 % 0.20 % 0.10 %
${\bar{\nu}_e}$ time cut 0.40 % 0.10 % 0.10 % 0.10 % 0.10 % 0.03 %
${\bar{\nu}_e}$ distance cut 0.30 %
n multiplicity 0.50 % 0.05 % 0.05 %
[**Total**]{} [**2.72 %**]{} 2.88 % [**0.44 %**]{} 2.82 % 0.39 % 0.20 %
------------------------------ ---------------- -------------- ---------------- -------------- -------------- --------------
: Comparison of Systematic errors for the CHOOZ analysis and estimates of the relative and absolute errors in Double Chooz and Daya Bay, as tabulated by Mention et al. in Reference 19.
\[tab:syst\]
Sites for reactor experiments
=============================
Projects that were previously considered
----------------------------------------
Four reactor $\nu$ projects have received some funding and are moving forward. These four experiments, which will be described in the following sections, are Double Chooz, Daya Bay, RENO and Angra. As members of the International Working Group considered locations for new reactor experiments, a number of other possible locations were studied which have since been dropped. It is instructive to consider some of the strengths and weaknesses of sites that are not currently being pursued.
The first idea for a two-detector experiment to measure ${\theta_{13}}$ was KR2DET[@bib:kr2det]. This would have been built at the Krasnoyarsk reactor in Russia, which was originally built underground for producing weapons-grade plutonium. Two 46 ton detectors would have been 115 m and 1000 m from the reactor. Since the whole complex is underground, the 600 meters of water equivalent (m.w.e.) overburden would shield against a high rate of cosmogenic nuclei, such as $^9Li$, and the near and far detector would have the same low backgrounds. Unfortunately, local officials were not cooperative at the prospect of an international collaboration at their formerly secret soviet city.
The reactor complex in the United States with the highest power, and the site of a former reactor neutrino experiment, is Palo Verde, in Arizona. The previous collaboration had a poor decommissioning experience and the reactor company was not approached about a new project. There was a collaboration which did an extensive site study at the Diablo Canyon reactor on the coast of California. The hills there offered an opportunity for considerable overburden for both the near and far detector. However PG&E, the reactor power company, had recently gone through a politically motivated bankruptcy, and they decided not to cooperate with the collaboration after the initial studies. The most complete proposal in the United States was put forward by a large collaboration at the Braidwood reactor, about an hour’s drive from Chicago[@bib:braidwood]. Good cooperation with the Exelon Corporation was obtained after efforts from the Directors of Fermilab and Argonne. In Illinois, the overburden would need to be achieved by a vertical shaft rather than horizontal tunnels. Although the per-foot cost for a shaft is higher than a tunnel, the shaft height to reach a given overburden can be obtained with less digging than for the length of a typical mountain tunnel, and civil construction costs are comparable. An experiment was designed which could move two pairs of 65 ton detectors between two 180 m shafts about 1 km apart outside the reactor’s security fence. After consideration by the Neutrino Scientific Assessment Group (NuSAG)[@bib:nusag], the DOE decided not to fund the Braidwood experiment, presumably because it was more expensive than the alternative, which was support for U.S. participation in the Daya Bay project.
In Japan, a collaboration formed to prepare an experiment called KASKA at the Kashiwazaki-Kariwa complex south of Niigata[@bib:kaska]. With seven 3.4 GW$_{th}$ nuclear power plants, it is the world’s most powerful reactor complex. The plants are located in two small clusters, so two near detectors were planned. The absence of hard rock at the desired depth led to a design in which the detectors were placed in deep narrow shafts. However the economics of shafts limits their size and hence the size of the detector that could be placed in them. The collaboration developed an excellent relationship with the nuclear power company and conducted extensive boring studies to plan the shafts for 4.5 ton detectors. The Japanese funding agencies, which also support the KamLAND and Super-Kamiokande experiments, decided not to support this project.
Double Chooz in France {#sec:dc}
----------------------
Double Chooz will use the location of the CHOOZ experiment as its site for its far detector. By avoiding civil construction costs for the far site, Double Chooz will be less expensive and will be able to get started more quickly than the alternatives. There will be a near detector location 270 m from the middle of the two reactor cores, with an overburden of about 90 m.w.e., which conservatively maintains a similar signal to background as the far detector. Engineering for a near site has been provided by the French Electricity Company, Ed.F. The final design will be completed during 2007 and the lab will be ready in 2009.
The design for the three volume Double Chooz detector was shown in Figure \[fig:choozdet\]. A three volume prototype was built for the R&D stage, and the project is now entering the construction stage. Key parameters of the Double Chooz Experiment are given in Table \[tab:detsum\]. Initial tenders for the steel shielding and for the scintillator have already been issued, and the far detector will be installed and operated while the near detector lab is under construction. With just the 10 ton far detector, the CHOOZ limit on ${\theta_{13}}$ can be passed in a few months. When the near detector is operational, the full sensitivity can be reached quickly, as shown in Figure \[fig:choozsched\].
As the site of a former reactor neutrino experiment with extensive reactor off running, Chooz is one place where backgrounds have been measured. Accidental backgrounds in Double Chooz will be much lower than CHOOZ because sand will be replaced by 170 mm steel shielding, and because of the buffer. At the far detector, where 60 neutrinos per day will be measured, accidental backgrounds will be about 2 per day, while correlated backgrounds from fast neutrons will be an order of magnitude smaller. The estimate for $^9$Li is 1.4 per day, based on measurements in Chooz. The near detector, which should measure 1012 neutrino events per day, will have accidental backgrounds of about 22 per day, and 1.3 per day from fast neutrons. The estimate for $^9$Li is 9 per day. While Double Chooz will be both the cheapest experiment and the first to provide new knowledge on ${\theta_{13}}$, its lower ultimate sensitivity has been used by some funding agencies to deny it the resources that could have provided that knowledge in a more timely way.
--------------------------------------------- ----------------- -------------------------
Thermal power 4.27 GW each of 2 cores
Electric power 1.5 GWe each of 2 cores
${\bar{\nu}_e}$ target volume 10.3 m$^3$ Gd loaded LS (0.1%)
$\gamma$-catcher thickness 55 cm Gd-free LS
Buffer thickness 105 cm nonscintillating
Total liquid volume $\sim$237 m$^3$
Number of phototubes per detector 534 8["]{} 13% coverage
Far detector distance 1050 m averaged
Near detector distance 280 m averaged
Far detector overburden 300 m.w.e. hill topology
Near detector overburden 70–80 m.w.e. shaft
${\bar{\nu}_e}$ 5 years far detector events 75,000 with a 60.5% efficiency
${\bar{\nu}_e}$ 5 near detector events 789,000 with a 43.7% efficiency
Relative systematic error 0.6%
Effective bin-to-bin error 1% background systematics
Running time with far detector only 1–1.5 year
Running time with far+near detector 3 years
${\sin^2(2\theta)}$ goal in 3 years 0.02–0.03 (90% CL)
--------------------------------------------- ----------------- -------------------------
: Summary of the some parameters of the proposed Double Chooz experiment.[]{data-label="tab:detsum"}
Daya Bay in China
-----------------
The Daya Bay Complex, located near Hong Kong in Guangdong Province China, currently consists of two pairs of reactors, called Daya Bay and Ling Ao. The centers of each pair of reactors are about 1100 m apart. In addition, two more reactor cores near Ling Ao are under construction (Ling Ao II) and should be in operation by 2011, resulting in a total 17.4 GW$_{th}$ reactor power. With this geometry, two near detectors are needed to monitor the reactor power, as well as a far detector, as shown in Figure \[fig:daya1\]. Important factors for the near sites are the estimated muon induced backgrounds, which are a function of overburden. The near sites were optimized using a global $\chi^2$, which takes into account backgrounds, mountain profile, detector systematics and residual reactor related systematics. A summary of distances obtained is provided in Table \[tab:db\].
----------------- --------- --------- ------
Detectors DB near LA near far
Reactors (m) (m) (m)
DB cores 363 1347 1985
LA cores 857 481 1618
LA II cores 1307 526 1613
----------------- --------- --------- ------
: Distances between reactors and planned detectors at Daya Bay.
\[tab:db\]
The cylindrical Daya Bay detector will contain three zones, with a target, $\gamma$-catcher and buffer, as described above. The 224 phototubes will be located on the sides of each 20 ton detector, with reflective surfaces at the top and bottom. The multiple detectors at each site will be used to cross-calibrate each other, and the possibility of movable detectors is being studied. In Daya Bay’s baseline design, the detectors at each site are placed inside a large water buffer/water Cerenkov muon detector. For the far hall, this is similar to a swimming pool with dimensions 16 m $\times$ 16 m $\times$ 10 m (high). In addition, water tanks of 1 m $\times$ 1 m are used as an outer muon tracker.
![\[fig:daya3\] Daya Bay Detectors in the water buffer/Veto System](daya2.eps){width="8cm"}
![\[fig:daya3\] Daya Bay Detectors in the water buffer/Veto System](daya3.eps){width="4cm"}
The large depth of the Daya Bay detectors will be used to keep cosmogenic backgrounds at a small level. Currently, tests involving muon and neutron backgrounds are taking place with a number of detectors at the Aberdeen tunnel in Hong Kong, which has a similar overburden.
With the large reactor power and large overburden to reduce backgrounds, Daya Bay is an excellent choice for a reactor ${\theta_{13}}$ experiment. With support from the Chinese government and the U.S. Department of Energy, it is poised to be an excellent reactor experiment. With a baseline detector systematic error of 0.38% and a goal of 0.18%, they hope to take full advantage of the statistical uncertainty of 0.2%. Data taking with two near halls and far hall could begin in June 2010. With three years of running, Daya Bay will reach ${\sin^2(2\theta_{13})}< 0.008$ or better.
RENO in South Korea
-------------------
The South Korean Reactor Experiment for Neutrino Oscillation (RENO) collaboration is working on an experimental project at the YoungGwang reactor complex, which consists of six equally spaced reactors in a line on the west coast of South Korea. A schematic setup showing the topography and the proposed location of the near and far detectors is shown in Figure \[fig:skreno\]. The near detector at a distance of about 150 m would be under a 70 m high hill, and the far detector at a distance of 1.5 km would be under a 260 m high mountain.
After getting a \$9M funding approved by the government of Korea, the RENO collaboration[@bib:reno] has been undertaking detector design since May 2006. Various samples of liquid scintillator are under investigation with respect to the long-term stability of their optical properties. Other tests include compatibility with stainless steel and mylar and an acrylic cracks test. A RENO prototype contained 50 liters of Gd loaded scintillator with a 400 liter $\gamma$-catcher and a 60 cm $\times$ 100 cm stainless steel dark container. The prototype was used to do performance tests and background studies, R&D for the detector structure and phototube mounting scheme, and to establish a data analysis effort. The phototube layout in a simulation of a three-volume detector is shown in Figure \[fig:reno\]. Each detector would have a fiducial mass of 15.4 ton using scintillator with density 0.73 gm/cm$^3.$
RENO has received support from the South Korean government and good cooperation from the Y.K. Power Plant Company. The expected number of ${\bar{\nu}_e}$’s is about 5000 per day at the near detector and 100/day at the far detector. With a systematic error near 1%, the project could reach ${\sin^2(2\theta_{13})}< 0.03$ in 3 years.
Angra in Brazil
---------------
The Angra dos Reis reactor complex in Brazil, about 150 km south of Rio de Janeiro contains two reactors, Angra-I and Angra-II, which have 2 and 4 GW thermal powers and up times 83% and 94% respectively. The nearby site has high terrain consisting of granite, so both near and far detectors could have a substantial overburden. Initial designs for a ${\theta_{13}}$ experiment involve a near detector, 300 m from Angra-II, with 250 m.w.e. overburden, and a far detector, under the peak of a mountain called “Morro de Frade", which would provide 2000 m.w.e. at a distance of 1.5 km. Thoughts are to build a 50 ton near detector and a 500 ton far detector, and concentrate on reducing any bin-to-bin systematic errors. The 1000 ton KamLAND detector is a proof that large reactor neutrino detectors are possible. Unlike KamLAND but like the other new ${\theta_{13}}$ projects, the Angra collaboration plans to build a three-volume detector. For such a large detector, phototube costs scale as $V^{2/3}$. A statistical precision of ${\sin^2(2\theta_{13})}< 0.006$ could be obtained in three years.
The Angra experiment was originally conceived as a large ${\theta_{13}}$ detector under a considerable overburden together with a single reactor in order to obtain an large luminosity but still have substantial reactor off running. A funding request to the Brazilian Minister of Science and Technology in 2006 was approved for initial stages of the project. The experiment will be a long term project which will take advantage of lessons learned at Double Chooz, Daya Bay and RENO. In the meantime, smaller detectors are being constructed with possible applications toward the monitoring of reactor operations. The collaboration is establishing a formal agreement with Eletronuclear for permanent access to the site. They are already authorized to place one ton of Gd-loaded scintillator provided by LVD near to the reactor for muon background measurements. Other tests have measured noise and singles rate in the vicinity of the proposed detectors.
The next stage is a very near detector with three volumes of scintillator and a muon to be placed between 50 and 100 m from the core. The current design is a cylinder, 1.3 m high and with a 0.5 m radius for the target, 1.9 m high and 0.8 m radius for the $\gamma$-catcher, and 3.1 m high with a 1.4 m radius for the buffer.
A comparison of current reactor $\nu$ projects
----------------------------------------------
Two summaries of some features of the four current projects are given in Tables \[tab:1\] and \[tab:2\]. These tables were prepared with input from each collaboration’s management in October 2005[@bib:pc] and may not be up to date. In any case, the exact size and location of the detectors is subject to further modifications in design, and the “optimistic start dates" need to be taken with a huge clump of salt.
Project Power (P) $<P>$ Location Detectors
---------- ------------------- ------------------- ---------- -------------------
$GW_{th}$ $GW_{th}$ km/ton/m.w.e.
0.05/1/20
Angra 6.0 5.3 Brazil 0.3/50/250
1.5/500/2000
RENO 17.3 16.4 Korea 0.15/20/230
1.5/20/675
Daya Bay 11.6 9.9 China 0.36/40/260
(17.4 after 2010) (14.8 after 2010) 0.50/40/260
1.75/\[40x2\]/910
Double 8.7 7.4 France 0.27/10.2/90
Chooz 1.067/10.2/300
: Comparison of Detectors for four reactor $\nu$ projects.
\[tab:1\]
---------- ------------ ---------- -------------------------- ------------------------- ------------------ ----------------
Project Start Date GW-t-yr 90% CL for ${\Delta m^2_{32}}$ efficiences Far event
optimistic (yr) $\sin^2(2{\theta_{13}})$ ($10^{-3}eV^2$) rate
3900(1) 0.0070
Angra 2013(full) 9000(3) 0.0060 2.5 0.8$\times$0.9 350,000/yr
15000(5) 0.0055 2.5
RENO 2009 340(1) 0.03 2.0 0.8 18,000/yr
Daya Bay 2009 3700(3) 0.008 2.5 0.75$\times$0.83 70,000/yr
110,000/yr$^*$
Double 2007(far) 29(1) 0.08
Chooz 2008(near) 29(1+1) 0.04 2.5 0.8$\times$0.9 15,000/yr
80(1+3) 0.025
---------- ------------ ---------- -------------------------- ------------------------- ------------------ ----------------
: Comparison of Physics for reactor experiments. (\* For Daya Bay after 2010)
\[tab:2\]
A comparison of the philosophy of the new reactor projects can be discerned by a critical examination of Figure \[fig:lindner\]. The thick curve shows the evolution of ${\theta_{13}}$ sensitivity with reactor luminosity for a particular set of assumptions about the detector locations, ${\Delta m^2_{32}}$ and systematic error. That curve showed a transition from near the sensitivity of the rate-only test to near the sensitivity of the shape-only test between 200 and 2000 GW-ton-year. The Double Chooz and RENO projects aim to quickly improve the limit by reaching the “transition" near 200 GW-ton-year. Daya Bay has adopted a goal to work hard to reduce systematic errors below the assumptions of Figure \[fig:lindner\]. It will reach ${\sin^2(2\theta_{13})}\sim 0.01$ with 2000 -ton-year, perhaps by using movable detectors. Angra’s strategy is to build a much larger far detector with $ > 10,000$ -ton-year to make it less sensitive to systematic error. Depending, of course, on ${\Delta m^2_{32}}$, it is reasonable for the field to have a sensitivity goal of ${\sin^2(2\theta_{13})}\sim 0.01$, as might be achievable with the Daya Bay or Angra experiments. However, as can be seen for Figure \[fig:lindner\], the luminosity requirement for 0.01 is 70 times larger than for 0.03, following the thick curve. In that sense, a 0.01 experiment is 70 times harder than an 0.03 experiment, and the earlier and less expensive Double Chooz and RENO projects can be valuable steps on the learning curve for a successful 0.01 experiment.
It would be desirable to compare the real schedules of these four projects. All four projects have some funding, though not necessarily enough to reach their design goals (yet). Even though it is the cheapest experiment, Double Chooz’ schedule is limited only by funding. The other four projects, which will require considerably more civil construction, also have schedules that are probably limited both by funding and by technical considerations.
The near future
===============
The earliest results from reactor experiments may be three years or more away. However, at the next Neutrino Telescopes meeting in 2009, observers will be able to gauge progress by paying attention to the following subjects:
- Updated estimates of GW-ton-year as the final detector design and efficiencies are completed,
- Liquid scintillator production and stability and attenuation length studies using large amounts of liquid scintillator,
- Civil construction issues and, in particular, experience with costs and schedules,
- Improved estimates of the background for cosmogenic sources such as $^9Li$, (It may be possible to achieve an improved understanding of the possible production mechanisms for cosmogenic sources. In any case, each experiment should carefully estimate the range of uncertainty of their background estimates, the impact that uncertainty would have on the ${\theta_{13}}$ sensitivity, and quantitative methods for measurements that will lead to a reduction of the uncertainty when data taking is underway.)
- Calibration system development and the results of relative calibration measurements between two or more detectors.
- Progress in the implementation of movable detectors as a calibration technique, and evidence as to whether this is a reliable method, given the progress or absence of changes when the detectors are moved.
The longer term
===============
Due to the importance of ${\theta_{13}}$ for CP violation and the mass hierarchy, a potential long-term program of reactor neutrino measurements lies ahead of us. Results from Double Chooz, Daya Bay, RENO, and later Angra, will be used to determine the value of upgrades, additional detectors, and new projects. An important factor will be whether the goal becomes further limits on a small value of ${\theta_{13}}$, or more precise measurements of a non-zero value. Statistical precision better than $\delta({\sin^2(2\theta_{13})}) < 0.01$ can be imagined, but experience with systematic errors and backgrounds must be weighed along with the capabilities and needs of accelerator experiments. Ideas already exist for more ambitious reactor experiments to study $\theta_{13}$ further, as well as $\theta_{12}$. Some examples are Triple Chooz[@bib:triple], R2D2[@bib:r2d2] and Hano Hano[@bib:learned]. If such projects become reality, they will certainly be based on lessons not yet learned by Double Chooz, Daya Bay, RENO and Angra.
Acknowledgments
===============
Thanks to the organizers of Neutrino Telescopes 2007 for the opportunity to discuss reactor neutrinos. I am indebted to many colleagues for information in the preparation of this paper, including Jo[ã]{}o dos Anjos, Milind Diwan, Karsten Heeger, Ernesto Kemp, Soo-Bong Kim, Thierry Laserre, Manfred Lindner, Kam-Biu Luk, Guillaume Mention, David Reyna, Michael Shaevitz and Patricia Vahle.
[99]{} M. Apollonio et al., Eur. Phys. J. C27 (2003) 331-374. F. Boehm et al., Phys. Rev. Lett. 84, 3764(2000). Tokai to Kamiokande (T2K) Letter of intent, “Neutrino Oscillation Experiment at JHF", Y. Hayato et al., January 21, 2003; Available at http://neutrino.kek.jp/jhfnu/loi/loi$\_$JHFcor.pdf. Neutrino Oscillation $\nu_e$ Appearance (NO$\nu$A) Proposal, “Proposal to Build a 30 Kiloton Off-Axis Detector to Study ${\nu_\mu \rightarrow \nu_e}$ Oscillations in the NuMI Beamline, D.S. Ayres et al., March 21, 2005; Available at http://nova-docdb.fnal.gov/0005/000593/001/NOvA$\_$P929$\_$March21$\_$2005.pdf.
P. Adamson et al., Phys. Rev. D73 (2006) 072002. Patricia Vahle for the MINOS collaboration, “The Status of MINOS in the Second Year of Beam Running", these proceedings. W.-M.-Yao et al., J. Phys. G: Nucl. Part. Phys. 33 (2006), 1; also known as the Review of Particle Physics from the Particle Data Group. P. Huber et al., Phys.Rev. D70 (2004) 073014. K. Anderson et al., “A New Nuclear Reactor Neutrino Experiment to Measure ${\theta_{13}}$", Also known as the reactor neutrino White Paper, Available at hep-ex/0402041. Double Chooz Proposal, F. Ardellier et al., “Double Chooz: A Search for the Neutrino Mixing Angle ${\theta_{13}}$", available at hep-ex/0405032. Daya Bay Proposal, “A Precision Measurement of the Neutrino Mixing Angle ${\theta_{13}}$ Using Reactor Antineutrinos at Daya Bay.", December 1, 2006, available at hep-ex/0701029. The Braidwood Collaboration, “Braidwood Project Description", available at http://mwtheta13.uchicago.edu/project$\_$web.ps V. Martemyanov et al., “The KR2DET project: Search for mass-3 state contribution to $|U_{e3}|^2$ to the electron neutrino using a one reactor - two detector oscillation experiment at Krasnoyarsk underground site", Phys.Atom.Nucl. 66 (2003) 1934-1939; Yad.Fiz. 66 (2003) 1982-1987; Available at arXiv:hep-ex/0211070. Neutrino Scientific Assessment Group, “Recommendations to the Department of Energy and the National Science Foundation on a U.S. Program of Reactor- and Accelerator-based Neutrino Oscillation Experiments", February 28, 2006; available at http://www.science.doe.gov/hep/NuSAG2ndRptFeb2006.pdf. KASKA Letter of Intent, M. Aoki et al., “Letter of Intent for KASKA: High Accuracy Neutrino Oscillation Measurements with ${\bar{\nu}_e}$’s from Kashiwazaki-Kariwa Nuclear Power Station", 11 July 2006, available at hep-ex/0607013. See http://neutrino.snu.ac.kr/RENO/ Private communication with spokespersons of Angra, Daya Bay, Double Chooz and RENO during September 2005 for the 8th ICFA Seminar on “Future Perspectives in High Energy Physics", Kyungpook National University, Daegu South Korea, October 2005. http://chep.knu.ac.kr/ICFA-Seminar/. J.C. Anjos et al., Brazilian Journal of Physics, 36 4A(2006) G. Mention et al., “A unified analysis of the reactor neutrino program toward the measurement of the ${\theta_{13}}$ mixing angle.", hep-ex/0704.0498/. P. Huber et al., JHEP 05 (2006) 072. P. Huber et al., JHEP 0502 (2005) 029. J. Learned, “Hano Hano", these proceedings.
[^1]: submitted to the proceedings of the XIIth International Workshop on Neutrino Telescopes, Venice 2007
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'This paper presents a method to square up a generic MIMO system that already possesses transmission zeros. The proposed method is developed based on and therefore can be incorporated into the existing method that has been proven effective on a system without transmission zeros. It has been shown that for the generic system considering here, the squaring-up problem can be transformed into a state-feedback pole-placement problem with uncontrollable modes.'
author:
- 'Zheng Qu, Daniel Wiese, Anuradha M. Annaswamy and Eugene Lavretsky [^1] [^2]'
bibliography:
- 'Ref/VFA\_Refworks\_v4.bib'
title:
- ' **On Output Feedback Adaptive Control**'
- 'Squaring-Up Method In the Presence of Transmission Zeros'
---
Introduction
============
Square system plays key role in control theory development because of some unique properties it may possess, such as invertibility[@RefWorks:96] and strictly positive realness[@RefWorks:98]. Some control technique is first developed based on square systems and later on, extended into a more general plant using square systems as a leverage. Such extension usually requires a squaring method. By squaring, we particularly define it as a way to coincide the number of inputs and outputs and also, make the squared system minimum phase. Minimum phase system is of particular interest because its inversion is also stable, which is the necessary conditions of some advanced control design[@RefWorks:55].
Two different squaring methods have been presented in the previous literature, i.e. squaring-down and squaring-up. The squaring-down method is first attempted in 1970s[@RefWorks:74; @RefWorks:68] and is revealed to be equivalent to pole-placement using output feedback in the transformed coordinate. Pole-placement using output feedback has been shown only achievable under some specific conditions and therefore can be restrictive. On the other hand, the research on squaring-up method has been sparse until Misra’s work in 1990s[@RefWorks:70; @RefWorks:102]. It has been shown the squaring-up method is equivalent to pole-placement using state feedback in the transformed coordinate and therefore is much more feasible. Actually, pole-placement using state feedback has been considered a solved problem as long as the controllability condition in the transformed coordinate is met. In terms of control design, however, squaring-up method does not prevail over squaring-down method because it brings pseudo-inputs or outputs into the system that cannot be used in the physical world.
Recently, the squaring-up method has gained increasing interest in the new development of adaptive control theory when a minimum phase system is assumed[@RefWorks:55], or is required as an intermediate step[@RefWorks:90; @RefWorks:36]. Some properties the square system is having, such as strictly positive realness, can be transmitted to the original system using proper partition. Thus the pseudo-inputs and outputs are only used in the gain design but never used in the actual physical action. The final results of the squaring-up treatment produces a squaring-down matrix, which can be realized in the real world. One problem arises, however, that when the given system already possesses transmission zeros, the existing squaring-up method fails. Previous literature warns the reader but did not specify a solution[@RefWorks:70]. Although the existence of transmission zeros in a non-square system is especially rare[@RefWorks:92], the failure of squaring-up method in such case draws interest because the existing transmission zero, as long as it is nonminimum-phase, has been demonstrated to be nonpreventive in the adaptive control design. Such observation motivates us to rationalize the failure and its countermeasure. This paper will show the pre-existing nonminimum phase zeros are the only case of our interest that the existing squaring-up method cannot work. Section \[sub:Countermeasure\] will provide a remedy to the method and Section \[sec:Example\] will present a numeric example.
Preliminaries
=============
Given a system $\{A,B,C,D\}$, $A\in\mathbb{R}^{n\times n}$, $B\in\mathbb{R}^{n\times m}$, $C\in\mathbb{R}^{p\times n}$, $D\in\mathbb{R}^{p\times m}$. The system $\{A,B,C,D\}$ is square if $m=p$. It is tall if $m<p$. It is fat if $m>p$. For a non-square system, the procedure to square-up is defined as adding more inputs or outputs to make $m=p$. The procedure to square-down is defined as abandoning inputs or outputs to make $m=p$. Given the slight disagreement in the definition of the zeros of a MIMO system[@RefWorks:104; @RefWorks:106], this paper begins with a series of zero definitions that are widely accepted in recent literature and relevant to our investigation.
The normal rank of a matrix function $X(s)$ is defined as the rank of $X(s)$ for almost all the values of $s$.
It is the interest of this paper to study the few $s$ such that $X(s)$ loses its normal rank.
[@RefWorks:91]The Rosenbrock matrix of a system $\{A,B,C,D\}$ is defined as: $$R(s)=\left[\begin{array}{cc}
sI-A & -B\\
C & D
\end{array}\right]\label{eq:Rosenbrock}$$
The Rosenbrock matrix is first proposed in the reference[@RefWorks:91] and has been widely used to study the zeros of MIMO systems.
If for a system $\{A,B,C,D\}$, the rank of $R(s)$ is strictly less than the $min(n+m,\ n+p)$ for any complex values of $s$, then the system is degenerate.
It is noted some controllable and observable system can be degenerate. The system is degenerate when there are some repeated states, inputs or outputs. For example, if $m<p$ and the system has repeated inputs, the $B$ has identical columns and naturally, $R(s)$ has a rank less than $n+m$ for any $s$.
For a system $\{A,B,C,D\}$, the input decoupling zeros are the values of $s$ such that the following $n\times(n+m)$ matrix loses its normal rank: $$R_{I}(s)=\left[\begin{array}{cc}
sI-A & -B\end{array}\right]$$
The input decoupling zeros are a subset of system poles. They are actually the uncontrollable modes of the system.
\[Def:Output Decoup\]For a system $\{A,B,C,D\}$, the output decoupling zeros are the values of $s$ such that the following $(n+p)\times n$ matrix loses its normal rank: $$R_{O}(s)=\left[\begin{array}{c}
sI-A\\
C
\end{array}\right]$$
The output decoupling zeros are a subset of system poles. They are actually the unobservable modes of the system.
For a system $\{A,B,C,D\}$, the invariant zeros are the values of $s$ such that $R(s)$ loses its normal rank.
This definition is first proposed by Rosenbrock[@RefWorks:91]. The name “invariant” comes from the fact that such zeros are invariant under feedback action, either state feedback or observer feedback.
\[Def:Tzero1\]For a system $G(s)=C(sI-A)^{-1}B+D$, the transmission zeros are defined as the values of $s$ such that the rank of $G(s)$ is less than its normal rank.
The definition is first proposed by Desoer and Schulman and later on generalized by MacFarlane[@RefWorks:107]. It is intended to describe a special property of the system that it blocks the transmission from input to output at some specific frequencies. Such frequencies are named as the “transmission zeros” of the system. Recently, a state-space form of transmission zero definition has been proposed.
\[Def:Tzero2\][@RefWorks:92]For a non-degenerate system $\{A,B,C,D\}$ that is controllable and observable, the transmission zeros are the values of $s$ such that $rank[R(s)]<min(n+m,n+p)$.
It is easy to show that for the system considered here, the set of {Invariant zeros} = the set of {transmission zeros + input decoupling zeros + output decoupling zeros - input and output decoupling zeros}.
It has been proved in the reference[@RefWorks:92] that Definition \[Def:Tzero1\] and Definition \[Def:Tzero2\] are equivalent. The idea is that the $G(s)$ of a non-degenerate system has a normal rank of $min(n+m,n+p)$, and therefore the invariant zeros are the $s$ such that $rank[R(s)]<min(n+m,n+p)$. The condition of controllability and observability ensures the invariant zeros excludes the input-decoupling zeros and output-decoupling zeros that are not shown in the $G(s)$. Then the invariant zeros are indeed the transmission zeros. Then $G(s)$ loses rank at the exactly $s$ at which $R(s)$ loses rank.
The system is minimum phase if all its transmission zeros are in strictly left-hand plane.
Squaring-Up Method
==================
Problem Definition
------------------
For the particular interest of adaptive control[@RefWorks:90][@RefWorks:36], we are dealing with a system $\Sigma_{p}=\{A,B,C\}$ that satisfies following assumptions.
\[enu:01\] The system is fat and satisfies $n>>m>p$;
The given system $\{A,B,C\}$ is strictly proper, i.e. $D=0$;
\[enu:ABC\] $(A,B)$ is a controllable pair, and $(A,C)$ is an observable pair;
\[enu:B\] $B$ has full column rank, i.e. $rank(B)=m$;
\[enu:CB\] $rank(CB)=p$
The observability of $(A,C)$ is not necessarily required for the following procedure to work[@RefWorks:70]. It only serves for the purpose of analysis simplification. Otherwise, the system will have output-decoupling zeros that makes $R(s)$ rank deficient according to Definition \[Def:Output Decoup\]. We excludes this case since unobservable modes are not our interest in the context of adaptive control design. With the observability condition, the only $s$ that makes $R(s)$ rank deficient is its transmission zeros.
The goal is to find an augmentation $C_{a}\in\mathbb{R}^{(m-p)\times n}$ such that the system $\{A,B,\bar{C}\}$ is square and minimum phase, where $\bar{C}^{T}=[C^{T},C_{a}^{T}]$. Assumption \[enu:ABC\], \[enu:B\] and \[enu:CB\] guarantee $\Sigma_{p}$ is non-degenerate. Assumption \[enu:ABC\] guarantee the $\{A,B,C\}$ is the minimal realization of the system. By Definition \[Def:Tzero2\], the number of transmission zeros should be zero or a finite number. Based on Definition \[Def:Tzero2\], Misra proposes a method to find $C_{a}$ using the technique of pole placement in a special coordinate[@RefWorks:70]. The following section briefly summarizes the steps. For expediency, we will not distinguish between the term “rank” and the term “normal rank”.
The Existing Method
-------------------
The Rosenbrock matrix $R(s)$ can be transformed into a special coordinate where the controllable states and the uncontrollable states are separated. $$\tilde{R}(s)=TR(s)T^{T}=\left[\begin{array}{cc|c}
sI_{m}-A_{11} & -A_{12} & -B_{1}\\
-A_{21} & sI_{n-m}-A_{22} & 0\\
\hline C_{11} & C_{12} & 0
\end{array}\right]\label{eq:Tran Rs}$$ Since $T$ is an invertible matrix, $rank[\tilde{R}(s)]=rank[R(s)]$ for all $s\in\mathbb{C}$. Assuming a $C_{a}=[C_{21},C_{22}]$ is found and the augmented system is following: $$\tilde{R}_{a}(s)=\left[\begin{array}{cc|c}
sI_{m}-A_{11} & -A_{12} & -B_{1}\\
-A_{21} & sI_{n-m}-A_{22} & 0\\
\hline C_{11} & C_{12} & 0\\
C_{21} & C_{22} & 0
\end{array}\right]\label{eq:Tran Ras}$$ Group the column of $\bar{C}$ and denote $C_{1}=\left[\begin{array}{c}
C_{11}\\
C_{21}
\end{array}\right]$ and $C_{2}=\left[\begin{array}{c}
C_{12}\\
C_{22}
\end{array}\right]$. Choose $C_{21}$ such that $C_{1}$ is an invertible matrix. Without loss of generality, we choose: $$C_{21}=null(C_{11}^{T}))\label{eq:C21}$$ where $null$ stands for the null space. $C_{21}$ can be made unity. The following equality holds: [ $$\begin{aligned}
& & rank[\tilde{R}_{a}(s)]=rank\left(\tilde{R}(s)\left[\begin{array}{cc|c}
I_{m} & -C_{1}^{-1}C_{2} & 0\\
0 & I_{n-m} & 0\\
\hline 0 & 0 & I_{m}
\end{array}\right]\right)\nonumber \\
& & =rank\left[\begin{array}{cc|c}
sI_{m}-A_{11} & -sC_{1}^{-1}C_{2}-A_{12}+A_{11}C_{1}^{-1}C_{2} & -B_{1}\\
-A_{21} & sI_{n-m}-A_{22}+A_{21}C_{1}^{-1}C_{2} & 0\\
\hline C_{1} & 0 & 0
\end{array}\right]\label{eq:rank tran}\end{aligned}$$ ]{}It is easy to show that: $$\begin{aligned}
& rank[\tilde{R}_{a}(s)] & =rank(C_{1})+rank(B_{1})\nonumber \\
& & +rank(sI_{n-m}-A_{22}+A_{21}C_{1}^{-1}C_{2})\label{eq:rank cond}\end{aligned}$$ Since $rank(C_{1})=rank(B_{1})=m$ by design and assumption \[enu:B\] and \[enu:CB\], $\tilde{R}(s)$ loses rank only if $(sI_{n-m}-A_{22}+A_{21}C_{1}^{-1}C_{2})$ loses rank. From Definition \[Def:Tzero2\], the transmission zeros of the system is exactly the poles of $(A_{22}-A_{21}C_{1}^{-1}C_{2})$. From Eq.(\[eq:Tran Rs\]), it is easy to see that the assumption $(A,B)$ is controllable implies $(A_{22},A_{21})$ is controllable, which in turn implies $(A_{22},A_{21}C_{1}^{-1})$ is controllable (since $C_{1}$ is invertible). This implies state feedback technique can be used on the pair $(A_{22},A_{21}C_{1}^{-1})$ to place the zeros of the system. The remaining problem is to deal with the fact that $C_{2}$ is not totally free (since $C_{12}$ is given). Perform partition on $C_{2}$: $$\tilde{C}_{2}=\left[\begin{array}{c}
C_{12}\\
O_{(m-p)\times(n-m)}
\end{array}\right]\; and\;\hat{C}_{2}=\left[\begin{array}{c}
O_{p\times(n-m)}\\
C_{22}
\end{array}\right]$$ Correspondingly: $$A_{22}-A_{21}C_{1}^{-1}C_{2}=\underset{\tilde{A}_{22}}{\underbrace{A_{22}-A_{21}C_{1}^{-1}\tilde{C}_{2}}}-A_{21}C_{1}^{-1}\hat{C}_{2}$$ We only have freedom in designing $\hat{C}_{2}$. That means only last $(m-p)$ pseudo inputs of $A_{21}C_{1}^{-1}$ are available for pole placement. Denote $B_{ps}=A_{21}C_{1}^{-1}$ and perform corresponding partition: $$B_{ps}\triangleq A_{21}C_{1}^{-1}\triangleq[B_{ps1},B_{ps2}]$$ Now the problem becomes pole placement using feedback on the pair $(\tilde{A}_{22},B_{ps2})$ where $\tilde{A}_{22}\triangleq A_{22}-A_{21}C_{1}^{-1}\tilde{C}_{2}$. However, from all above derivation, there is no guarantee that $(\tilde{A}_{22},B_{ps2})$ is controllable. It is found that for some special systems satisfying all assumptions list above, $(\tilde{A}_{22},B_{ps2})$ can be uncontrollable. For such system, the existing square-up procedure won’t work. Following context will elaborate the properties of such system and propose the countermeasure.
Presence of Transmission Zeros
------------------------------
Suppose $(\tilde{A}_{22},B_{ps2})$ is not controllable. Then there exists a scalar $s_{0}$ and a vector $w_{0}$ such that: $$w_{0}^{T}[s_{0}I-\tilde{A}_{22},B_{ps2}]=0$$ It follows: $$\begin{aligned}
& w_{0}^{T}s_{0}-w_{0}^{T}\tilde{A}_{22}=0\label{eq:w01}\\
& w_{0}^{T}B_{ps2}=0\label{eq:w02}\end{aligned}$$ Substituting the definition of $\tilde{A}_{22}$ transforms Eq.(\[eq:w01\]) into: $$w_{0}^{T}s_{0}-w_{0}^{T}A_{22}+w_{0}^{T}A_{21}C_{1}^{-1}\tilde{C}_{2}=0\label{eq:w01t}$$ Also it is noted: $$w_{0}^{T}A_{21}C_{1}^{-1}=w_{0}^{T}B_{ps}=w_{0}^{T}[B_{ps1},B_{ps2}]=w_{0}^{T}[B_{ps1},0]\label{eq:Bps1}$$ The last equality is true because of Eq.(\[eq:w02\]). Now let’s examine the form of $C_{1}^{-1}$. With loss of generality, $C_{1}^{-1}$ can be written as: $$C_{1}^{-1}=[C_{11}^{\dagger},C_{21}^{T}]\label{eq:C1inv}$$ where $C_{11}^{\dagger}$ stands for the right inverse of $C_{11}$. One can easily verify $C_{1}C_{1}^{-1}=I_{m}$ using the facts: $$\begin{cases}
& C_{11}C_{21}^{T}=O_{p\times(m-p)}\\
& C_{11}C_{11}^{\dagger}=I_{p}\\
& C_{21}C_{11}^{\dagger}=O_{(m-p)\times p}\\
& C_{21}C_{21}^{T}=I_{m-p}
\end{cases}$$ Using Eq.(\[eq:C1inv\]), $B_{ps1}$ and $B_{ps2}$ can be rewritten as: $$B_{ps1}=A_{21}C_{11}^{\dagger}\qquad and\qquad B_{ps2}=A_{21}C_{21}^{T}\label{eq:Bps}$$ Eq.(\[eq:Bps1\]) and Eq.(\[eq:Bps\]) can transform Eq.(\[eq:w01t\]) into: $$w_{0}^{T}s_{0}-w_{0}^{T}A_{22}+w_{0}^{T}A_{21}C_{11}^{\dagger}C_{12}=0\label{eq:rankdef}$$ Now we will examine what Eq.(\[eq:rankdef\]) implies of the original system (\[eq:Tran Rs\]). Following equality takes place: [ $$\begin{aligned}
& & rank[\tilde{R}(s)]=rank\left(\tilde{R}(s)\left[\begin{array}{cc|c}
I_{m} & -C_{11}^{\dagger}C_{12} & 0\\
0 & I_{n-m} & 0\\
\hline 0 & 0 & I_{m}
\end{array}\right]\right)\nonumber \\
& & =rank\left[\begin{array}{cc|c}
sI_{m}-A_{11} & -sC_{11}^{\dagger}C_{12}-A_{12}+A_{11}C_{11}^{\dagger}C_{12} & -B_{1}\\
-A_{21} & sI_{n-m}-A_{22}+A_{21}C_{11}^{\dagger}C_{12} & 0\\
\hline C_{11} & 0 & 0
\end{array}\right]\end{aligned}$$ ]{}Similar to Eq.(\[eq:rank cond\]), the rank of $\tilde{R}(s)$ fully depends on $C_{11}$, $B_{1}$ and $sI_{n-m}-A_{22}+A_{21}C_{11}^{\dagger}C_{12}$: $$\begin{aligned}
& rank[\tilde{R}(s)] & =rank(C_{11})+rank(B_{1})\nonumber \\
& & +rank(sI_{n-m}-A_{22}+A_{21}C_{11}^{\dagger}C_{12})\label{eq:Ori rank cond}\end{aligned}$$ Eq.(\[eq:rankdef\]) says there exists a $s_{0}$ such that $sI_{n-m}-A_{22}+A_{21}C_{11}^{\dagger}C_{12}$ loses rank. Eq.(\[eq:Ori rank cond\]) says such $s_{0}$ will make the original system $\tilde{R}(s)$ loses rank. By definition \[Def:Tzero2\], the system $R(s)$ has a transmission zero at $s_{0}$. It can be concluded now that given assumption \[enu:01\] to \[enu:CB\], the only case Misra’s method can not solve is the case when the given system already posses a transmission zero. Comparing Eq.(\[eq:Tran Rs\]) and Eq.(\[eq:Tran Ras\]), it is easy to see that any $s_{0}$ that makes $\tilde{R}(s)$ lose rank will also make $\tilde{R_{a}}(s)$ loses rank. In other words, any transmission zeros of the given system will become the transmission zeros of the squared-up system. That is one important limitation of the square-up procedure.
Countermeasure\[sub:Countermeasure\]
------------------------------------
The countermeasure is following. The above derivation can be reversed and the sufficient condition argument is true, i.e. the transmission zeros of the given system is indeed the uncontrollable mode of the pair $(\tilde{A}_{22},B_{ps2})$. Even if uncontrollable modes exist, other controllable modes can be placed in the strictly left-hand plane using the remaining feedback action. As a result, it can be concluded that if the given system satisfies one additional condition:
\[enu:Tran\]The system has only nonminimum-phase transmission zeros.
the pair $(\tilde{A}_{22},B_{ps2})$ is stabilizable and we are still be able to design a $C_{a}$ such that the squared-up system is minimum phase. LQR technique is immediately available for such problem. We summarized our improved method as following:
Check if the given system satisfies all assumption \[enu:01\] to \[enu:Tran\];
Transform it into a controllable canonical form as in Eq.(\[eq:Tran Rs\]);
Find $C_{21}$ using Eq.(\[eq:C21\]);
Calculate the stabilizable pair $(\tilde{A}_{22},B_{ps2})$;
Perform LQR technique on $(\tilde{A}_{22},B_{ps2})$ to find $C_{22}$;
Augment $C$ and transform the system back to its original coordinate.
Example\[sec:Example\]
======================
Following context gives an example of a MIMO system with a transmission zero and the results of our squaring-up procedure. It is a linearized model for the lateral dynamics of Boeing 747-100 aircraft. We transposed the system for the illustration of a fat system. [ $$R(s)=\left[\begin{array}{cccc|ccc}
s+0.0605 & 0.0015 & -0.0011 & 0 & 0 & 0 & 0\\
0 & s+0.4603 & 0.0208 & 1 & -1 & 0 & 0\\
871 & -0.28 & s+0.141 & 0 & 0 & -1 & 0\\
32.3 & 0 & 0 & 0 & 0 & 0 & -1\\
\hline 0 & -0.1860 & 0.0061 & 0 & 0 & 0 & 0\\
4.0380 & 0.1 & -0.4419 & 0 & 0 & 0 & 0
\end{array}\right]$$ ]{}A quick check can confirm the given system satisfies assumption \[enu:01\] to \[enu:CB\], and has an transmission zero at $-0.0511$. Coordinate transformation $T=[B^{T},(null(B^{T}))^{T}]$gives: [ $$\tilde{R}(s)=\left[\begin{array}{cccc|ccc}
s+0.4603 & 0.0208 & -1 & 0 & -1 & 0 & 0\\
-0.28 & s+0.1410 & 0 & -871 & 0 & -1 & 0\\
0 & 0 & 0 & -32.3 & 0 & 0 & -1\\
-0.0015 & 0.0011 & 0 & 0.0605 & 0 & 0 & 0\\
\hline -0.1860 & 0.0061 & 0 & 0 & 0 & 0 & 0\\
0.1 & -0.4419 & 0 & -4.038 & 0 & 0 & 0
\end{array}\right]$$ ]{}Quick examination sees: $$A_{22}=-0.0605\quad and\quad A_{21}=[0.0015,-0.0011,0]$$ By our choice of $C_{21}$, $C_{1}$ and $C_{1}^{-1}$ is given: $$\begin{aligned}
& C_{1}=\left[\begin{array}{ccc}
-0.186 & 0.0061 & 0\\
0.1 & -0.4419 & 0\\
0 & 0 & 1
\end{array}\right]\\
& C_{1}^{-1}=\left[\begin{array}{ccc}
-5.4171 & -0.0744 & 0\\
-1.2263 & -2.2796 & 0\\
0 & 0 & 1
\end{array}\right]\end{aligned}$$ And $C_{21}$ and $\tilde{C}_{2}$ are: $$C_{21}=[0,0,1]\quad and\quad\tilde{C}_{2}=\left[\begin{array}{c}
0\\
-4.038\\
0
\end{array}\right]$$ Then $\tilde{A}_{22}=A_{22}-A_{21}C_{1}^{-1}\tilde{C}_{2}$ and $B_{ps2}=A_{21}C_{21}^{T}$ gives: $$\tilde{A}_{22}=-0.0511\quad and\quad B_{ps2}=0$$ It is verified that the pair $(\tilde{A}_{22},B_{ps2})$ is uncontrollable and the uncontrollable mode is exactly the transmission zero of the given system. Simply put $\hat{C}_{2}=O_{m\times(n-m)}$ and transform the $C_{a}$ back into the original coordinate. The augmented system will become: [ $$R_{a}(s)=\left[\begin{array}{cccc|ccc}
s+0.0605 & 0.0015 & -0.0011 & 0 & 0 & 0 & 0\\
0 & s+0.4603 & 0.0208 & 1 & -1 & 0 & 0\\
871 & -0.28 & s+0.141 & 0 & 0 & -1 & 0\\
32.3 & 0 & 0 & 0 & 0 & 0 & -1\\
\hline 0 & -0.1860 & 0.0061 & 0 & 0 & 0 & 0\\
4.0380 & 0.1 & -0.4419 & 0 & 0 & 0 & 0\\
0 & 0 & 0 & 1 & 0 & 0 & 0
\end{array}\right]$$ ]{}The last row is the designed pseudo-output. Quick examination will verify the augmented system $R_{a}(s)$ has only one transmission zero at $-0.0511$.
Conclusions
===========
This paper proves that the uncontrollable modes in the existing squaring-up method are actually the transmission zeros of the given system. In other words, we are not able to move the locations of the existing transmission zeros using the proposed method. Systems with minimum phase zeros has a stabilizable pair in the transformed coordinate and therefore can be squared-up using LQR technique. It is noted that by transposing the system, the proposed method can be applied on a tall system with more outputs than inputs.
[^1]: Z. Qu, D. Wiese and A. M. Annaswamy are with the Department of Mechanical Engineering, Massaschusetts Institute of Technology, Cambridge, MA, 02139 e-mail: ([[email protected]]{}).
[^2]: E. Lavretsky is with the Boeing Company, Huntington Beach CA, 92648.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'We derive the basic physical parameters of the field double-mode RR Lyrae star BSCom from its observed periods and the requirement of consistency between the pulsational and evolutionary constraints. By using the current solar-scaled horizontal branch evolutionary models of @pietrinferni04apj and our linear non-adiabatic purely radiative pulsational models, we get $M / M_{\odot} = 0.698 \pm 0.004$, $\log(L/L_{\odot}) = 1.712 \pm 0.005$, $T_{\rm eff} = 6840 \pm 14$K, \[Fe/H\]=$-1.67 \pm 0.01$, where the errors are standard deviations assuming uniform age distribution along the full range of uncertainty in age. The last two parameters are in a good agreement with the ones derived from the observed $BVI_{\mathrm{C}}$ colours and the updated ATLAS9 stellar atmosphere models. We get $T_{\rm eff}=6842 \pm 10$K, \[Fe/H\]=$-1.58 \pm 0.11$, where the errors are purely statistical ones. It is remarkable that the derived parameters are nearly independent of stellar age at early evolutionary stages. Later stages, corresponding to the evolution toward the asymptotic giant branch are most probably excluded because the required high temperatures are less likely to satisfy the constraints posed by the colours. We also show that our conclusions are only weakly sensitive to nonlinear period shifts predicted by current hydrodynamical models.'
date: 'Released: will be inserted later'
title: 'Pulsational and evolutionary analysis of the double-mode RR Lyrae star BSCom'
---
\[firstpage\]
stars: individual: BSCom – stars: variables: RR Lyrae – stars: abundances – stars: evolution – stars: fundamental parameters
Introduction {#sect:introduction}
============
Double-mode radial pulsators play a very significant role in the study of pulsating stars. This is because we can determine all fundamental stellar parameters just by using a few observed parameters and relying on the rather solid theory of linear radial pulsation (assuming of course, that the mode identification is correct – a question often raised for typically low-amplitude pulsators whereas rarely bothered with in the case of high/moderate-amplitude ones). In a series of papers [@kw99apj; @kovacs2000smc; @kovacs2000lmc] we investigated the consistency of the method by using the observed colours and assumed cluster metallicities in deriving distances for globular clusters and for the Magellanic Clouds. These distances have been confirmed by independent studies, for example by direct HST parallax measurements of @vL07mnras and @benedict07aj, by *Hipparcos*-based $\delta$ Scuti PL-relation of @mcnamara07aj, by mixtures of parallax and interferometric methods of @fouque07aap and by Baade-Wesselink analyses [@storm04aap; @kovacs03mnras]. In yet another application, @buchler07apj put limits on the metallicities of Magellanic Cloud beat Cepheids through stellar evolution constraints.
Double-mode radial pulsators are also viable for nonlinear hydrodynamical modelling. Although early attempts have already shown that sustained double-mode pulsation is possible with purely radiative hydrodynamical codes [at least for RR Lyrae stars, see @kovacs93apj], the modelling was put on a better sounding physical ground only after time-dependent convection was more carefully included (for Cepheids by @kollath98apjl, for RR Lyrae stars by @feuchtinger98aap and @szabo04aap). Although these models are capable of producing stable double-mode pulsation in the appropriate neighbourhood of the physical parameter space and the model light curves are in reasonable agreement with the observed ones, star-by-star modelling and application of more stringent observational constraints are still lacking.
Large-scale surveys aimed at microlensing and variability studies have already yielded a considerable increase in the number of known double-mode variables. In particular, before 2000, only four double-mode RR Lyrae (RRd) stars were identified in the Milky Way (see @garcia-melendo97aj and @kovacs01book for additional statistics). At the moment of this writing we have 26 RRd stars known in the Galactic field according to the literature, mostly due to the NSVS and ASAS full-sky variability surveys, and many more in the Galactic Bulge, due to the OGLE survey (@mizerski03aca; @moskalik03aap and @pigulski03aca).
The field RR Lyrae star BSCom ($\alpha = 13^{\rm h} 34^{\rm m} 39^{\rm s}$, $\delta = +24^{\circ} 16' 38''$) was originally put on the observing list of the 60cm automated telescope of the Konkoly Observatory (see @sodor07an for details on the ongoing photometric survey), due to the suspected Blazhko behaviour of this star reported by @clementini95mnras. While gathering data, the star was analysed by @wils06ibvs, using the NSVS database. He found that BSCom was an RRd star[^1].
This paper describes the results of the frequency analysis of our data and the pulsational/evolutionary analysis. The result of this latter, purely theoretical study will be confronted with additional constraints posed by the effective temperature and metallicity obtained from our multi-colour data. We show that the purely theoretical approach yields stellar parameters in agreement with the colour data and can be used for very accurate parameter estimation once we are able to pose an age limit for the object.
Photometry {#sect:photometry}
==========
Observations and data reduction {#subsect:obs}
-------------------------------
We obtained $BV(RI)_C$ photometric observations of BSCom with the 60 cm automated telescope of the Konkoly Observatory, Budapest. The telescope is equipped with a Wright $750\times1100$ CCD providing a $17'\times24'$ field of view (detailed parameters are given by @bakos99). The observations were carried out between 2005 March 1 and 2006 April 17 in 3 seasons, on a total number of 35 nights. Observational statistics are shown in Table \[tab:obstat\]. BSCom magnitudes were measured relative to BD+24 2598 by aperture photometry[^2]. GSC00454-00454 and USNO-B11143-0206728 served as check stars. We used standard IRAF[^3] packages for data reduction procedures. Coefficients for linear transformation of the instrumental data into standard magnitudes were derived each year.
BSCom was tied to the standard photometric system in $B,V$ and $I_{\rm C}$ colours through the comparison star BD+24 2598. The magnitudes of BD+24 2598 were measured relative to HD117876 ($V = 6$$082$, $B-V = 0$$968$, $V-R = 0$$522$, $V-I = 1$$017$, see @holtzman84aj), the closest standard star to BSCom. The observations were carried out near two subsequent culminations, under very good weather conditions. Because of the brightness of HD117876, a large fraction of the incoming light had to be blocked in order to avoid saturation. We reduced the effective aperture of the telescope to the $\sim\,7 \%$ of its original size by placing a circular diaphragm at the front end of the telescope tube. Because of the $\sim\,24'$ angular distance between HD117876 and BD+242598, it was not possible to observe the stars in a common field of view. In order to keep our photometric precision, the two distinct fields were observed repeatedly. We obtained 5, 10, and 9 frames in $B$, $V$, and $I_{\mathrm{C}}$, respectively, for each of the two stars. Data were corrected for atmospheric extinction. The standard magnitudes and their formal statistical errors are shown in Table \[tab:mags\]. For BSCom, magnitude averages are shown, calculated over the complete light curve solution given in Sect. \[subsect:fourier\].
--------------------------- -------- ------- ------- ------------------ ------------------
date nights
$(\mathrm{JD} - 2450000)$ $B$ $V$ $R_{\mathrm{C}}$ $I_{\mathrm{C}}$
$3431$ – $3467$ $19$ $568$ $566$ $561$ $543$
$3744$ – $3762$ $7$ $0$ $148$ $149$ $149$
$3824$ – $3843$ $9$ $0$ $256$ $214$ $243$
total: $35$ $568$ $970$ $924$ $935$
--------------------------- -------- ------- ------- ------------------ ------------------
: BS Com observational statistics[]{data-label="tab:obstat"}
\[tab:mags\]
Star
------------ --------- ------ --------- ------ --------- ------
BD+24 2598 10 608 0 603 0 662
BSCom 12 755 0 286 0 433
errors: $\pm 0$ 0015 $\pm 0$ 0016 $\pm 0$ 0032
: Standard $BVI_{\mathrm{C}}$ magnitudes
[ Magnitude averages are given. The errors are standard deviations of the mean differences between BD+24 2598 and HD117876.]{}
Fourier analysis: the double-mode solution {#subsect:fourier}
------------------------------------------
We performed a standard frequency analysis on the $V$ data by discrete Fourier-transformation. We derived $P_0 = 0.48791$d and $P_1 = 0.36307$d for the periods of the fundamental and the first overtone modes, respectively. Besides the frequencies of the two pulsation modes, $5$ harmonics and $8$ linear combinations of the frequencies corresponding to the above periods were identified in the amplitude spectra. In order to increase the detection probability of potentially existing additional components, we applied the spectrum averaging method described by @nagy06aa. Due to the averaging of the frequency spectra in the $V(RI)_C$ colours, the noise of the residual spectra decreased by a factor of $\sim\,\sqrt{3}$, down to the $\sim\,2$mmag level. Except for the above two components, their harmonics, and linear combinations, no additional components were found. More details on the frequency analysis are given by @dekany07an. Our 15-term Fourier-sum fit gives a full description of the light curves within the observational accuracy. The Fourier-parameters and their errors are given in Table \[tab:fourier\].
Photometric abundance and temperature from $\bld{B, V, I_{\mathrm{C}}}$ colours {#sect:abundances}
=================================================================================
Knowledge of the metallicity and temperature is very important in the determination of the physical parameters of double-mode stars (@cox91apj; @kovacs92aap and @kovacs2000lmc). There are two former estimates of the metallicity of BSCom. @smith90pasp obtained two measurements of $\Delta S = 6.0$ and $8.2$. The second value was taken at lower brightness than the first one, so this, being closer to minimum light yields $\mathrm{[Fe/H]} = -1.53$ on the scale of . @clementini95mnras acquired one spectrum of BSCom with a resolution of 1.4Å and a signal-to-noise ratio of $\sim\,20$. They derived a very low abundance of ${\rm [Fe/H]} \sim\,-2.0$ from the equivalent width of the Calcium K-line. As explained by @clementini95mnras, they took the spectrum ‘at minimum light’, according to their photometric $V$ time series consisting of $97$ points. However, taking into consideration the double-mode behaviour of BSCom yet unknown at that time, the real pulsational phase of the aforementioned observations is ambiguous. (Unfortunately, the pulsational parameters of BSCom derived in this paper do not enable us to predict the light curve accurately enough at the time of these observations.) Therefore, we decided to make an independent estimation of the metallicity and also the effective temperature of BSCom based on our photometric data.
The idea of deriving heavy element abundances from magnitudes measured in proper photometric bands is not new. In his early study, [@sturch] used $UBV$ magnitudes around minimum light to estimate the line blanketing of RRab stars. In a similar application, [@lub] employed the reddening-free Walraven \[$B-L$\] colour index for metallicity determination. Also, we refer to a more recent paper by @twarog03aj, who obtained more precise abundances for open clusters using intermediate-band photometry.
The method represents the most economical approach in terms of getting a first order estimate on \[M/H\]. In a completely formal approach, the method of photometric abundance determination comes from the following idea. In the static stellar atmosphere computations the models are determined by fixing basic input parameters, such as chemical composition, convective parameters (mixing length over atmospheric scale height, turbulent velocity, overshooting, etc.), gravitational acceleration $g$ and effective temperature $T_\mathrm{eff}$. If we fix the admittedly poorly known convective parameters and assume that solar-type abundance distribution is, in general, a good approximation of the true distribution, we are left with three parameters only: overall metal abundance \[M/H\], $\log{g}$, and $T_\mathrm{eff}$. Then, it is obvious that once we have an independent estimate on $\log{g}$, then we can compute both $T_\mathrm{eff}$ and \[M/H\] from two colour indices, assuming that they depend differently on the above two parameters. Because of the availability of the average $B, V, I_{\mathrm{C}}$ magnitudes of BSCom and because the $B-V$ colour index relatively strongly depends on \[M/H\] [see e.g. @kw99apj], whereas $V-I_{\mathrm{C}}$ is mostly sensitive to $T_\mathrm{eff}$, it is useful to examine the possibility of determining \[M/H\] and $T_{\rm eff}$ through the $B-V$ and $V-I_{\mathrm{C}}$ colour indices.
Because of the poor representation of the $B-V$, \[M/H\], $\log{g}$, and $T_\mathrm{eff}$ dependence by a single linear or low-order polynomial formula, we decided to use directly the grids available for stellar atmosphere models. We employ quadratic interpolation in order to get accurate model colour indices for each set of physical parameters. More precisely, having the value of $\log{g}$ fixed, we scan the parameter range in $B-V$, $V-I_{\mathrm{C}}$ and find the best \[M/H\], $\log{T_\mathrm{eff}}$ parameters that minimise the following function:
[ $$\begin{aligned}
\label{discriminator}
\mathcal{D} & = &
\left[ \log{T_\mathrm{eff}} (B-V) -
\log{T_\mathrm{eff}} (V-I_{\mathrm{C}}) \right]^2
\nonumber \\
& & + \alpha^2_{B-V}\left[ (B-V)_{\rm obs} - (B-V) \right]^2
\nonumber \\
& & + \alpha^2_{V-I_{\mathrm{C}}}\left[ (V-I_{\mathrm{C}})_{\rm obs} -
(V-I_{\mathrm{C}}) \right]^2~,
\end{aligned}$$]{}
where $T_\mathrm{eff}(B-V)$ and $T_\mathrm{eff}(V-I_{\mathrm{C}})$ are the $T_\mathrm{eff}$ values at the respective colour indices, the subscript ‘obs’ means the observed values, whereas the scanned values are denoted without subscripts. The weights $\alpha_{B-V}$ and $\alpha_{V-I_{\mathrm{C}}}$ are set equal to $0.33$ and $0.25$, respectively, in order to account for the proportionality of $\log{T_\mathrm{eff}}$ to $0.33(B-V)$ and $0.25(V-I_{\mathrm{C}})$ in formulae obtained by linear regressions [see @kw99apj]. The introduction of the last two terms in $\mathcal{D}$ is necessary, because of the high sensitivity of the method for observational noise.
We employ the stellar atmosphere models of @castelli99aap [see also @castelli97aap] computed with the opacity distribution functions based on the solar abundances of @grevesse98ssr (grids quoted as ‘ODFNEW’ at Kurucz’s web site[^4]). These new grids differ from the previous ones as described in @pietrinferni04apj. All models have a microturbulent velocity of $2\,\mathrm{km\,s^{-1}}$, a convective mixing length parameter of $1.25$, scaled solar heavy element distribution and no convective overshooting. For the computation of $\log{g}$ we used a similar formula to that in @kovacs2000smc, adapted to the appropriate parameter regime of RR Lyrae stars: $$\begin{aligned}
\label{logg}
\log{g} = 2.47 - 1.23\log{P_0}~.
\end{aligned}$$
We note two possible sources of ambiguity when using Eq. (\[discriminator\]) for the determination of $T_{\rm eff}$ and \[M/H\]. First, the zero points of the various colour—$T_{\rm eff}$ calibrations (e.g. infrared flux method, IRFM versus calibrations based on Vega or Sun-like stars) may be slightly different. In earlier applications we used zero points defined by IRFM [see e.g. @kovacs2000smc]. For pulsating stars, there is also a problem of the difference between the static and pulsation-phase-averaged colours [@bono95apjs]. Since the sizes of these effects are still debatable, we opted to use the stellar atmosphere data by @castelli99aap without applying any corrections to the magnitude-averaged colours. As we shall see in the subsequent test, the derived metallicities are in a good agreement with the spectroscopic ones, indicating that the theoretical colours are fairly accurate.
![ \[Fe/H\]$_{\mathrm{ph}}$ values computed from the $BVI_{\mathrm{C}}$ colours plotted against the spectroscopic \[Fe/H\]$_{\mathrm{sp}}$ values taken from the literature for 24 RRab stars. Error bars denote standard deviations for each star, obtained by adding uncorrelated Gaussian random noise with $\sigma = 0$$005$ to the observed $B-V$, $V-I_{\mathrm{C}}$ colour indices. Open circles show the outlying stars XAri, SSLeo, and VYSer. []{data-label="fig:bvifeh"}](fig1_revised.eps){width="84mm"}
To test the reliability of the method we collected 24 fundamental mode RR Lyrae (RRab) stars from the literature with known periods, $B-V$ and $V-I_{\mathrm{C}}$ average colour indices, and spectroscopic \[Fe/H\] values[^5]. Details on this test sample are given in Appendix \[appendix:teststars\]. In Fig. \[fig:bvifeh\] we show the derived \[Fe/H\]$_{\mathrm{ph}}$ values versus the \[Fe/H\]$_{\mathrm{sp}}$ spectroscopic metallicities of the test stars. By the exclusion of the three stars XAri, SSLeo, and VYSer with discrepant photometric metallicities (open circles in Fig. 1), the standard deviation of the fit decreases from $0.26$ dex to $0.15$ dex. We consider the latter as the $1\,\sigma$ error of the method. Estimates of the errors of the individual data points were computed by adding Gaussian random noise of $\sigma = 0$$005$ to the input colour indices. The error bars in Fig. \[fig:bvifeh\] indicate that the sensitivity of the method to observational errors increases with decreasing \[Fe/H\] (see also Kovács 2008 for more details).
We estimated the metallicity and effective temperature of BSCom by using $\log g = 2.85$ from Eq. (\[logg\]) and the average $B-V$ and $V-I_{\mathrm C}$ values given in Table \[tab:mags\]. The amount of reddening at the high Galactic latitude of BSCom ($l = 19\fdg00$, $b = +79\fdg86$) is expected to be low. Indeed, according to @schlegel98apj, $\mathrm{E}(B-V) = 0.014$. This gives a maximum value for the real reddening of the star, and therefore poses minimum values to its intrinsic colours. The estimated metallicities and effective temperatures of BSCom with different possible reddening values are shown in Table \[tab:bsctfeh\]. Formal statistical errors are $10$K and $0.11$ dex, computed the same way as for the test objects, but using the photometric errors listed in Table \[tab:mags\].
Pulsation models
================
We have generated linear non-adiabatic (LNA), fully radiative stellar models to derive the fundamental physical parameters of BSCom. The pulsation code is basically identical to the one described by @buchler90, originally developed by @stellingwerf75 [see also @castor71]. As in @kovacs06mmsai, we used the Rosseland mean opacity values taken from @iglesias96apj and @alexander94 interpolated to the appropriate chemical composition. The heavy element distribution is solar-scaled, as given by @grevesse91. The stellar envelopes were densely sampled with 500 mass shells down to the inner radius and mass values, where $Q_{\rm in}\equiv (R_{\rm in}/R_{\rm surf})(M_{\rm in}/M_{\rm surf})=0.05$, with $R_{\rm surf}$ and $M_{\rm surf}$ values at the outermost grid point. This zoning resulted in a deep envelope down to $(6-7)\cdot10^6$K and ensured a sufficient (better than $10^{-4}$ d) stability in the periods (as compared to shallower envelopes of $Q_{\rm in} = 0.10$).
\[tab:bsctfeh\]
$\mathrm{E}(B-V)$ \[Fe/H\] $T_{\mathrm{eff}}$
------------------- ---------- --------------------
$0.000$ $-1.47$ $6774$ K
$0.005$ $-1.51$ $6798$ K
$0.010$ $-1.54$ $6823$ K
$0.014^*$ $-1.58$ $6842$ K
: Reddening dependence of the photometric \[Fe/H\] and $T_{\mathrm{eff}}$ values of BSCom.
[$^{*}$ @schlegel98apj]{}
We covered the metallicity and temperature ranges well beyond the ones derived from the $BVI_{\rm C}$ colours in Sect. \[sect:abundances\]. By scanning these ranges and matching the observed periods, we obtained unique solutions for the possible physical parameters of the star. These solutions are listed in Table \[tab:lnagrid\]. To be more compatible with the evolutionary models derived by @pietrinferni04apj, with the change of the overall metal content $Z$, we have also changed the hydrogen abundance $X$. We note that our best estimated \[Fe/H\] and $T_{\rm eff}$ values are closest to the item at $Z=0.0006$ and $T_{\rm eff}=6850$K. It is noticeable, how sensitive the solution is to the metal content. For example, by changing \[Fe/H\] by $\sim \,0.2$dex, the mass changes by $0.1$–$0.15$$M_{\odot}$. However, a $200$K change in $T_{\rm eff}$ causes a change in the mass less than $0.06$$M_{\odot}$. We also draw attention to the remarkable stability of the gravity and density. While the \[Fe/H\] change mentioned above causes $\sim\,15$ % change in mass, the same figure is less than $0.8$% and $0.3$% in $\log g$ and $\log \rho$, respectively. The effect of temperature change is even smaller. It is also important to note that all models listed in the table are linearly excited in their fundamental and first overtone modes. Therefore, they are viable for sustained double-mode pulsations. More detailed tests by direct nonlinear hydrodynamic simulations are beyond the scope of this paper (see, however Section \[sect:esps\]).
$T_{\rm eff}$ $M$ $\log L$ $R$ $\log g$ $\log \rho$
--------------- ------- ---------- ---------------- ---------- -------------
[$Z=0.0003$]{}
6700 0.639 1.647 4.940 2.856 $-2.127$
6750 0.646 1.664 4.964 2.857 $-2.128$
6800 0.656 1.682 4.993 2.858 $-2.130$
6850 0.663 1.699 5.018 2.858 $-2.131$
6900 0.673 1.717 5.049 2.860 $-2.133$
[$Z=0.0004$]{}
6700 0.669 1.661 5.021 2.861 $-2.128$
6750 0.678 1.679 5.050 2.862 $-2.130$
6800 0.688 1.697 5.080 2.863 $-2.131$
6850 0.698 1.715 5.111 2.865 $-2.133$
6900 0.709 1.733 5.143 2.866 $-2.134$
[$Z=0.0006$]{}
6700 0.732 1.690 5.188 2.872 $-2.132$
6750 0.743 1.708 5.221 2.873 $-2.133$
6800 0.754 1.726 5.253 2.875 $-2.135$
6850 0.768 1.745 5.291 2.876 $-2.137$
6900 0.780 1.763 5.324 2.877 $-2.138$
[$Z=0.0010$]{}
6700 0.872 1.745 5.530 2.893 $-2.139$
6750 0.885 1.763 5.563 2.894 $-2.140$
6800 0.901 1.782 5.603 2.896 $-2.142$
6850 0.914 1.800 5.637 2.897 $-2.143$
6900 0.929 1.818 5.672 2.898 $-2.145$
: Pulsation models matching the periods of BSCom[]{data-label="tab:lnagrid"}
[ $T_{\rm eff}$ is given in \[K\], $M$, $\log L$ and $R$ are measured in solar units, $g$ and $\rho$ are in \[CGS\]. Observed periods match within $10^{-5}$ d.]{}
Combination of the pulsational and evolutionary models {#sect:esps}
======================================================
Stellar evolutionary models play a crucial role in understanding basic stellar physics and developing more coherent models. For example, the ‘Cepheid mass discrepancy’ problem before 1992 appeared as a basic confrontation between the ‘standard’ evolutionary and pulsational theories. A desperate search for the clue of this discrepancy led Norman Simon to a plea for the re-visitation of the heavy element opacities [@simon82apjl], that has resulted finally in a success in 1992 (@rogers92apjs, see also @seaton94mnras). In another application of the evolutionary results, in the current search for extrasolar planets, a generally used method for determining the physical parameters of the host star is matching stellar isochrones in the $\rho_{\rm star}$ – $T_{\rm eff}$ plane to the observed values (assuming that $T_{\rm eff}$ and \[Fe/H\] are reliably determined from e.g. spectroscopy, see @sozzetti07apj).
From a completely formal point of view, the method of parameter determination by the combination of evolutionary and pulsational models is equivalent to the following problem. At a certain pair of periods, the LNA pulsational model grid (Table \[tab:lnagrid\]) gives stellar mass and luminosity as a function of temperature and metallicity. Moreover, HB evolutionary tracks of @pietrinferni04apj also give stellar mass and luminosity for arbitrary temperature, metallicity, and age values. Therefore, we have the following two vector functions sampled in discrete points: $$\begin{aligned}
\label{esps}
\nonumber &{\mathbf E}:&~(\mathrm{[Fe/H]}, T_{\rm eff}, t) \rightarrow (M, L)\\
\nonumber &{\mathbf P}:&~(\mathrm{[Fe/H]}, T_{\rm eff}, P_0, P_1) \rightarrow (M, L)~,
\end{aligned}$$ where ${\mathbf E}$ stands for the evolutionary models, ${\mathbf P}$ denotes the pulsational models, and $t$ is the time elapsed from the start of the core helium burning at the zero age horizontal branch (ZAHB). If we assume consistency between the two kinds of models, then the subspace of stellar parameters that fulfil the constraints posed by both models is given by the intersection of the two vector functions $({\mathbf E} = {\mathbf P})$. This will be (at the most) a 1D subspace, i.e. a set of solutions that has age $t$ as a free parameter. The \[Fe/H\] and $T_{\rm eff}$ ranges corresponding to our multitude of solutions will then be compared with the ones obtained from the photometry in Section \[sect:abundances\]. The method is implemented as follows.
\[fig:onesolution\] ![Stellar mass and luminosity as functions of the metallicity. Age and temperature are fixed at $\log(t[\mathrm{yr}])= 7.424$ and $T_{\rm eff} = 6851$K. Continuous and dashed lines show the quadratic interpolants in Z for the evolutionary (squares) and pulsational (circles) models. For better visibility, the intersections are connected by a vertical grey line. ](fig2_revised.eps "fig:"){width="84mm"}
(a)
: We construct isochrones from the solar-scaled horizontal branch models of @pietrinferni04apj. Physical ingredients and details of the model building are described in their paper and references therein. Here we note only some relevant features. All models are canonical in the sense that they have been calculated without atomic diffusion and convective overshooting. They include up-to-date input physics, most importantly, they use recent and accurate results for plasma-neutrino processes and electron conduction opacities which lead to more accurate He core masses at evolutionary stages after the core helium flash. And last, but not least, their HB models have a fine mass spacing, particularly suitable for investigating stellar parameters in the instability strip.
(b)
: From the pulsational solution (PS) we get ‘near straight lines’ for any fixed $T_{\rm eff}$ in the $Z~\rightarrow~(M, \log L)$ planes (hereafter ML planes).
(c)
: For a given age we can search for the evolutionary solution (ES) by using the above fixed $T_{\rm eff}$. These solutions yield another set of ‘near straight lines’ in the ML planes.
(d)
: The ES and PS lines may or may not intersect in the appropriate $Z$ range in the ML planes. If they do, then we have two $Z$ values at which intersection occurs. Obviously, the consistent $\mathrm{ES} = \mathrm{PS}$ solution should yield the same $Z$. Since this is generally not the case, the best solution is searched for by minimising the difference between the two metallicities: $$\begin{aligned}
\label{eqn:deltaz}
\Delta Z & = & |\log Z_M - \log Z_L|\hskip 2mm ,
\end{aligned}$$ where $Z_M$ and $Z_L$ are the metallicities obtained in the M and L planes, respectively. To illustrate the solution obtained at a given age and temperature, we show the PS and ES curves and their intersections in Fig. 2. It is seen that these curves are indeed near linear. Therefore, we can employ quadratic interpolation to compute the positions of the intersections.
(e)
: By changing the age, with the above method we can map the solution as a function of the evolutionary stage. As it follows from the above construction, at each stage we have different stellar parameters that yield the same periods and satisfy evolutionary constraints. The grey-scale plot of ${\Delta Z}$ on the $\log(t)$—$T_{\rm eff}$ plane is shown in Fig. 3. It is remarkable how much the solution is limited in a narrow $T_{\rm eff}$ regime. It is also observable that we get solutions (i.e. ${\Delta Z}<10^{-4}$) at nearly all ages up to $\log(t)=7.77$ ($\approx60$ Myr). We could, in principle, fix the age by using very accurately (and independently) determined $T_{\rm eff}$. However, the precision of the currently available methods for temperature estimation is insufficient for posing a strong constraint on the age.
\[tab:bscparam\]
\[K\] $M/M_{\odot}$ $\log(L/L_{\odot})$
------------------ --------------- ------------------- ---------------------
$-1.67 \pm 0.01$ $6840 \pm 14$ $0.698 \pm 0.004$ $1.712 \pm 0.005$
$-1.58 \pm 0.11$ $6842 \pm 10$
: Physical parameters of BSCom derived in this paper
[ Errors for the ES/PS solution (first line) are the standard deviations of the various age-dependent solutions in the full age interval of $60$ Myr. Items derived from the $BVI_{\rm C}$ colours (second line) are taken from Table \[tab:bsctfeh\], errors are formal statistical ones. ]{}
The ridge line in Fig. 3 selects the best fitting models. The age dependencies of the corresponding $\log L$, $M$, and \[Fe/H\] are shown in Fig. 4. The small range of stellar parameters allowed by the solution is remarkable. The stronger topological change occurs at $\log t=7.6$ ($\approx 40$ Myr) just after the blueward loop where evolution starts toward the asymptotic giant branch. Considerably larger ages seem to be less probable because the associated high temperatures can be excluded by the ones derived from the colours. From the above result we can estimate the parameters of BSCom. Assuming equal probability of ages in the range plotted, we can compute the distributions of the various physical parameters. Although these distributions are not symmetric around their most probable values, for the present case this effect is of secondary importance and we compute the standard deviations around the averages as a first approximation of the errors introduced by the effect of age uncertainty. Table \[tab:bscparam\] displays the so-derived stellar parameters, which are compared with the independently estimated $T_{\rm eff}$ and \[Fe/H\] values from Table \[tab:bsctfeh\]. It is reassuring that there is a comfortable overlap between the two sets of parameters, indicating a good level of consistency among the very different ingredients of the above parameter determination.
\[fig:grayscale\] ![ Grey-scale plot of the metallicity difference $\Delta Z$ \[see Eq. (\[eqn:deltaz\])\] on the age—$T_{\rm eff}$ plane. ](fig3_revised.eps "fig:"){width="84mm"}
\[fig:ridge\] {width="84mm"}
The employment of the LNA pulsation models in the above solution presumes that the theoretical periods are applicable to the observed ones that are known to be inherently nonlinear due to the longevity of the observable double-mode state. Since our method strongly relies on this assumption, it is important to examine if the currently available nonlinear models predict some considerable change in the asymptotic nonlinear periods with respect to the LNA periods. We note that this problem is also present in any asteroseismological investigation if the multimode state is the result of some nonlinear phenomenon. However, nonlinear modelling is usually not easy to attempt since most of the multimode pulsators are of non-radial, therefore it would require full (i.e. 3D) hydrodynamics. We can pose the above question only in the case of RRd stars and double-mode Cepheids, since these variables are thought to be pure radial pulsators and we have successful nonlinear modelling available on these variables.
Data on the nonlinear/linear period difference for asymptotic hydrodynamical models are nearly non-existent. In an early assessment [@kovacs01book] we used the somewhat limited and preliminary result obtained by @kollath01book. The nonlinear periods were longer than the linear ones in the order of $\sim\,0.001$ days. The amounts of the period shifts were such that they resulted in a [*decrease*]{} of [*less*]{} than $0.002$ in the period ratio. An ongoing, more extended study by Szabó et al. (to be published) basically confirms the above result. As it is detailed in Appendix \[appendix:lna-nl\], we generated another set of LNA models by artificially decreasing the observed periods of BSCom by the amount predicted from the models of Szabó et al. After repeating the same computation as for the LNA periods, we ended up with stellar parameters that were remarkably close to the ones obtained with the LNA assumption. The solution curves (age vs. stellar parameters) exhibit small overall shifts that result in differences between the parameters obtained with the LNA and modified LNA solutions. The parameter shifts are approximately as follows: $+0.01$, $+0.006$, $+10$K, $-0.08$, for $M/M_{\odot}$, $\log(L/L_{\odot})$, $T_{\rm eff}$ and \[Fe/H\], respectively, in the sense of modified LNA minus non-modified LNA. We conclude that considering the period shifts predicted by the currently available nonlinear simulations does not effect significantly the agreement with the independently estimated $T_{\rm eff}$ and \[Fe/H\].
Conclusions
===========
We performed a pulsational/evolutionary analysis on one of the most extensively observed Galactic double-mode RR Lyrae (RRd) star BSCom. Our earlier analysis incorporating all of our measurements in the $BV(RI)_{\rm C}$ colours has shown that the light variation does not contain any other components except for those of the fundamental and first overtone modes together with their linear combinations [@dekany07an]. Therefore, the presence of possible non-radial components is excluded above the $\sim$mmag level. (This may not be the case for AQ Leo, where the analysis of the high precision data from the MOST satellite suggests the presence of two possibly non-radial components at the sub-mmag level – see @gruberbauer07mnras.) Based on the result of the frequency analysis, we decided to derive the physical parameters of BSCom using the assumption of radial double-mode pulsation.
Earlier studies on double-mode variables were mostly based on the period ratio (Petersen) diagram, which is an effective (but not complete) representation of the main (mass and heavy element) dependences of the periods [see e.g. @popielski2000aca]. It is also possible to employ other available pieces of information, such as colour or overall cluster metallicity [e.g. @kovacs2000smc]. In this paper we followed a [*purely*]{} theoretical approach, where we derived all basic stellar parameters by matching the evolutionary and pulsational properties. In the solutions age remains a free parameter.
It is important to confront the above [*theoretical*]{} solution with other independent pieces of information. For this goal, we utilised the observed $BVI_{\rm C}$ colours in deriving metal abundance and effective temperature with the help of updated solar-scaled stellar atmosphere models by @castelli99aap. We tested the reliability of the method through the comparison of the published spectroscopic abundances of 21 Galactic field fundamental mode RR Lyrae (RRab) stars. The standard deviation of the residuals between the spectroscopic and photometric metallicities is $0.15$ dex, which we consider as a good support for the reliability of the photometric method based on the $BVI_{\rm C}$ colours. By applying this method on BSCom, we get \[Fe/H\]$ = -1.58\pm 0.11$, $T_{\rm eff} = 6842\pm 10$K, where the errors are purely statistical ones.
Taking into consideration the above independently estimated parameters, we can most probably exclude the late evolutionary stages after the core helium exhaustion. This is because the high temperature required by the solution is incompatible with the temperature computed from the observed colours. Therefore, in the acceptable set of solutions the age spans only an approximately 60 Myr range from the zero age horizontal branch till slightly after the start of the final redward evolution. Because we do not have independent information about the evolutionary status, all of these solutions are viable. However, the ranges covered by the parameters in the possible time interval are remarkably narrow. Furthermore, assuming uniform age distribution, most of the values are clumped in an even narrower regime at the earlier evolutionary stages. By considering this age ambiguity, we get $M / M_{\odot} = 0.698 \pm 0.004$, $\log(L/L_{\odot}) = 1.712 \pm 0.005$, $T_{\rm eff} = 6840 \pm 14$K, \[Fe/H\]=$-1.67 \pm 0.01$ where the errors are $1\,\sigma$ ranges due to the above age interval.
It is interesting to compare the above values with the compilation of the average physical parameters of the RRd stars found in various stellar systems. In Table 2 of the paper by @kovacs2001aap we summarised the result for four systems, where the parameter determinations were performed without the use of evolutionary models. There are IC4499 and LMC that have average metallicities similar to that of BSCom. Omitting LMC, due to its very wide population range, we see that the average RRd parameters in IC4499 ($T_{\rm eff} = 6760$K, $\log L / L_{\odot} = 1.708$, $M / M_{\odot} = 0.755$) are remarkably similar to those of BSCom. The largest difference occurs in the mass, but this is due to the slightly longer period of BSCom than the average value of the RRd stars in IC4499. Since the other stellar parameters are nearly the same, this period difference results in a larger average mass for the cluster stars.
Another indirect check of the consistency of our theoretical parameters of BSCom is the computation of the zero point of the period-luminosity-colour (PLC) relation of @kw01aap. Although this relation was derived from fundamental mode RR Lyrae stars, because of its apparent generality, we use it with the period of the fundamental mode of BSCom. By using $M_{\rm bol,\odot} = 4.75$, a bolometric correction interpolated to the appropriate stellar parameters of ${\rm BC} = -0.15$ [see @castelli99aap] and a reddening value given by @schlegel98apj [see Table 3], we get $-1.02$ for the PLC zero point. In comparing this value with the one published by @kovacs03mnras, we see that the above value is larger by $0.06$, yielding lower distance modulus for the LMC by the same amount (we recall that the LMC distance modulus with the zero point of @kovacs03mnras comes out to be $18.55$). We think that the above result confirms the consistency of the derived physical parameters, especially if we add the ambiguity due to the yet to be shown fit of the RRd variables to the PLC relation derived from RRab stars.
We have also tested the effect of a possible period shift due to convection and nonlinearity. Based on the extensive set of models by Szabó et al. (in preparation), our conclusion is that the effect in the derived parameters is relatively small ($\sim\,1.5\,\%$). Once this correction is proved to be important by future studies, taking it into consideration is fairly easy, since the overall period shift seems to be almost independent of the particular model parameters. Therefore, there is no need to run specific nonlinear models in each case but it is quite enough to fit the relatively inexpensive linear models to the modified periods based on the above general corrections.
The remarkably small ambiguity of the derived stellar parameters from the purely theoretical method employed in this paper brings up several intriguing questions. One of these is the compatibility of the theoretical abundance and temperature and the ones obtained by spectroscopic and various semi-empirical methods (e.g. by the widely used infrared flux method). Accuracies such as $1\%$ (or better) are quoted for the temperature in the current papers on this topic [e.g. @masana06aap]. Unfortunately, for the chemical composition we still lack this level of accuracy. Most of the abundances available for RR Lyrae stars are based on some calibration of low-dispersion spectra by a low number of high-dispersion ones. Double-mode stars in globular clusters are obvious targets for the above-type of studies, due to the expected chemical homogeneity and same distance. Last, but not least, we note the need of putting more stringent constraints on the nonlinear hydrodynamical models. Our current understanding is that double-mode behaviour is caused by the non-resonant interaction of two normal modes with a delicate level of physical dissipation (understood as the result of convection). Although the global observational and theoretical properties are in reasonable agreement, the accuracy of the presently available stellar parameters is insufficient to lend further support to these models. Considering the general astrophysical importance of convection and, in particular, the role of that in double-mode pulsation, it is clear that pinning down the physical parameters is of great significance.
Acknowledgements {#acknowledgements .unnumbered}
================
We are very grateful to Zoltán Kolláth for lending his opacity interpolation code. Thanks are also due to László Szabados for his help in the language correction of the paper. R.Sz. is grateful to the Hungarian NIIF Supercomputing Center for providing resources for the nonlinear model computations (project No. 1107). We are also grateful to Csaba Komáromi for his valuable help in constructing the aperture stop to obtain standard magnitudes of BSCom. This work has been supported by the Hungarian Research Foundation (OTKA) grants K-60750 (to G.K.) and K-68626 (to J.J.).
D. R., [Ferguson]{} J. W., 1994, ApJ, 437, 879
G. [Á]{}., 1999, Konkoly Obs. Occ. Techn. Notes, 11 http://www.konkoly.hu/Mitteilungen/Mitteilungen.html
G. F. et al. 2007, AJ, 133, 1810
V. M., 1992, AJ, 104, 734
G., [Caputo]{} F., [Stellingwerf]{} R. F., 1995, ApJS, 99, 263
J. R., 1990, NATO ASI series, 302, 1
J. R., [Szab[ó]{}]{} R., 2007, ApJ, 660, 723
F., 1999, A&A, 346, 564
F., [Gratton]{} R. G., [Kurucz]{} R. L., 1997, A&A, 318, 841
J. I., 1971, ApJ, 166, 109
G., [Carretta]{} E., [Gratton]{} R., [Merighi]{} R., [Mould]{} J. R., [McCarthy]{} J. K., 1995a, AJ, 110, 2319
G., [Tosi]{} M., [Bragaglia]{} A., [Merighi]{} R., [Maceroni]{} C., 1995b, MNRAS, 275, 929
A. N., 1991, ApJ, 381, L71
I., 2007, AN, 328, 833
M. U., 1998, A&A, 337, L29
P. et al., 2007, A&A, 476, 73
E., [Clement]{} C. M., 1997, AJ, 114, 1190
N., [Anders]{} E., 1991, in Solar interior and atmosphere. University of Arizona Press, 1991, p. 1227
N., [Sauval]{} A. J., 1998, Space Sci. Rev., 85, 161
M. et al., 2007, MNRAS, 379, 1498
J. A., [Nations]{} H. L., 1984, AJ, 89, 391
C. A., [Rogers]{} F. J., 1996, ApJ, 464, 943
D. H. P., 1973, ApJS, 25, 487
J., 1995, AcA, 45, 653
J., [Kovács]{} G., 1996, A&A, 312, 111
Z., [Beaulieu]{} J. P., [Buchler]{} J. R., [Yecko]{} P., 1998, ApJ, 502, L55
Z., [Buchler]{} J. R., 2001, ASSL, 257, 29
G., 2000a, A&A, 360, L1
G., 2000b, A&A, 363, L1
G., 2001a, A&A, 375, 469
G., 2001b, ASSL, 257, 61
G., 2003, MNRAS, 342, L58
G., 2006, MmSAI, 77, 160
G., 2008, to appear in Goupil M. J. et al., eds, Nonlinear Stellar Hydrodynamics and Pulsations of Cepheids, EDP Sci.
G., [Buchler]{} J. R., 1993, ApJ, 404, 765
G., [Jurcsik]{} J., 1997, A&A, 322, 218
G., [Kanbur]{} S. M., 1998, MNRAS, 295, 834
G., [Walker]{} A. R., 1999, ApJ, 512, 271
G., [Walker]{} A. R., 2001, A&A, 371, 579
G., [Buchler]{} J. R., [Marom]{} A., [Iglesias]{} C. A., [Rogers]{} F. J., 1992, A&A, 259, L46
J., 1979, AJ, 84, 383
E., [Jordi]{} C., [Ribas]{} I., 2006, A&A, 450, 735
D. H., [Clementini]{} G., [Marconi]{} M., 2007, AJ, 133, 2752
T., 2003, AcA, 53, 307
P., [Poretti]{} E., 2003, A&A, 398, 213
A., [Kov[á]{}cs]{} G., 2006, A&A, 454, 257
A., [Cassisi]{} S., [Salaris]{} M., [Castelli]{} F., 2004, ApJ, 612, 168
A., [Kolaczkowski]{} Z., [Kopacki]{} G., 2003, AcA, 53, 27
B. L., [Dziembowski]{} W. A., [Cassisi]{} S., 2000, AcA, 50, 491
F. J., [Iglesias]{} C. A., 1992, ApJS, 79, 507
D. J., [Finkbeiner]{} D. P., [Davis]{} M., 1998, ApJ, 500, 525
M. J., [Yan]{} Y., [Mihalas]{} D., [Pradhan]{} A. K., 1994, MNRAS, 266, 805
N. R., 1982, ApJ, 260, L87
H. A., 1990, PASP, 102, 124
., 2007, AN, 328, 829
A., [Torres]{} G., [Charbonneau]{} D., [Latham]{} D. W., [Holman]{} M. J., [Winn]{} J. N., [Laird]{} J. B., [O’Donovan]{} F. T., 2007, ApJ, 664, 1190
R. F., 1975, ApJ, 195, 441
J., [Carney]{} B. W., [Gieren]{} W. P., [Fouqu[é]{}]{} P., [Latham]{} D. W., [Fry]{} A. M., 2004, A&A, 415, 531
C., 1966, ApJ, 143, 774
R., [Koll[á]{}th]{} Z., [Buchler]{} J. R., 2004, A&A, 425, 627
B. A., [Anthony-Twarog]{} B. J., [De Lee]{} N., 2003, AJ, 125, 1383
F., [Feast]{} M. W., [Whitelock]{} P. A., [Laney]{} C. D., 2007, MNRAS, 379, 723
P., 2006, IBVS, 5685, 1
Fourier decomposition of BS Comae {#appendix:decomp}
=================================
Table \[tab:fourier\] shows the frequencies, amplitudes and phases of the 15-term Fourier sum fitted to the light curves of BSCom. Phases refer to sine-term decomposition with an epoch $t_0 = 2453431.0$. The Fourier-sums fit the $B, V, R_C, I_{\mathrm{C}}$ data with $0.0116$, $0.0097$, $0.0091$, $0.0098$ mag rms scatters, respectively. These values are in the appropriate proximity of the formal errors of the individual photometric data points as derived from the IRAF package. To determine the errors of the fitted amplitudes and phases, we performed Monte Carlo simulations of the light curves. Synthetic data were generated by adding independent Gaussian random noise to the Fourier solution, with $\sigma$ made equal to the rms scatter of the residual data. Errors in Table \[tab:fourier\] are the standard deviations of the Fourier parameters obtained from 100 independent realisations of the light curve in each colour.
Frequency
-------- ----------- ------------- ------------------- ------------------- ------------------- -------------------
$f_0$ – $2.04955$ $164.26 \pm 0.77$ $127.05 \pm 0.47$ $101.18 \pm 0.45$ $79.63 \pm 0.51$
$2f_0$ – $4.09910$ $26.51 \pm 0.70$ $20.87 \pm 0.47$ $17.26 \pm 0.46$ $14.88 \pm 0.48$
$3f_0$ – $6.14865$ $4.43 \pm 0.78$ $1.82 \pm 0.47$ $1.44 \pm 0.45$ $2.48 \pm 0.47$
– $f_1$ $2.75432$ $259.43 \pm 0.76$ $201.66 \pm 0.46$ $161.36 \pm 0.42$ $123.06 \pm 0.51$
– $2f_1$ $5.50864$ $35.55 \pm 0.83$ $29.08 \pm 0.49$ $22.76 \pm 0.50$ $17.42 \pm 0.47$
– $3f_1$ $8.26296$ $15.57 \pm 0.68$ $12.48 \pm 0.46$ $9.87 \pm 0.49$ $6.86 \pm 0.44$
– $4f_1$ $11.01728$ $3.75 \pm 0.65$ $3.21 \pm 0.42$ $2.94 \pm 0.45$ $2.48 \pm 0.47$
$f_0$ $f_1$ $4.80387$ $62.50 \pm 0.65$ $50.23 \pm 0.48$ $40.87 \pm 0.54$ $32.82 \pm 0.54$
$f_0$ $-f_1$ $0.70477$ $36.19 \pm 0.68$ $29.65 \pm 0.43$ $24.98 \pm 0.49$ $18.10 \pm 0.53$
$2f_0$ $f_1$ $6.85342$ $11.59 \pm 0.78$ $9.75 \pm 0.51$ $8.54 \pm 0.42$ $6.86 \pm 0.52$
$f_0$ $2f_1$ $7.55819$ $21.67 \pm 0.74$ $20.31 \pm 0.47$ $15.30 \pm 0.49$ $11.38 \pm 0.57$
$2f_0$ $2f_1$ $9.60774$ $10.85 \pm 0.77$ $6.08 \pm 0.46$ $4.99 \pm 0.41$ $3.14 \pm 0.45$
$f_0$ $3f_1$ $10.31251$ $8.04 \pm 0.79$ $6.46 \pm 0.46$ $4.28 \pm 0.42$ $3.61 \pm 0.50$
$f_0$ $-3f_1$ $6.21341$ $5.09 \pm 0.83$ $2.63 \pm 0.50$ $3.52 \pm 0.47$ $2.24 \pm 0.50$
$f_0$ $4f_1$ $13.06683$ $3.35 \pm 0.56$ $2.56 \pm 0.45$ $1.78 \pm 0.43$ $1.46 \pm 0.43$
– $f_1$ $2.754320$ $3.121 \pm 0.004$ $3.080 \pm 0.004$ $3.034 \pm 0.005$ $2.960 \pm 0.006$
– $2f_1$ $5.508640$ $2.255 \pm 0.030$ $2.199 \pm 0.019$ $2.304 \pm 0.026$ $2.210 \pm 0.032$
– $3f_1$ $8.262960$ $1.023 \pm 0.176$ $1.562 \pm 0.273$ $2.202 \pm 0.355$ $1.433 \pm 0.260$
– $4f_1$ $11.017280$ $2.216 \pm 0.003$ $2.200 \pm 0.003$ $2.172 \pm 0.003$ $2.104 \pm 0.004$
$f_0$ – $2.049550$ $1.560 \pm 0.020$ $1.534 \pm 0.016$ $1.502 \pm 0.019$ $1.569 \pm 0.030$
$2f_0$ – $4.099100$ $0.438 \pm 0.046$ $0.433 \pm 0.035$ $0.426 \pm 0.048$ $0.457 \pm 0.074$
$3f_0$ – $6.148650$ $5.553 \pm 0.180$ $5.992 \pm 0.133$ $5.746 \pm 0.137$ $5.727 \pm 0.169$
$f_0$ $f_1$ $4.803870$ $1.564 \pm 0.014$ $1.576 \pm 0.008$ $1.602 \pm 0.013$ $1.542 \pm 0.014$
$f_0$ $-f_1$ $0.704770$ $4.576 \pm 0.022$ $4.616 \pm 0.015$ $4.657 \pm 0.018$ $4.682 \pm 0.024$
$2f_0$ $f_1$ $6.853420$ $0.875 \pm 0.054$ $0.829 \pm 0.046$ $0.931 \pm 0.059$ $0.793 \pm 0.066$
$f_0$ $2f_1$ $7.558190$ $0.234 \pm 0.035$ $0.251 \pm 0.023$ $0.250 \pm 0.027$ $0.316 \pm 0.045$
$2f_0$ $2f_1$ $9.607740$ $0.319 \pm 0.071$ $0.159 \pm 0.077$ $0.113 \pm 0.071$ $0.178 \pm 0.135$
$f_0$ $3f_1$ $10.312510$ $0.698 \pm 0.097$ $0.359 \pm 0.082$ $0.102 \pm 0.109$ $6.266 \pm 0.114$
$f_0$ $-3f_1$ $6.213410$ $4.988 \pm 0.162$ $5.180 \pm 0.199$ $5.342 \pm 0.145$ $5.418 \pm 0.236$
$f_0$ $4f_1$ $13.066830$ $5.980 \pm 0.164$ $5.303 \pm 0.170$ $5.341 \pm 0.232$ $4.824 \pm 0.358$
[ The light curves are represented in the from of $A_0 + A_1\cdot\sin(2\pi f (t - t_0) + \Phi_1) + \dots$\
with $t_0 = 2453431.0$. Frequency is given in d$^{-1}$, amplitudes and phases are in mmag and radians, respectively. We show $1\sigma$ formal statistical errors.]{}
The sample for testing the $\bld{BVI_C}~-~\bld{[F\lowercase{e}/H],T_{\lowercase{eff}}}$ relation {#appendix:teststars}
================================================================================================
Here we present the parameters of the 24 RRab used in Sect. \[sect:abundances\] to test the method of metallicity and temperature estimation from the $BVI_{\rm C}$ colours and fundamental periods. Table \[tab:teststars\] shows fundamental periods, colour indices, interstellar reddenings and spectroscopic \[Fe/H\]$_{\mathrm{sp}}$ values taken from the literature. We relied mainly on the compilations by @kj97aap for colour indices and on @jk96 for \[Fe/H\], except for three stars. For the abundances of UUCet and V440Sgr we refer to @clementini95aj and for RVPhe to @jones73apjs. Reddenings were taken from @blanco92aj, except for BBPup for which we used the E($B-V$) value given by @schlegel98apj. The derived \[Fe/H\]$_{\mathrm{ph}}$ photometric abundances for each star are listed in the last column of Table \[tab:teststars\]. For XAri, SSLeo, and VYSer, we get discrepant \[Fe/H\]$_{\mathrm{ph}}$ values. These might be due to systematic errors in their colour indices, which have a particularly large effect on the estimated \[Fe/H\]$_{\mathrm{ph}}$ at lower metallicities (see Kovács 2008). We also note that SSLeo and VYSer have also been met as outliers in various earlier studies [e.g. @kovacskanbur98; @kw01aap; @kovacs03mnras].
$P_{\mathrm FU}$ $B-V$ $V-I_{\mathrm{C}}$ $\mathrm{E}(B-V)$ \[Fe/H\]$_{\mathrm{sp}}$ \[Fe/H\]$_{\mathrm{ph}}$
--------- ------------------ --------- -------------------- ------------------- -------------------------- --------------------------
SWAnd $0.4422659$ $0.434$ $0.544$ $0.09$ $-0.06$ $-0.22$
WYAnt $0.5743365$ $0.384$ $0.549$ $0.05$ $-1.39$ $-1.13$
XAri $0.6511597$ $0.490$ $0.724$ $0.16$ $-2.10$ $-2.50$
RRCet $0.5530288$ $0.371$ $0.537$ $0.05$ $-1.29$ $-1.28$
UUCet $0.6053409$ $0.390$ $0.567$ $0.01$ $-1.38$ $-1.33$
WCrt $0.4120139$ $0.370$ $0.466$ $0.09$ $-0.45$ $-0.22$
DXDel $0.4726182$ $0.454$ $0.571$ $0.09$ $-0.32$ $-0.26$
SUDra $0.6604200$ $0.348$ $0.519$ $0.02$ $-1.56$ $-1.50$
SWDra $0.5696710$ $0.369$ $0.520$ $0.01$ $-0.95$ $-0.93$
RXEri $0.5872475$ $0.415$ $0.585$ $0.10$ $-1.07$ $-1.13$
RRGem $0.3973082$ $0.406$ $0.511$ $0.11$ $-0.14$ $-0.23$
RRLeo $0.4523926$ $0.334$ $0.486$ $0.03$ $-1.30$ $-1.23$
SSLeo $0.6263438$ $0.350$ $0.488$ $0.02$ $-1.56$ $-0.82$
TTLyn $0.5974332$ $0.381$ $0.550$ $0.05$ $-1.50$ $-1.26$
V445Oph $0.3970232$ $0.620$ $0.772$ $0.29$ $+0.01$ $-0.18$
AVPeg $0.3903760$ $0.424$ $0.528$ $0.10$ $+0.08$ $-0.17$
ARPer $0.4255489$ $0.676$ $0.853$ $0.35$ $-0.14$ $-0.31$
RVPhe $0.5964182$ $0.370$ $0.535$ $0.03$ $-1.50$ $-1.19$
BBPup $0.4805468$ $0.459$ $0.582$ $0.11$ $-0.35$ $-0.47$
V440Sgr $0.4774788$ $0.404$ $0.568$ $0.09$ $-1.14$ $-1.09$
VYSer $0.7140962$ $0.377$ $0.579$ $0.02$ $-1.71$ $-2.33$
WTuc $0.6422370$ $0.329$ $0.490$ $0.01$ $-1.37$ $-1.39$
TUUMa $0.5576570$ $0.364$ $0.522$ $0.05$ $-1.15$ $-1.15$
UUVir $0.4756062$ $0.349$ $0.475$ $0.02$ $-0.60$ $-0.66$
[ Periods are in \[days\], colour indices and reddenings are in \[mag\]. See Appendix \[appendix:teststars\] for references on the sources of data.]{}
Effect of nonlinearity {#appendix:lna-nl}
======================
\[testlna\]
$T_{\rm eff}$ $M$ $\log L$ $R$ $\log g$ $\log \rho$
--------------- ------- ---------- ---------------- ---------- -------------
[$Z=0.0003$]{}
6700 0.672 1.662 5.026 2.862 $-2.128$
6750 0.679 1.679 5.050 2.863 $-2.129$
6800 0.689 1.697 5.080 2.864 $-2.131$
6850 0.697 1.714 5.105 2.865 $-2.132$
6900 0.707 1.732 5.137 2.866 $-2.134$
[$Z=0.0004$]{}
6700 0.705 1.677 5.114 2.868 $-2.130$
6750 0.715 1.695 5.144 2.869 $-2.131$
6800 0.724 1.713 5.172 2.870 $-2.133$
6850 0.733 1.730 5.200 2.871 $-2.134$
6900 0.745 1.748 5.232 2.872 $-2.135$
[$Z=0.0006$]{}
6700 0.772 1.706 5.287 2.879 $-2.133$
6750 0.786 1.725 5.325 2.881 $-2.135$
6800 0.798 1.743 5.356 2.882 $-2.136$
6850 0.810 1.761 5.389 2.883 $-2.138$
6900 0.820 1.778 5.416 2.884 $-2.139$
[$Z=0.0010$]{}
6700 0.919 1.761 5.633 2.900 $-2.140$
6750 0.930 1.778 5.660 2.901 $-2.141$
6800 0.944 1.796 5.694 2.902 $-2.142$
6850 0.959 1.814 5.728 2.903 $-2.144$
6900 0.971 1.831 5.757 2.904 $-2.145$
: Pulsation models matching the modified periods of BSCom
[ $T_{\rm eff}$ is given in \[K\], $M$, $\log L$ and $R$ are measured in solar units, $g$ and $\rho$ are in \[CGS\]. Observed periods match within $10^{-5}$ d. The modified periods are $P_0=0.486448$ and $P_1=0.362340$ days.]{}
In Szabó et al. (in preparation) we examine the period shifts between the asymptotic nonlinear period of the convective models and those of the purely radiative linear non-adiabatic models (i.e. the ones we use in this paper). The results are based on many different models, including various combinations of physical and model construction parameters (notably depth of the inner boundary and number of mass shells constituting the stellar model). All models show that both modes have longer periods in the asymptotic double-mode regime than those in the LNA approximation. The amount of period increase is smaller than $0.001$ days. The shifts lead to period ratio values that are lower by up to $0.0015$ than the LNA ones. For the overall relative period shifts we get: $$\begin{aligned}
P_{\rm LNA}(FU) / P_{\rm NL}(FU) &=& 0.997 \\
P_{\rm LNA}(FO) / P_{\rm NL}(FO) &=& 0.998~.
\end{aligned}$$ Assuming that the observed periods of BSCom correspond to the nonlinear asymptotic values, the predicted [*modified*]{} LNA values are as follows: $$\begin{aligned}
P_0 (\rm{LNA}, mod.) & = & 0.486448 \\
P_1 (\rm{LNA}, mod.) & = & 0.362340 \\
P_1/P_0 (\rm{LNA}, mod.) & = & 0.744869~.
\end{aligned}$$ With these periods we performed the same type of LNA survey as described in Sect. 5 for the observed periods. The result is shown in Table C1. By comparing the items at the same chemical composition and temperature with those matching the observed periods, we see that $M$ and $\log L$ are shifted to larger values by $\sim\,0.04$ and $\sim\,0.015$, respectively. However, in the final solution the difference between the parameters obtained by the predicted and directly computed LNA periods will be considerably smaller, due to the compensating effect of changes in $T_{\rm eff}$ and \[Fe/H\] (see Sect. 5 for additional details).
\[lastpage\]
[^1]: We note however, that @wils06ibvs selected the $\sim\,1$ yr alias of the fundamental period.
[^2]: Photometric data are available electronically on-line at CDS.
[^3]: IRAF is distributed by the National Optical Astronomy Observatory, which is operated by the Association of Universities for Research in Astronomy, Inc., under cooperative agreement with the National Science Foundation.
[^4]: http://kurucz.harvard.edu
[^5]: For solar-scaled heavy element distribution we have $[\mathrm{M/H}] = [\mathrm{Fe/H}]$.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'We exactly solve a four-site spin model with site-dependent Kitaev’s coupling in a tetrahedron by means of an analytical diagonalization. The non-abelian fusion rules of eigen vortex excitations in this small lattice model are explicitly illustrated in real space by using Pauli matrices. Comparing with solutions of Kitaev models on large lattices, our solution gives an intuitional picture using real space spin configurations to directly express zero modes of Majorana fermions, non-abelian vortices and non-abelian fusion rules. We generalize the single tetrahedron model to a chain model of tetrahedrons on a torus and find the non-abelian vortices become well-defined non-abelian anyons. We believe these manifest results are very helpful to demonstrate the nonabelian anyon in laboratory.'
author:
- Yue Yu
- Tieyan Si
title: 'Explicit illustration of non-abelian fusion rules in a small spin lattice '
---
The spin lattice models of Kitaev-type have attracted many research interests because of the abelian and nonabelian anyons in these exactly soluble two-dimensional models [@kitaev; @kitaev1], which are of the potential application in the topological quantum memory and fault-tolerant topological quantum computation [@rev].
The abelian anyons can be explicitly shown in Kitaev toric code model [@kitaev] or Levin-Wen model [@lw]. Fusion rules and braiding matrix can be easily illustrated in real space. This has simulated many attempts to design and process experiments to demonstrate these abelian anyons in laboratory [@exp].
In solving Kitaev honeycomb model [@kitaev] and its ramifications [@yuwang; @yao1; @wuc], a key technique is the usage of the Majorana fermion representation of the spin-1/2 operators. The systems then are transferred into bilinear fermion systems and the ground state sector can be diagonalized in the momentum space. The shortcoming to use the momentum space is that the ground state and the elementary excitations are hard to be expressed by the original spin operators, i.e., Pauli matrices. On the other hand, the sectors with vortex excitations can only be treated numerically. Then the nonabelian fusion rules and statistics may not be directly shown in Pauli matrices’ language. Lahtinen et al have derived the nonabelian fusions through the spectrum analysis [@la]. However, they are still not directly related to Pauli matrices. A real space form of the ground state of the Kitaev honeycomb model in the insulating phase with abelian anyon was studied [@cn]. We tried to present the non-abelian fusion rules for high energy excitations in this Kitaev model [@yusi]. There was an attempt to use toric code abelian anyons to superpose the Ising non-abelian anyons without involving a Hamiltonian[@wang].
In this paper, we solve a spin model with Kitaev’s coupling in a small lattice, i.e., a tetrahedron (Fig.\[fig:Fig.1\]). We perform Majorana fermions and their zero modes, non-abelian vortices and their fusion rules in real space by means of Pauli matrices. Because the system is finite and everything can be deduced in an elemental way, it will be very helpful to intuitionally understand these concepts which were used to be expressed in those deep mathematical language. Since the spin configurations of these excitations and the fusion rules are explicitly shown, the experimental techniques with cold atoms, finite photon graph states and nuclear magnetic resonance systems [@exp] are possible to be applied to demonstrate these quantum states and then to design quantum bits and gates for a topological quantum computer.
The single tetrahedron model is too small for non-abelian anyons to be well-defined. We generalize it to be a chain model of tetrahedrons on a torus. This chain model is also exactly solvable and the non-abelian anyons can be well-defined. They are still of a simple form like the non-abelian vortex in the single tetrahedron model.
![\[fig:Fig.1\] (a) A tetrahedron in whose points 1,2,3 and 4, the spins live. The four surfaces are labelled by $A, B, C$ and $ D$ or 124, 134, 234 and 123. (b) The top view of the tetrahedron. $x,y,z$ are the links with different Kitaev couplings.](fig1.EPS){width="5cm"}
[*The model and symmetries* ]{} Kitaev model in a tetrahedron is given by $ H_K=\sum_{x}J_{x,ij}\sigma_i^x
\sigma_j^x+\sum_{y}J_{y,ij}\sigma_i^y
\sigma_j^y+\sum_{z}J_{z,ij}\sigma_i^z \sigma_j^z$ where $x,y,z$ are the links shown in Fig. \[fig:Fig.1\](b) and $J_{a,ij}$ are link- and site-dependent coupling constants. $\sigma^a_j$ are spin-1/2 operators obeying Pauli matrix algebra, e.g., $\sigma^x_i\sigma^y_i=i\sigma_i^z$ and $(\sigma^x_i)^2=(\sigma^y_i)^2=(\sigma^z_i)^2=1$. The spin operators on the different sites commute, i.e., $[\sigma^a_i,\sigma^b_j]=0$ for $i\ne j$. An intuitive imagination is the model may be easily solved if $J_{a,ij}$ are not site-dependent as that in Kitaev model in an infinite lattice [@kitaev1]. However, a direct check finds that, unlike Kitaev model in an infinite lattice, this model can not be reduced to a bilinear fermion theory in such a coupling constant choice. This is because the tetrahedron is topologically equivalent to a sphere, Kitaev model defined on this compact space is very different from the model on infinite or periodic lattices. In this paper, we consider these coupling constants are site-dependent. We also include some three- and four-spin coupling terms as those in a generalized Kitaev model [@yuwang]. The model Hamiltonian we will study is given by $$\begin{aligned}
&&H=J_x\sigma_1^x\sigma^x_2+J_x\sigma_3^x\sigma^x_4
+J_y\sigma^y_2\sigma^y_3 \nonumber\\&&+\kappa
\sigma_1^x\sigma_2^z\sigma_3^y+\kappa\sigma_2^y\sigma_3^z\sigma_4^x
+\lambda\sigma_1^x\sigma_2^z\sigma_3^z\sigma_4^x.\label{1}\end{aligned}$$ There are four conserved operators which live in triangular plaquettes: $$\begin{aligned}
&&P=\{P_A=P_{124}=\sigma^z_1\sigma_2^y\sigma^x_4,~
P_B=P_{134}=\sigma^x_1\sigma_3^y\sigma^z_4,\nonumber\\&&P_C=P_{234}=\sigma^x_2\sigma_3^z\sigma^y_4,~P_D=P_{123}
=\sigma^y_1\sigma_2^z\sigma^x_3 \},\end{aligned}$$ which are mutual commutative. They obey $[P_{ijk},H]=0$ and $P_AP_BP_CP_D=1$. Three-spin couplings break time reversal symmetry. The Hamiltonian is time-reversal invariant if $\kappa=0$, .
[*Bilinearization, diagonalization and states* ]{} This Hamiltonian can be transferred into a bilinear fermion Hamiltonian. In the deduction process, we can illustrate some abstract concepts in an elementary way. For example, we can define eight Majorana fermions corresponding to four sites: $$\begin{aligned}
&&\psi_1=\sigma_1^y,\psi_2=\sigma^x_2\sigma^z_1,
\psi_3=\sigma^y_3\sigma_2^z\sigma^z_1,
\psi_4=\sigma^x_4\sigma^z_3\sigma_2^z\sigma^z_1 \label{3}\\
&&b_1=-\sigma_1^x,b_2=-\sigma^y_2\sigma^z_1,
b_3=-\sigma^x_3\sigma_2^z\sigma^z_1,
b_4=-\sigma^y_4\sigma^z_3\sigma_2^z\sigma^z_1\nonumber\end{aligned}$$ They obey $\{\psi_i,\psi_j\}=\{b_i,b_j\}=2\delta_{ij}$ and $\{\psi_i,b_j\}=0$. Remarkably, $[H,b_i]=0$. This can be directly checked or be seen by writing the Hamiltonian in terms of Majorana fermions, $$\begin{aligned}
H=-t(d^\dag_a d_b+d^\dag_b d_a)-\mu(d^\dag_ad_a+ d^\dag_bd_b)
+\Delta^* d_ad_b+\Delta d_b^\dag d^\dag_a,\label{2}\end{aligned}$$ where $ d_a=\frac{1}2(\psi_1+i\psi_2)$ and $d_b=\frac{1}2(\psi_3+i\psi_4)$. The parameter relations are given by $\mu=2J_x,~t=J_y+\lambda,~\Delta=\Delta_1+i\Delta_2
=J_y-\lambda+i\kappa.$ An easy way to identify (\[1\]) and (\[2\]) is substituting (\[3\]) into (\[2\]). One can check $d_a$ and $d_b$ are conventional fermions, i.e., $\{d_a,d^\dag_a\}=\{d_b,d^\dag_b\}=1$ and $\{d_a,d_a\}=\{d_b,d_b\}=\{d^\dag_a,d_b\}=\{d_a,d^\dag_b\}=0$ [@cn; @yuwang]. This is a BCS $p$-wave pairing Hamiltonian in a finite system and can be diagonalized by rewriting (\[2\]) as $$H=\frac{1}2(d_a^\dag,d_b^\dag,d_a,d_b)\left(\begin{array}{cccc}
-\mu&-t&0&-\Delta\\
-t&-\mu&\Delta&0\\
0&\Delta^*&\mu&t\\
-\Delta^*&0&t&\mu\\\end{array}\right)\left(\begin{array}{c}
d_a\\
d_b\\
d^\dag_a\\
d^\dag_b\\\end{array}\right)$$ The eigen vaules of the Hamiltonian matrix are $$\begin{aligned}
&&E_0=-\frac{1}2(\sqrt{|\Delta|^2+\mu^2}+t),
E_t=-\frac{1}2(\sqrt{|\Delta|^2+\mu^2}-t),\nonumber\\
&&E_\mu=\frac{1}2(\sqrt{|\Delta|^2+\mu^2}-t),
E_2=\frac{1}2(\sqrt{|\Delta|^2+\mu^2}+t),\end{aligned}$$ Diagonalizing the Hamiltonian, one has $$\begin{aligned}
H&=&E_2\tilde d^\dag_b\tilde d_b+E_\mu\tilde d^\dag_a\tilde
d_a+E_0\tilde d_b\tilde d^\dag_b+E_t\tilde d_a\tilde d^\dag_a.\end{aligned}$$ The generalized Bogoliubov fermions $\tilde d_{a,b}$ can be obtained in a standard way with $\tilde
d_{a,b}=x^{(a,b)}_1d_{a,b}+x^{(a,b)}_2d_{b,a}+x^{(a,b)}_3d^\dag_{a,b}+x^{(a,b)}_4d^\dag_{b,a}
$ and the coefficients $x^{a,b}_i$ are normalized eigen vector of the Hamiltonian matrix. The Bogoliubov fermion operators obey the standard fermion commutation relations. A subspace of quantum states are $\{|G\rangle,\tilde d_a|G\rangle,\tilde
d^\dag_a|G\rangle, \tilde d^\dag_b\tilde d^\dag_a|G\rangle\}$ where $|G\rangle=\tilde d_a\tilde d_b|0\rangle$ with $
d_a|0\rangle= d_b|0\rangle=0$. The vacuum $|0\rangle=d_ad_b|\phi\rangle$ for a reference state $|\phi\rangle$, e.g., $|$$\uparrow\uparrow\uparrow\uparrow
\rangle$. The eigen energies of this set of quantum states are $\{-\sqrt{|\Delta|^2+\mu^2},-t,t,\sqrt{|\Delta|^2+\mu^2}\}$. When $\sqrt{|\Delta|^2+\mu^2}>t$, i.e., $4J_x^2+\kappa^2-4J_y\lambda>0,$ $|G\rangle$ is the ground state. Because $[H,b_i]=0$, each energy level is formally sixteen-fold degenerate, e.g., the ground states are $ \{|G\rangle,
b_i|G\rangle,b_ib_j|G\rangle,b_ib_jb_k|G\rangle\}. $ That is , $b_i$ play a role of zero modes of Majorana fermions. However, there are only four independent, which, e.g., are $$\begin{aligned}
\{|G\rangle,c_1^\dag|G\rangle,c_2^\dag|G\rangle, c_1^\dag
c_2^\dag|G\rangle\}\label{ground}\end{aligned}$$ where $c_1^\dag=\frac{1}2(b_1-ib_3)$ and $c^\dag_2=\frac{1}2(b_2-ib_4)$ with $
c_1|G\rangle=c_2|G\rangle=0$. The total Hilbert space is sixteen-dimensional as expected. Four degenerate states in a given energy level are distinguished by quantum number $P=(P_A,P_B,P_C,P_D)$, which are shown in Tab. 1.
--------------------------------------------------------------------------------------
$1$ $c^\dag_1$ $c^\dag_2$ $c^\dag_1c^\dag_2$
------------------------ ------------- ------------ ------------- --------------------
$|G\rangle$ (1,1,1,1) (-1,1,1-1) (1,-1,-1,1) (-1,-1,-1,-1)
$\tilde (-1,-1,1,1) (1,-1,1-1) (-1,1-1,1) (1,1,-1,-1)
d^\dag_a|G\rangle$
$\tilde (-1,-1,1,1) (1,-1,1-1) (-1,1-1,1) (1,1,-1,-1)
d^\dag_b|G\rangle$
$\tilde d^\dag_a\tilde (1,1,1,1) (-1,1,1-1) (1,-1,-1,1) (-1,-1,-1,-1)
d^\dag_b||G\rangle$
--------------------------------------------------------------------------------------
[Table 1: The eigen values of $P$ of the quantum states.]{}
Due to the constraint $P_AP_BP_CP_D=1$, there are eight different $P$ which are carried by the states in the first two levels or in the last two levels as shwon in Tab. 1.
[*Fusion rules: abelian and non-abelian* ]{} We now go to illustrate the fusion rules of these eigen excitations. First, we define Majorana fermions corresponding to $\tilde d_{a,b}$: $$\begin{aligned}
&&\tilde \psi_1=\tilde d_a+\tilde d^\dag_a,~\tilde
\psi_2=-i(\tilde d_a-\tilde d^\dag_a),\nonumber\\&&
\tilde
\psi_3=\tilde d_b+\tilde d^\dag_b,~\tilde \psi_4=-i(\tilde
d_b-\tilde d^\dag_b).\end{aligned}$$ They obey $\{\tilde\psi_i,\tilde\psi_j\}=2\delta_{ij}$ and $\{\tilde\psi_i,b_j\}=0$. There are four sets of states which obey abelian fusion rules, as those in Kitaev toric code model, $$\begin{aligned}
&&\sigma^{(1)}_i\sigma^{(2)}_i\sim \tilde \psi_i,~
\sigma^{(1)}_i\tilde \psi_i\sim \sigma^{(2)}_i,~
\sigma^{(2)}_i\tilde \psi_i\sim \sigma^{(1)}_i,\nonumber
\\&&
\sigma^{(1)}_i\sigma^{(3)}_i\sim b_i,~~\sigma^{(1)}_ib_i\sim
\sigma^{(3)}_i,~~ \sigma^{(3)}_ib_i\sim \sigma^{(1)}_i,\end{aligned}$$ for $i=1,2,3,4$. These operators are $$\begin{aligned}
&& \sigma^{(1)}_1=ib_1\tilde \psi_1,~\sigma^{(2)}_1=ib_1,
~ \sigma^{(3)}_1=-i\tilde \psi_1;\nonumber\\
&&\sigma^{(1)}_2=-ib_2\tilde \psi_2,~\sigma^{(2)}_2=-ib_2,~
\sigma^{(3)}_2=i\tilde \psi_2;
\\
&& \sigma^{(1)}_3=-b_1\tilde \psi_1b_3\tilde \psi_3,~
\sigma^{(2)}_3=-b_3b_1\tilde \psi_1,~\sigma^{(3)}_3
=\tilde \psi_3b_1\tilde \psi_1;\nonumber\\
&&\sigma^{(1)}_4=-b_2\tilde \psi_2b_4\tilde \psi_4,~
\sigma^{(2)}_4=-b_4b_2\tilde \psi_2, ~\sigma^{(3)}_4 =-\tilde
\psi_4b_2\tilde \psi_2.\nonumber\end{aligned}$$ Acting on the ground state $|G\rangle$, they are eigen states of $P$ and their eigen values can be read out from Tab. 1. They are also eigen states of the Hamiltonian and their eigen energies can be read out from the number of $\tilde\psi_i$ in a given operator. Since each energy level is four-fold degenerate, we find that the linear combination of these degenerate states may obey non-abelian fusion rules. They can be thought as non-abelian vortices. For example, $$\begin{aligned}
&&\eta_A=\frac{i}{\sqrt 2} (b_2\tilde \psi_2-b_4b_2\tilde
\psi_2),~\eta_B=\frac{i}{\sqrt
2}(b_1\tilde\psi_1-b_3b_1\tilde\psi_1),\nonumber\\
&&\eta_C=\frac{i}{\sqrt
2}(b_4\tilde\psi_4-b_1b_4\tilde\psi_4),~\eta_D=\frac{i}{\sqrt
2}(b_3\tilde\psi_3-b_2b_3\tilde\psi_3).\end{aligned}$$ They are in fact the superposition of those toric code abelian vortcies. (We will be back to this point when we study braiding of anyons.) Acting these operators on $|G\rangle$, they are also eigen states of $H$. The details of $P$ and $H$’s eigen values list on Tab. 2. The subscript indices are used because, e.g., $\eta_A|G\rangle$ has an eigen value $P_A=-1$ and other three either are 1 or do not have a definite eigen value. Therefore, $\eta_A$ can be thought as a vortex excitation on the surface $A$, and so on.
$~~P_A~~$ $~~P_B~~$ $~~P_C~~$ $~~P_D~~$ $~~H~~$
-------------- ----------- ----------- ----------- ----------- ---------
$~~\eta_A~~$ -1 \* \* 1 $-t$
$\eta_B$ \* -1 1 \* $-t$
$\eta_C$ \* 1 -1 \* $t$
$\eta_D$ 1 \* \* -1 $t$
[Table 2: The eigen values of $P$ and $H$ of $\eta$. ‘$*$’ means the vortex does not have a definite eigen value. ]{}
Using the algebra of Pauli matrices or equivalently, the anti-commutation relations between the Majorana fermions, one may directly prove that these operators obey the following non-abelian fusion rules $$\begin{aligned}
&&\eta_A\eta_A\sim 1+b_4, ~~\eta_A b_4\sim \eta_A,~~
b_4b_4=1,\nonumber\\&&\eta_B\eta_B\sim 1+b_3, ~~\eta_B b_3\sim
\eta_B,~~
b_3b_3=1,\label{naf}\\
&&\eta_C\eta_C\sim 1+b_1, ~~\eta_C b_1\sim \eta_C,~~
b_1b_1=1,\nonumber\\ &&\eta_D\eta_D\sim 1+b_2, ~~\eta_D b_2\sim
\eta_D,~~ b_2b_2=1,\nonumber\end{aligned}$$ which are standard non-abelian Ising fusion rules. Equations (\[naf\]) are one of central results in this paper. Only when the zero modes of Majorana fermions exist, the non-abelian vortices are eigen excitations [@iv]. We hope this illustration can help condensed matter physicists have a direct impression to these elusive mathematical relations and understand them in an elementary way.
[*Breaking of time reversal symmetry* ]{} The Ising non-abelian fusion rules and time reversal symmetry are concomitant. The three-spin coupling terms in Hamiltonian (\[1\]) explicitly break the time reversal symmetry. However, our deduction of the non-abelian fusion rules does not rely on a non-zero $\kappa$. They hold even $\kappa=0$. The only change is $\Delta_2=\kappa=0$ and the gap $\Delta=\Delta_1$. Actually, when $\kappa=0$, the time reversal symmetry is spontaneously broken. The ground states are four-fold degenerate, which are given by (\[ground\]). Since under the time reversion $T$, $T\sigma^a
T^{-1}=-\sigma^a$, one has $T|G\rangle=|G\rangle,~
Tc^\dag_2|G\rangle=c^\dag_2|G\rangle,
~Tc_1^\dag|G\rangle=-c_1^\dag|G\rangle,~Tc_1^\dag
c^\dag_2|G\rangle=-c_1^\dag c^\dag_2|G\rangle.$ We see two sectors which have different eigen values of $T$, i.e., the time reversal symmetry is spontaneously broken. This is because of the geometric frustration of the tetrahedron. This spontaneous breaking of the time reversal symmetry first observed in Kitaev model on a triangle-honeycomb lattice and leads to a chiral spin liquid [@yao].
[*Non-abelian anyons* ]{} A frequently quoted result is that the excitations obey non-abelian fusion rules like (\[naf\]) is equivalent to the braidings of these vortices $\eta$ are non-abelian and then these vortices are called anyons with non-abelian statistics or non-abelian anyons [@rev]. However, it is based on these anyons are well-defined and they are energetically degenerate. In this small system, to identify an $\eta$ vortex as an anyon is reluctant because we see that, e.g., $\eta_A$ is not an eigen state of $P_B$ and $P_C$, which means this is not a particle-like isolated excitation. On the other hand, the vortices $\eta_{A,B}$ and $\eta_{C,D}$ are not in the same energy level if $t\ne 0$ and braiding two vortices with different energy do not make sense in statistics. Therefore, for this small system, we only emphasize the non-abelian fusion rules of these vortex excitations but not call them non-abelian anyons.
![\[fig:Fig.2\] left panel: A chain of tetrahedrons on a torus. Right panel: unwind the torus to a periodic lattice. The thick lines carry $J_x$ while the dash lines carry $J_y$.](fig2a.EPS "fig:"){width="2.3cm"} ![\[fig:Fig.2\] left panel: A chain of tetrahedrons on a torus. Right panel: unwind the torus to a periodic lattice. The thick lines carry $J_x$ while the dash lines carry $J_y$.](fig2b.EPS "fig:"){width="6.5cm"}
To well define an anyon, we generalize the single tetrahedron model to a chain model of tetrahedrons on a torus (Fig. \[fig:Fig.2\]). The model Hamiltonian is given by $$\begin{aligned}
&&H_{\rm chain}=
J_x\sum_{i=1}^{n}\sigma^x_{i_1}\sigma^x_{i_2}+J_y\sum_{i=1}^{n}
\sigma^y_{i_2}\sigma^y_{i+1_1}\nonumber\\
&&\label{gh}
+\lambda\sum_{i=1}^{n}\sigma^x_{i_1}\sigma^z_{i_2}\sigma^z_{i+1_1}
\sigma^x_{i+1_2}\\
&&
+\kappa\sum_{i=1}^{n}(\sigma^x_{i_1}\sigma^z_{i_2}\sigma^y_{i+1_1}
+\sigma^y_{i_2}\sigma^z_{i+1_1}\sigma^x_{i+1_2})\nonumber\end{aligned}$$ with $n+1\equiv 1$. Only a half of triangular plaquette operators are conserved, which are $
P=\{P_{2a-1}=P_{2a-1_12a-1_22a_1}=\sigma^y_{2a-1_1}\sigma^z_{2a-1_2}\sigma^x_{2a_1},
P_{2a}=P_{2a-1_12a_12a_2}=\sigma^y_{2a-1_2}\sigma^z_{2a_1}\sigma^x_{2a_2}\}.$ The Hamiltonian can be diagonalized as $$\begin{aligned}
&&H_{\rm chain}=-t\sum_{i=1}^{n}(d^\dag_i d_{i+1}+d^\dag_{i+1}
d_i)-\mu\sum_{i=1}^{n}d^\dag_id_i\nonumber\\&&+\sum_{i=1}^{n}(\Delta^*
d_id_{i+1}+\Delta d_{i+1}^\dag d^\dag_i),\end{aligned}$$ where $ d_i=\frac{1}2(\psi_{i_1}-i\psi_{i_2}) $ and the Majorna fermions are given by [@jw][@cn] $\psi_{i_1}=\sigma^x_{i_1}\prod_{j<i}\sigma^z_{j_1}\sigma^z_{j_1},
\psi_{i_2}=\sigma^y_{i_2}\sigma^z_{i_1}\prod_{j<i}\sigma^z_{j_1}\sigma^z_{j_1},~b_{i_1}=-\sigma^y_{i_1}\prod_{j<i}\sigma^z_{j_1}\sigma^z_{j_1},
b_{i_2}=-\sigma^x_{i_2}\sigma^z_{i_1}\prod_{j<i}\sigma^z_{j_1}\sigma^z_{j_1},
$ where Majorana fermions $b_i$ commute with $H_{\rm chain}$. Therefore, $b_ib_j\cdots|\rangle$ are degenerate states if $|\rangle$ is an eigen state. Since $[P,\psi_i]=0$ for all $i$ but not all $b_i$, the eigen values of $P$ for the vortex states $\eta_i=\frac{i}{\sqrt 2}(b_i\tilde\psi_p-b_jb_i\tilde\psi_p)$, $j\ne i$ are determined by $b_{i,j}$. Here $\tilde\psi_p$ are the linear combination of $\psi_l$ after diagonalizing $H_{\rm chain}$ as those in the single tetrahedron model. If $j$ is far from $i$, $\eta_i$ has only one minus $P$ near the $i$-th site. Then, $\eta_i$ is a well-defined single vortex operator and can be thought as a non-abelian anyon due to $\eta_i\eta_i=1+b_j$. There are $n$ such anyons which are degenerate. Each anyon brings a zero mode $b_i$ of Majorana fermion in its center [@iv][@kitaev].
[*Non-abelian braiding matrices*]{} Rewriting $\eta_i=\frac{1}{\sqrt{2}}(e_i-m_i)$ with $e_i=ib_i\tilde\psi_p$ and $m_i=ib_jb_i\tilde\psi_p$, one has abelian fusuion rules $em\sim b$, $mb\sim e$, $eb\sim m$ and $e^2=m^2=b^2=1$. $e$ and $m$ are toric code mutual abelian anyons with degenerate energy. Thus, the Ising anyon is in fact the superposition of the toric code abelian anyons. This result has been recognized in refs. [@yusi; @wang] but without involving the Hamiltonian. Therefore, the toric code braiding matrices determine the non-abelian braiding matrices of $\eta_i$ [@wang], which are Ising-like braiding matrices $$R^{bb}_1=-1,~(R^{\eta\eta}_1)^2=1,~ (R^{b\eta}_\eta)^2=-1,
~(R^{\eta\eta}_b)^2=-1.$$Missing of the complex phases $e^{-i\pi/8}$ in $R^{\eta \eta}_1$ and $e^{i\pi/8}$ in $R^{\eta\eta}_b$ is because $R=R^\dag$ for the toric code anyons [@wang]. A framing therefore is needed [@lw]. We do not intend to propose a framing and ancillary qubits to implement the non-trivial chirality but refer to Wootton et al [@wang].
In conclusions, we presented a simple model in which there are a set of vortices obeying non-abelian fusion rules which were explicitly illustrated in an elementary way without using deep mathematical tools. Finally, we generalized the single tetrahedron model to a chain model of tetrahedrons on a torus and showed that the non-abelian vortices defined in the single tetrahedron model become well-defined non-abelian anyons.
This work was supported in part by the national natural science foundation of China, the national program for basic research of MOST of China and a fund from CAS.
A. Kitaev, Ann. Phys. [**303**]{}, 2(2003). A. Kitaev, Ann. Phys. [**321**]{}, 2(2006). C. Nayak, S. H. Simon, A. Stern, M. Freedman and S. Das Sarma, Rev. Mod. Phys. [**80**]{}, 1083 (2008). M. A. Levin and X.-G. Wen, Phys. Rev. B [**71**]{}, 045110 (2005) Y.-J. Han, R. Raussendorf and L.-M. Duan, Phys. Rev. Lett. [**98**]{}, 150404 (2007). C. -Y. Lu, W. -B. Gao, O. Gühne, X. -Q. Zhou, Z. -B. Chen, and J. -W. Pan, arXiv:0710.0278. J. K. Pachos, W. Wieczorek, C. Schmid, N. Kiesel, R. Pohlner, and H. Weinfurter, arXiv:0710.0895. L. Jiang, G. K. Brennen, A. V. Gorshkov, K. Hammerer, M. Hafezi, E. Demler, M. D. Lukin, and P. Zoller, Nature Physics [**4**]{}, 482 (2008). J. -F. Du,. J. Zhu, M. -G. Hu, and J. -L. Chen, arXiv:0712.2694. M. Aguado, G. K. Brennen, F. Verstraete, and J. I. Cirac, arXiv:0802.3163. B. Paredes and I. Bloch, arXiv:0711.3796. S. Dusuel, K. P. Schmidt, J. Vidal, Phys. Rev. Lett. [**100**]{}, 177204 (2008).
Yue Yu and Ziqiang Wang, arXiv: 0708.0631, to appear in Euro. Phys. Lett.
H. Yao, S. C. Zhang, and S. A. Kivelson, arXiv: 0810.5347.
C. J. Wu, D. Arovas, and H. -H. Hung, arXiv: 0811.1380.
V. Lahtinen, G. Kells, A. Carollo, T. Stitt, J. Vala and J. K. Pachos, Ann. of Phys. [**323**]{}, 2286 (2008). H. D. Chen and Z. Nussinov, J. Phys. A [**41**]{}, 075001 (2008).
Yue Yu and Tieyan Si, arXiv:0804.0483.
J. R. Wootton, V. Lahtinen , Z. Wang, J. K. Pachos, Phys. Rev. B [**78**]{}, 161102(R) (2008).
D. A. Ivanov, Phys. Rev. Lett. [**86**]{}, 268 (2001).
H. Yao and S. A. Kivelson, Phys. Rev. Lett. [**99**]{}, 247203 (2007).
X. Y. Feng, G. M. Zhang, and T. Xiang, Phys. Rev. Lett. [**98**]{}, 087204 (2007). H. D. Chen and J. P. Hu, Phys. Rev. B [**76**]{}, 193101 (2007).
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'We study the multiplicative Hilbert matrix, i.e. the infinite matrix with entries $(\sqrt{mn}\log(mn))^{-1}$ for $m,n\geq2$. This matrix was recently introduced within the context of the theory of Dirichlet series, and it was shown that the multiplicative Hilbert matrix has no eigenvalues and that its continuous spectrum coincides with $[0,\pi]$. Here we prove that the multiplicative Hilbert matrix has no singular continuous spectrum and that its absolutely continuous spectrum has multiplicity one. Our argument relies on the tools of spectral perturbation theory and scattering theory. Finding an explicit diagonalisation of the multiplicative Hilbert matrix remains an interesting open problem.'
address:
- 'Department of Mathematics and Statistics, University of Reading, Reading RG6 6AX, United Kingdom'
- 'Department of Mathematics, King’s College London, Strand, London WC2R 2LS, United Kingdom'
author:
- 'Karl-Mikael Perfekt'
- Alexander Pushnitski
title: On the spectrum of the multiplicative Hilbert matrix
---
Introduction {#sec.a}
============
Let $\{h(n)\}_{n=0}^\infty$ be a sequence of complex numbers. A *Hankel matrix* is an infinite matrix of the form $$\{h(n+m)\}_{n,m=0}^\infty,$$ considered as a linear operator on $\ell^2({{\mathbb Z}}_+)$, ${{\mathbb Z}}_+=\{0,1,2,\dots\}$. One of the key examples of Hankel matrices is the *Hilbert matrix*: $$\{(n+m+a)^{-1}\}_{n,m=0}^\infty,\quad a>0.
\label{a1}$$ Magnus [@Magnus] studied the spectrum of the Hilbert matrix for $a=1$, proving that it is given by the interval $[0,\pi]$ and that it is purely continuous, i.e. there are no eigenvalues. Later, Rosenblum [@Rosenblum] proved the following theorem.
For any $a>0$, the absolutely continuous (a.c.) spectrum of the Hilbert matrix is $[0,\pi]$ with multiplicity one. There is no singular continuous spectrum.
1. If $a\geq1/2$, there are no eigenvalues.
2. If $0<a<1/2$, there is one simple eigenvalue $\pi/\cos(\pi(a-1/2))$, and no other eigenvalues.
In fact, Rosenblum gave an explicit diagonalisation of the Hilbert matrix for all $a\in{{\mathbb R}}$, $a\not=0,-1,-2,\dots$, from which one can read off the above description of the spectrum.
The Hilbert matrix is remarkable, in particular, for being the simplest bounded non-compact Hankel matrix. It exhibits the following borderline behaviour:
- If $nh(n)\to0$ as $n\to\infty$, then $\{h(n+m)\}_{n,m=0}^\infty$ is compact on $\ell^2({{\mathbb Z}}_+)$;
- If $nh(n)\to\infty$ as $n\to\infty$, then $\{h(n+m)\}_{n,m=0}^\infty$ is unbounded on $\ell^2({{\mathbb Z}}_+)$.
In this paper we discuss a similarly remarkable borderline operator in the class of *multiplicative Hankel matrices,* i.e. in the class of infinite matrices of the form $$M(g)=\{g(nm)\}_{n,m=1}^\infty,$$ considered as linear operators on $\ell^2({{\mathbb N}})$, ${{\mathbb N}}=\{1,2,\dots\}$. Here the $(n,m)$’th entry of the matrix depends on the product $nm$ instead of the sum $n+m$. Following [@QQ], we call such operators *Helson matrices,* in honour of H. Helson’s pioneering work [@Helson] on the subject.
The special Helson matrix we consider corresponds to the sequence $$g_a(n)=\frac1{\sqrt{n}(a+\log n)}, \quad n\geq1, \quad a>0.
\label{a0}$$ It is not difficult to see that $M(g_a)$ is a bounded self-adjoint operator on $\ell^2({{\mathbb N}})$ (see Theorem \[thm.a4\]). Similarly to the case of the Hilbert matrix, the Helson matrix $M(g_a)$ is borderline in the following sense:
- If $(\sqrt{n}\log n) g(n)\to0$ as $n\to\infty$, then $M(g)$ is compact on $\ell^2({{\mathbb N}})$;
- If $(\sqrt{n}\log n) g(n)\to\infty$ as $n\to\infty$, then $M(g)$ is unbounded on $\ell^2({{\mathbb N}})$.
Even though it is not possible to take $a=0$ in the definition , one can do so if the indices $n,m$ are restricted to ${{\mathbb N}}_2=\{2,3,\dots\}$. We will denote by $M_2(g)$ the operator on $\ell^2({{\mathbb N}}_2)$ corresponding to the matrix $\{g(nm)\}_{n,m\geq2}$. Hence, when $a=0$, one can consider $M_2(g_0)$, with $g_0$ defined as in , i.e. $$M_2(g_0)=\{(\sqrt{mn}\log(mn))^{-1}\}_{n,m=2}^\infty.$$ Following [@BPSSV], we call $M_2(g_0)$ the *multiplicative Hilbert matrix*.
We now state our main results.
\[thm.a1\] The multiplicative Hankel matrix $M_2(g_0)$ has a purely a.c. spectrum (no singular continuous spectrum, no eigenvalues) which coincides with $[0,\pi]$ and has multiplicity one.
\[thm.a2\] For any $a>0$, the a.c. spectrum of $M(g_a)$ coincides with $[0,\pi]$ and has multiplicity one. The singular continuous spectrum of $M(g_a)$ is absent. There is a critical value $a_*>0$ such that:
1. If $a\geq a_*$, then $M(g_a)$ has no eigenvalues.
2. If $0< a<a_*$, then $M(g_a)$ has one simple eigenvalue $\lambda(a)>\pi$, and no other eigenvalues. The eigenvalue $\lambda(a)$ is a continuous non-increasing function of $a\in(0,a_*)$, with $\lim_{a \to a_*^{-}} \lambda(a) = \pi$ and $\lim_{a \to 0^{+}} \lambda(a) = \infty$.
Despite the similarity with Theorem A, we do not have an explicit diagonalisation of either $M(g_a)$ or $M_2(g_0)$; to find one is an interesting open problem. We are also unable to explicitly compute the critical value $a_*$, and the same is true for the eigenvalue $\lambda(a)$. We only have the crude estimates $$\frac1\pi\leq a_*\leq 2,\quad \frac1a\leq \lambda(a)\leq \pi+\frac1a,$$ see Section \[sec.aaa\].
The multiplicative Hilbert matrix $M_2(g_0)$ was introduced in [@BPSSV], where it was placed in the context of the study of operators acting on Hardy spaces of Dirichlet series. The multiplicative Hilbert matrix is especially interesting, because although its analogy to the classical Hilbert matrix is unambiguous, questions about its place within the theory of Dirichlet series are open. In particular, it is not known whether $M_2(g_0)$ has a bounded symbol — see Problem 3.2 in [@SaksmanSeip] for a precise statement.
The fact that the spectrum of $M_2(g_0)$ is purely continuous and coincides with $[0,\pi]$ was already proven in [@BPSSV], but we give a more streamlined version of the proof.
Our main contributions in this paper are:
- the proof that the singular continuous spectra of $M_2(g_0)$ and $M(g_a)$ are absent and that the respective a.c. spectra have multiplicity one;
- a clarification of the connection between $M_2(g_0)$, $M(g_a)$ and certain integral Hankel operators. This lies at the heart of our proof — see Section \[sec.aa\].
We also attempt to push the analogy with the Hilbert matrix a little further by introducing the one-parameter family $M(g_a)$, with the hope of stimulating some further progress.
The strategy of the proof {#sec.aa}
=========================
Reduction to integral Hankel operators
--------------------------------------
Key to our analysis is a unitary equivalence between $M(g_a)$, $M_2(g_0)$ and certain integral Hankel operators. We start by recalling the definition of this class of operators. For a *kernel function* $h\in L^1_{{\mathrm{loc}}}({{\mathbb R}}_+)$, let us denote by $H(h)$ the *integral Hankel operator* in $L^2({{\mathbb R}}_+)$, formally defined by $$(H(h)f)(t)=\int_0^\infty h(t+s)f(s)ds, \quad t>0, \quad f\in L^2({{\mathbb R}}_+).$$ We need a simple sufficient condition for the boundedness of integral Hankel operators. Let $H(1/t)$ be the Carleman operator, i.e. the integral Hankel operator with the kernel function $h(t)=1/t$. Recall that $H(1/t)$ is bounded on $L^2({{\mathbb R}}_+)$ with norm $\pi$; we will come back to the spectral properties of $H(1/t)$ in Section \[sec.b3\]. From the boundedness of $H(1/t)$ we immediately get the following estimate, which is both well known and easy to prove.
\[lma.d1\] Let $h$ be a kernel function with ${\lverth(t)\rvert}\leq A/t$ for some $A>0$. Then ${\lVertH(h)\rVert}\leq A\pi$.
For $f_1,f_2\in L^2({{\mathbb R}}_+)$, we have $${\lvert(H(h)f_1,f_2)\rvert}\leq A(H(1/t){\lvertf_1\rvert},{\lvertf_2\rvert})\leq A\pi{\lVert{\lvertf_1\rvert}\rVert}{\lVert{\lvertf_2\rvert}\rVert}=A\pi{\lVertf_1\rVert}{\lVertf_2\rVert}.
\qedhere$$
Let $$\zeta(t)=\sum_{n=1}^\infty n^{-t}, \quad t>1,$$ be the Riemann zeta function. As usual, for a bounded operator $H$ in a Hilbert space, we say that $H$ is non-negative, $H\geq0$, if $(Hx,x)\geq0$ for all elements $x$ in the Hilbert space.
Now we are ready to state the unitary equivalence between $M_2(g_0)$, $M(g_a)$, and integral Hankel operators.
\[thm.a4\]
1. The operator $M_2(g_0)$ is bounded, non-negative, has trivial kernel, and is unitarily equivalent to $$H(h_0)|_{(\operatorname{Ker}H(h_0))^\perp}, \quad h_0(t)=\zeta(t+1)-1.$$
2. For any $a>0$, the operator $M(g_a)$ is bounded, non-negative, has trivial kernel, and is unitarily equivalent to $$H(h_a)|_{(\operatorname{Ker}H(h_a))^\perp}, \quad h_a(t)=\zeta(t+1)e^{-at/2}.$$
In fact, it is not difficult to see that the kernels of $H(h_0)$ and $H(h_a)$ are also trivial. However, we will not need this observation in our proof.
\(i) Consider the operator ${\mathcal{N}}_0: L^2({{\mathbb R}}_+)\to \ell^2({{\mathbb N}}_2)$, given by $$({\mathcal{N}}_0 f)_n=\int_0^\infty n^{-\frac12-t}f(t)dt,
\quad n\in{{\mathbb N}}_2,\quad f\in L^2({{\mathbb R}}_+).$$ For $f_1,f_2\in L^2({{\mathbb R}}_+)$, we have that $$\begin{aligned}
({\mathcal{N}}_0f_1,{\mathcal{N}}_0f_2)
&=
\sum_{n=2}^\infty \int_0^\infty \int_0^\infty n^{-1-t-s}f_1(s)\overline{f_2(t)}\, ds\, dt \\
&=
\int_0^\infty \int_0^\infty (\zeta(t+s+1)-1)f_1(s)\overline{f_2(t)} \, ds\, dt.\end{aligned}$$ Thus $$({\mathcal{N}}_0 f_1,{\mathcal{N}}_0f_2)=(H(h_0)f_1,f_2), \quad h_0(t)=\zeta(t+1)-1.$$ We have the elementary bound $$h_0(t)=\sum_{n=2}^\infty n^{-t-1}\leq\int_1^\infty\frac{dx}{x^{t+1}}=\frac1t, \quad t>0.
\label{d1}$$ Together with Lemma \[lma.d1\], this bound shows that $H(h_0)$ is bounded on $L^2({{\mathbb R}}_+)$. Thus, ${\mathcal{N}}_0$ is also bounded and $$H(h_0)={\mathcal{N}}_0^*{\mathcal{N}}_0.$$ Next, consider the adjoint ${\mathcal{N}}_0^*: \ell^2({{\mathbb N}}_2)\to L^2({{\mathbb R}}_+)$, given by $$({\mathcal{N}}_0^* x)(t)=\sum_{m\geq2} x_m m^{-\frac12-t},
\quad t>0, \quad x=\{x_m\}_{m=2}^\infty\in\ell^2({{\mathbb N}}_2).$$ Since $$g_0(n)=(\sqrt{n}\log n)^{-1}=\int_0^\infty n^{-\frac12-t}dt,\quad n\geq2,$$ we see that the operator ${\mathcal{N}}_0{\mathcal{N}}_0^*$ on $\ell^2({{\mathbb N}})$ is given by $$({\mathcal{N}}_0{\mathcal{N}}_0^* x)_n=\sum_{m\geq2} g_0(nm)x_m,
\quad x\in \ell^2({{\mathbb N}}),$$ i.e. ${\mathcal{N}}_0{\mathcal{N}}_0^*=M_2(g_0)$.
It is well known that, for any bounded operator ${\mathcal{N}}$, the operators $${\mathcal{N}}^*{\mathcal{N}}|_{(\operatorname{Ker}{\mathcal{N}})^\perp}
\quad \text{ and }\quad
{\mathcal{N}}{\mathcal{N}}^*|_{(\operatorname{Ker}{\mathcal{N}}^*)^\perp}$$ are unitarily equivalent. This gives the required unitary equivalence between the non-zero parts of $M_2(g_0)$ and $H(h_0)$.
To complete the proof of part (i), it remains to show that the kernel of ${\mathcal{N}}_0^*$ is trivial. This is easy to check: if ${\mathcal{N}}_0^* x=0$, then, inspecting the asymptotics of $({\mathcal{N}}_0^* x)(t)$ as $t\to\infty$, we inductively prove that $x_m=0$ for $m=2,3,\dots$.
\(ii) For a fixed parameter $a>0$, let ${\mathcal{N}}_a: L^2({{\mathbb R}}_+)\to \ell^2({{\mathbb N}})$ be the linear operator given by $$({\mathcal{N}}_a f)_n=\int_0^\infty e^{-at/2}n^{-\frac12-t}f(t)dt,
\quad n\in{{\mathbb N}},\quad f\in L^2({{\mathbb R}}_+).$$ Similarly to part (i), we have $${\mathcal{N}}_a^*{\mathcal{N}}_a=H(h_a), \quad h_a(t)=\zeta(t+1)e^{-at/2}.$$ Using , we get $$h_a(t)=e^{-at/2}\sum_{n=1}^\infty n^{-t-1}\leq e^{-at/2}(1+1/t)\leq C(a)/t, \quad t > 0,$$ and so $H(h_a)$ is bounded.
On the other hand, the adjoint ${\mathcal{N}}_a^*: \ell^2({{\mathbb N}})\to L^2({{\mathbb R}}_+)$ is given by $$({\mathcal{N}}_a^* x)(t)=e^{-at/2}\sum_{m\geq1} x_m m^{-\frac12-t},
\quad t>0, \quad x=\{x_m\}_{m\geq1}\in\ell^2({{\mathbb N}}).$$ Since $$g_a(n)=\frac1{\sqrt{n}(a+\log n)}=\int_0^\infty e^{-at}n^{-\frac12-t}dt,
\quad
n\geq1,$$ we obtain $${\mathcal{N}}_a{\mathcal{N}}_a^*=M(g_a).$$ This gives the required unitary equivalence. Again it is easy to see that the kernel of ${\mathcal{N}}_a^*$ is trivial.
Some heuristics
---------------
Theorem \[thm.a4\] reduces the question to the analysis of integral Hankel operators with specific kernel functions. Let us recall some basic facts about such operators.
It is well known that the Carleman operator has a purely a.c. spectrum $[0,\pi]$ of multiplicity *two*. In fact, $H(1/t)$ is explicitly diagonalised by the Mellin transform, see Section \[sec.b3\] below. From Lemma \[lma.d1\] it is easy to conclude that $$h(t)=o(1/t) \text{ as } t\to0 \text{ and as } t\to\infty \quad \Rightarrow \quad H(h) \text{ is compact.}$$ Heuristically, the behaviour $1/t$ of the kernel function is *singular* both as $t\to0$ and as $t\to\infty$. The spectrum of $H(1/t)$ has multiplicity two because each of these singularities generates an interval of a.c. spectrum of multiplicity one. J. S. Howland has made this observation more precise in [@Howland] by proving, brushing some technical details aside, that if $$h(t)=\begin{cases}
c_0/t+\text{error term}, & t\to0,
\\
c_\infty/t+\text{error term}, & t\to\infty,
\end{cases}
\label{a7}$$ then the a.c. spectrum of $H(h)$ is given by the union of intervals $$\sigma_{{\text{\rm ac}}}(H(h))=[0,\pi c_0]\cup [0,\pi c_\infty],$$ where each of the two intervals contributes multiplicity one to the spectrum. (Howland was motivated, on the one hand, by the earlier work [@Power] of S. Power, which concerns the essential spectrum of Hankel operators with piecewise continuous symbols, and on the other hand, by similar results in scattering theory for Schrödinger operators.) Howland’s results were further extended in [@PYa], where a more general class of kernel functions was considered.
Now let us come back to the kernel functions $h_a$ and $h_0$ of Theorem \[thm.a4\]. Recall that zeta function $\zeta(z)$ has a simple pole at $z=1$ with residue $1$ and so $$\zeta(1+t)-1/t\in C^\infty([0,\infty)),
\label{a11}$$ and $$\zeta(t)=1+O(2^{-t}), \quad t\to\infty.
\label{a12}$$ It follows that $h_a$ satisfies $$h_a(t)=
\begin{cases}
1/t+O(1), & t\to0,
\\
o(1/t), & t\to\infty,
\end{cases}$$ both for $a=0$ and for $a>0$. Thus, $h_a$ satisfies with $c_0=1$ and $c_\infty=0$ and so, according to Howland’s paradigm, we should expect $H(h_a)$ to have the a.c. spectrum $[0,\pi]$ of multiplicity one. This is indeed what we will prove.
The absolutely continuous spectrum
----------------------------------
The statements about the absolutely continuous and the singular continuous spectra of $H(h_a)$ are consequences of the following theorem.
\[thm.a4a\] Let $h(t)$, $t>0$, be a real-valued kernel function such that $$\begin{split}
{\lverth(t)-1/t\rvert}&\leq Ct^{-1+{\varepsilon}}, \quad 0<t\leq1,
\\
{\lverth(t)\rvert}&\leq Ct^{-1-{\varepsilon}}, \quad t\geq1,
\end{split}
\label{a8}$$ with some ${\varepsilon}>0$ and $C>0$. Then the Hankel operator $H(h)$ is bounded and has the essential spectrum $[0,\pi]$. The absolutely continuous spectrum of $H(h)$ has multiplicity one and coincides with the same interval $[0,\pi]$. The singular continuous spectrum of $H(h)$ is absent.
The proof is given in Section \[sec.c\]. In fact, this theorem follows from [@Howland Theorem 2], established with Mourre’s inequality, or from [@PYa Theorem 7.10], where the smooth method of scattering theory is used. However, in both of these references the argument is much more complicated than necessary for our purposes. The kernels considered in [@Howland; @PYa] have two or more singularities, which leads to an a.c. spectrum of multiplicity two or more, necessitating the use of multi-channel scattering theory. Here we give a much simpler single-channel argument.
Conditions can be relaxed somewhat by replacing $t^{\pm{\varepsilon}}$ by suitable powers of ${\lvert\log t\rvert}$, see [@PYa].
The absence of embedded eigenvalues
-----------------------------------
The absence of eigenvalues of $H(h_a)$ in the interval $(0,\pi]$ will be established through the following theorem.
\[thm.a5\] Let $h$ be as in Theorem \[thm.a4a\]. Assume in addition that $H(h)\geq0$, that the function ${\widetilde}h(t)=h(t)-1/t$ satisfies ${\widetilde}h \in C^2([0,\infty))$, and that $$\biggl(\frac{d}{dt}\biggr)^k {\widetilde}h(t)=O(t^{-1-k}), \quad t\to\infty, \quad k=1,2.
\label{a9}$$ Then $H(h)$ has no eigenvalues in $(0,\pi]$.
The proof, which is given in Section \[sec.b\], is an extension of an argument from [@BPSSV]. We note that the argument in [@BPSSV] was somewhat obscured by the fact that the equivalence between $M_2(g_0)$ and the integral Hankel operator $H(h_0)$ was not fully understood. Because of this, the argument in [@BPSSV] is presented in terms of functions given by Dirichlet series, which makes it a little more complicated than necessary. Here we give a more streamlined (and more general) version.
The conditions on $h$ in Theorem \[thm.a5\] are certainly not optimal, but they are sufficient for proving our main results, since the kernels $h_0$ and $h_a$ satisfy them.
The structure of the paper
--------------------------
In Section \[sec.c\] we prove Theorem \[thm.a4a\]. In Section \[sec.b\] we prove Theorem \[thm.a5\]. In Section \[sec.aaa\] we use these two results, and some additional concrete analysis related to the eigenvalues in the interval $(\pi,\infty)$, to prove Theorems \[thm.a1\] and \[thm.a2\].
The absolutely continuous spectrum: proof of Theorem \[thm.a4a\] {#sec.c}
================================================================
Preliminaries
-------------
The idea of the proof is as follows. Consider the integral Hankel operator $H(h_*)$ with the kernel function $$h_*(t)=\frac{e^{-t/2}}{t}, \quad t>0.$$ This kernel function satisfies with $c_0=1$ and $c_\infty=0$, and so, according to Howland’s paradigm, we expect the a.c. spectrum of $H(h_*)$ to be $[0,\pi]$ with multiplicity one. This is indeed the case, and in fact, an explicit diagonalisation of $H(h_*)$ is available [@Lebedev1; @Lebedev2; @Rosenblum; @Yafaev1]. This operator has a purely a.c. spectrum $[0,\pi]$ with multiplicity one (no singular continuous spectrum, no eigenvalues), and the family of generalised eigenfunctions of the continuous spectrum is known in explicit form, as described in the next subsection.
Let $h$ be as in Theorem \[thm.a4a\]; as a warm-up, let us check that $\sigma_{\text{\rm ess}}(H(h))=[0,\pi]$. We have $$H(h)=H(h_*)+H(w),$$ where $w(t)=h(t)-h_*(t)$ satisfies $$w(t)=\begin{cases}
O(t^{-1+{\varepsilon}}),& t\to0,
\\
O(t^{-1-{\varepsilon}}),&t\to\infty.
\end{cases}$$ It follows that $\int_0^\infty t{\lvertw(t)\rvert}^2dt<\infty$, and therefore the integral kernel $w(t+s)$ of $H(w)$ is in $L^2({{\mathbb R}}_+\times{{\mathbb R}}_+)$. Thus, $H(w)$ is Hilbert-Schmidt, and therefore compact. By Weyl’s theorem on the invariance of the essential spectrum under compact perturbations, we obtain that $\sigma_{\text{\rm ess}}(H(h))=\sigma_{\text{\rm ess}}(H(h_*))=[0,\pi]$, as required.
Now let us outline the idea of proof of the rest of Theorem \[thm.a4a\]. Under our assumptions, we will prove that $H(h)$ has the representation $$H(h)=H(h_*)+G^*AG,
\label{c0}$$ where $G$ is a strongly $H(h_*)$-smooth operator, in the terminology of [@Yafaev2], and $A$ is a compact operator. Roughly speaking, this means that the difference $H(h)-H(h_*)$ can be represented as an operator with a sufficiently regular integral kernel *in the spectral representation of $H(h_*)$*. The proof that holds consists of two ingredients: a detailed analysis of the explicit diagonalisation of $H(h_*)$ together with an identification of a class of $H(h_*)$-smooth operators, and an (easy) proof of the compactness of the operator $A$.
By standard results of smooth scattering theory, the representation implies the existence and completeness of wave operators for the pair of operators $(H(h),H(h_*))$, yielding the statement about the a.c. spectrum of $H(h)$. The same considerations of scattering theory also yield the absence of the singular continuous spectrum of $H(h)$.
Diagonalization of $H(h_*)$
---------------------------
Let $K_\nu(z)$ be the modified Bessel function of the third kind; for ${\hbox{{\rm Re}}\,}\nu>-1/2$ and ${\hbox{{\rm Re}}\,}z>0$ it is given by the integral representation [@BE Section 7.3.5, formula (15)] $$\Gamma(\nu+\tfrac12)K_\nu(z)
=
\sqrt{\pi}\bigl(\tfrac{z}{2}\bigr)^\nu
\int_1^\infty e^{-zu}(u^2-1)^{-\frac12+\nu}du.
\label{c1}$$ For $k>0$ and $t>0$, set $$\psi_k(t)=\frac1\pi \sqrt{2k\sinh(\pi k)}t^{-\frac12}K_{ik}(t/2).$$ Formally, $\{\psi_k\}_{k> 0}$ gives a complete normalised set of generalised eigenfunctions of $H(h_*)$: $$H(h_*)\psi_k=\lambda(k)\psi_k, \quad\text{ where }\quad \lambda(k)=\frac{\pi}{\cosh(\pi k)}, \quad k>0.$$ More precisely, we have the following statement [@Lebedev1; @Lebedev2; @Rosenblum; @Yafaev1].
For $f\in C_{\text{\rm comp}}^\infty({{\mathbb R}}_+)$, let $$(Uf)(k)=\int_0^\infty f(t)\psi_k(t)dt, \quad k>0.$$ Then $U$ extends to a unitary operator on $L^2({{\mathbb R}}_+)$. For any $f\in L^2({{\mathbb R}}_+)$, it holds that $$(UH(h_*)U^* f)(k)=\lambda(k)f(k), \quad k>0.
\label{c2}$$
One reads off the spectrum of $H(h_*)$ from : it is given by the closure of the range of $\lambda(k)$, $k>0$, which coincides with the interval $[0,\pi]$. Additionally, since $\lambda(k)$ is a continuous and strictly decreasing function of $k \geq 0$, $\lim_{k \to \infty} \lambda(k) = 0$, the spectrum is purely a.c. and has multiplicity one. More explicitly, the unitary operator $U_0:L^2({{\mathbb R}}_+)\to L^2(0,\pi)$, $$(U_0f)(\lambda)=\bigl(\tfrac{d\lambda(k)}{dk}\bigr)^{-1/2}(Uf)(k), \quad \lambda=\lambda(k)\in(0,\pi),$$ reduces $H(h_*)$ to the operator of multiplication by the independent variable on $L^2(0,\pi)$, $$(U_0 H(h_*)U_0^*f)(\lambda)=\lambda f(\lambda),
\quad \lambda\in(0,\pi).$$
Strong smoothness
-----------------
Let us fix the following function $q\in L^\infty({{\mathbb R}}_+)$, $$q(t)=
\begin{cases}
{\lvert\log t\rvert}^{-1}, & t\in(0,1/2),
\\
(\log 2)^{-1}, & t\geq 1/2.
\end{cases}$$ Denote by $Q$ the operator of multiplication by $q(t)$ in $L^2({{\mathbb R}}_+)$. For $\beta>0$, we will also consider the power $Q^\beta$. In this subsection we will prove the following result.
\[thm.b2\] Let $\beta>1/2$. Then, in the terminology of [@Yafaev2 Section 4.4], $Q^\beta$ is strongly $H(h_*)$-smooth with any exponent $\gamma$ in the range $0<\gamma<\min\{1,\beta-1/2\}$ on any compact sub-interval of $(0,\pi)$.
The strong smoothness here means the following. Let $\delta\subset(0,\pi)$ be a compact interval and let $\gamma$ be as in the theorem. Then there exists a constant $C=C(\delta,\gamma)$ such that for any $f\in L^2({{\mathbb R}}_+)$ we have the estimates $$\begin{gathered}
{\lvert(U_0Q^\beta f)(\lambda)\rvert}\leq C{\lVertf\rVert}, \quad \lambda\in\delta,
\\
{\lvert(U_0Q^\beta f)(\lambda)-(U_0Q^\beta f)(\lambda')\rvert}\leq C{\lvert\lambda-\lambda'\rvert}^\gamma{\lVertf\rVert},
\quad \lambda,\lambda'\in\delta.\end{gathered}$$ In other words, the linear functional $$L^2({{\mathbb R}}_+)\ni f\mapsto (U_0Q^\beta f)(\lambda)$$ is norm-Hölder continuous in $\lambda\in\delta$ with the exponent $\gamma$.
Since $\lambda=\lambda(k)$ is a $C^\infty$-smooth function of $k>0$ and the derivative $\lambda'(k)$ does not vanish for $k>0$, it suffices to prove that, for any compact interval $\Delta\subset (0,\infty)$ and for any $\gamma<\min\{1,\beta-1/2\}$, the linear functional $$L^2({{\mathbb R}}_+)\ni f\mapsto (UQ^\beta f)(k)$$ is norm-Hölder continuous in $k\in\Delta$ with the exponent $\gamma$. Recalling the formula for $U$, we see that this functional is given by $$f\mapsto \int_0^\infty f(t) \psi_k(t) q(t)^\beta dt, \quad k>0.$$ Thus, we need to prove the estimates $$\begin{gathered}
\int_0^\infty {\lvert\psi_k(t)\rvert}^2 q(t)^{2\beta} dt\leq C, \quad k\in\Delta,
\\
\int_0^\infty {\lvert\psi_k(t)-\psi_{k'}(t)\rvert}^2 q(t)^{2\beta} dt\leq C{\lvertk-k'\rvert}^{2\gamma}, \quad k,k'\in\Delta.\end{gathered}$$ By the explicit form of $\psi_k$, it suffices to prove the estimates $$\begin{gathered}
\int_0^\infty {\lvertK_{ik}(t/2)\rvert}^2 t^{-1}q(t)^{2\beta} dt\leq C, \quad k\in\Delta,
\label{c10}
\\
\int_0^\infty {\lvertK_{ik}(t/2)-K_{ik'}(t/2)\rvert}^2 t^{-1} q(t)^{2\beta} dt\leq C{\lvertk-k'\rvert}^{2\gamma}, \quad k,k'\in\Delta.
\label{c11}\end{gathered}$$ Let us split the domain of integration in and into $(0,1/2)$ and $(1/2,\infty)$ and estimate the corresponding integrals separately.
Consider first the integrals over $(0,1/2)$. We recall the following representation for the modified Bessel function $K_\nu$: $$K_\nu(z)=\frac{\pi}{2\sin(\nu\pi)}(I_{-\nu}(z)-I_\nu(z)),$$ where $I_\nu(z)$ is the modified Bessel function of the first kind, given by the convergent series $$I_\nu(z)
=
\sum_{m=0}^\infty \frac{(z/2)^{2m+\nu}}{m!\Gamma(m+\nu+1)}.$$ For $k>0$, let us write $$I_{\pm ik}(z)=(z/2)^{\pm ik}{\widetilde}I_{\pm ik}(z), \quad
{\widetilde}I_{\pm ik}(z)=\sum_{m=0}^\infty \frac{(z/2)^{2m}}{m!\Gamma(m\pm ik+1)}.$$ By inspection, both ${\widetilde}I_{\pm ik}(z)$ and $(d/dk){\widetilde}I_{\pm ik}(z)$ are entire functions of $z$, bounded uniformly for $z\in(0,1)$ and $k\in \Delta$. Using the elementary estimate ${\lverte^{ia}-e^{ib}\rvert}\leq 2{\lverta-b\rvert}^\gamma$ for $0<\gamma<1$, we get $${\lvert(z/2)^{ik}-(z/2)^{ik'}\rvert}
=
{\lverte^{ik\log(z/2)}-e^{ik'\log(z/2)}\rvert}
\leq
2 {\lvertk-k'\rvert}^\gamma{\lvert\log (z/2)\rvert}^\gamma.$$ Using this, we obtain the estimates $$\begin{aligned}
{\lvertK_{ik}(t/2)\rvert}&\leq C,
\quad k\in \Delta, \quad t\in (0,1/2),
\\
{\lvertK_{ik}(t/2)-K_{ik'}(t/2)\rvert}&\leq
C{\lvertk-k'\rvert}^\gamma {\lvert\log t\rvert}^\gamma,
\quad k,k'\in \Delta, \quad t\in (0,1/2).\end{aligned}$$ Now it is clear that for $\beta>\gamma+1/2$ the estimates and hold with the integrals taken over $(0,1/2)$.
Next, let us consider the integrals over $(1/2,\infty)$. Here we use the integral representation for $K_\nu$. Let us rewrite it as follows, $$K_{ik}(t/2)
=\frac{\sqrt{\pi}}{\Gamma(ik+\tfrac12)}(t/4)^{ik}
\int_1^\infty e^{-tu/2}(u^2-1)^{-\frac12+ik}du.$$ For $t\geq1/2$, one can estimate the exponential in the integral for $K_{ik}$ as $$e^{-tu/2}=e^{-tu/4}e^{-tu/4}\leq e^{-t/4} e^{-u/8}.$$ This allows one to conclude that $K_{ik}$ satisfies $${\lvertK_{ik}(t/2)\rvert}+{\lvert(d/dk) K_{ik}(t/2)\rvert}\leq C {\lvert\log (t/4)\rvert}e^{-t/4}, \quad t>1/2, \quad k\in \Delta.$$ Thus, we obtain the estimates and with the integrals taken over $(1/2,\infty)$. Here we do not need any restrictions on $\gamma$ and $\beta$.
Putting the results together
----------------------------
We use the following statement from scattering theory. For the proof and further details, see [@Yafaev2 Section 4.7].
\[prp.c3\] Let $H_0$ be a bounded self-adjoint operator in a Hilbert space. Assume that the spectrum of $H_0$ is purely a.c., has constant multiplicity $m$, and coincides with the interval $[a,b]$. Let $G$ be a bounded operator, which is strongly $H_0$-smooth with an exponent $\gamma>1/2$ on any compact sub-interval of $(a,b)$. Let $$H=H_0+G^* AG,$$ where $A$ is a compact self-adjoint operator. Then
1. the a.c. spectrum of $H$ coincides with $[a,b]$ and has a constant multiplicity $m$,
2. $H$ has no singular continuous spectrum,
3. and all eigenvalues of $H$, distinct from $a$ and $b$, have finite multiplicities and can accumulate only to $a$ and $b$.
We will not need part (iii) of this proposition, since the absence of eigenvalues is proven in Section \[sec.b\] by a different method.
We fix $\beta>1$ and write $$H(h)=H(h_*)+Q^\beta AQ^\beta,$$ where $A$ is the integral operator with the integral kernel $$a(t,s)=q(t)^{-\beta}(h(t+s)-h_*(t+s))q(s)^{-\beta}, \quad t,s\in{{\mathbb R}}_+.$$ In view of it is easy to see that $a\in L^2({{\mathbb R}}_+\times{{\mathbb R}}_+)$ and so $A$ is Hilbert-Schmidt, hence compact. Now it remains to use Proposition \[prp.c3\] with $H=H(h)$, $H_0=H(h_*)$ and $G=Q^\beta$.
Absence of embedded eigenvalues: proof of Theorem \[thm.a5\] {#sec.b}
============================================================
Preliminaries
-------------
The key element of the proof is the following lemma.
\[lma.b3\] Let $H(h)$ be as in Theorem \[thm.a5\]. If $H(h)f=Ef$ for some function $f\in L^2({{\mathbb R}}_+)$ and a constant $E$, $0<E\leq\pi$, then $f'\in L^2({{\mathbb R}}_+)$ and $f(0)=0$.
Before embarking on the proof of Lemma \[lma.b3\], let us show how it leads to a proof of Theorem \[thm.a5\].
Differentiating the eigenvalue equation $H(h)f=Ef$, we get $$\int_0^\infty h'(t+s)f(s)ds=Ef'(t), \quad t>0.
\label{b1a}$$ Integrating by parts, we obtain $$H(h)f'=-Ef';$$ the boundary term at zero vanishes because $f(0)=0$ by Lemma \[lma.b3\]. This means that $-E$ is an eigenvalue of $H(h)$ with the eigenfunction $f'$, and $f'\in L^2({{\mathbb R}}_+)$ by Lemma \[lma.b3\]. Since $H(h)\geq0$, this is impossible unless $f'\equiv0$, which implies that $f\equiv0$.
Lemmas on the Mellin transform {#sec.b3}
------------------------------
The remainder of this section is devoted to proving Lemma \[lma.b3\]. Our main tool, following [@BPSSV], is the Mellin transform. See for example [@Titchmarsh] for the background. For $f\in L^2({{\mathbb R}}_+)$, the Mellin transform is defined by $${\mathcal{M}}f(z)=\int_0^\infty s^{z-1}f(s)ds, \quad z\in{{\mathbb C}},$$ as long as the integral converges. It can be shown that the Mellin transform, initially defined on a suitable dense subset of $L^2({{\mathbb R}}_+)$, extends to an isometry between $L^2({{\mathbb R}}_+)$ and the $L^2$ space on the vertical line ${\hbox{{\rm Re}}\,}z=1/2$. In other words, the Plancherel identity $$\frac1{2\pi}\int_{-\infty}^\infty {\lvert{\mathcal{M}}f(\tfrac12+i\tau)\rvert}^2 d\tau
=
\int_0^\infty {\lvertf(s)\rvert}^2ds, \quad f\in L^2({{\mathbb R}}_+)
\label{b2}$$ holds. In this context, the inversion formula for the Mellin transform reads as $$f(s)=\frac1{2\pi i}\int_{\frac12-i\infty}^{\frac12+i\infty}s^{-z}{\mathcal{M}}f(z)dz, \quad f\in L^2({{\mathbb R}}_+).
\label{b3}$$ The Mellin transform is useful to us because it diagonalises the Carleman operator $H(1/t)$. More precisely, we have the identity $${\mathcal{M}}H(1/t)f(z)=\frac{\pi}{\sin(\pi z)}{\mathcal{M}}f(z), \quad {\hbox{{\rm Re}}\,}z=1/2,\quad f\in L^2({{\mathbb R}}_+),
\label{b3a}$$ which is the consequence of the elementary formula $$\int_0^\infty \frac{s^{z-1}}{t+s}ds=\frac{\pi}{\sin(\pi z)}t^{z-1}, \quad 0<{\hbox{{\rm Re}}\,}z<1.$$
\[lma.b4\] Let $h$, ${\widetilde}h$ be as in Theorem \[thm.a5\], and let $g=H({\widetilde}h)f$ for some $f\in \operatorname{Ran}H(h)$. Then the Mellin transform ${\mathcal{M}}g(z)$ extends to a meromorphic function in the strip $$-\frac12-{\varepsilon}<{\hbox{{\rm Re}}\,}z<\frac12+{\varepsilon}\label{b4}$$ where $0 < {\varepsilon}< 1/2$ is as in . This meromorphic extension has at most one pole in the strip; this pole is simple and is located at the origin. The function ${\mathcal{M}}g$ satisfies the estimate $$\int_{-\infty}^\infty {\lvert (\sigma+i\tau){\mathcal{M}}g(\sigma+i\tau)\rvert}d\tau\leq C({\varepsilon}'),
\quad
-\frac12-{\varepsilon}'<\sigma <\frac12+{\varepsilon}',
\label{b5}$$ with any $0 < {\varepsilon}' < {\varepsilon}$.
Since $f\in \operatorname{Ran}H(h)$, from and Cauchy-Schwarz we obtain $$f(t)=\begin{cases}
O(t^{-\frac12}), & t\to0,
\\
O(t^{-\frac12-{\varepsilon}}), & t\to\infty.
\end{cases}
\label{b11}$$ Next, we have $$g^{(k)}(t)=\int_0^\infty {\widetilde}h^{(k)}(t+s)f(s)ds,\quad k=0,1,2.$$ Combining with our assumptions on ${\widetilde}h$, we obtain that $g\in C^2([0,\infty))$ and $$g^{(k)}(t)=O(t^{-k-\frac12-{\varepsilon}}), \quad t\to\infty.
\label{b12}$$
Further, for ${\hbox{{\rm Re}}\,}z=1/2$, integrating by parts twice, we get $$z{\mathcal{M}}g(z)
=
-\int_0^\infty s^z g'(s)ds
=
\frac1{z+1}\int_0^\infty s^{z+1}g''(s)ds.
\label{b13}$$ The integrals here converge absolutely by the estimates ; the boundary terms vanish by the same estimates. The same estimates also show that the right side in has an analytic extension into the strip .
Finally, again by we have $$\int_0^\infty ({\lvertg''(s)\rvert}^2+{\lverts^{2+{\varepsilon}'}g''(s)\rvert}^2)ds\leq C({\varepsilon}'), \quad 0<{\varepsilon}'<{\varepsilon}.$$ By the Plancherel identity for the Mellin transform applied to $s^\alpha g''(s)$ with $0\leq \alpha\leq 2+{\varepsilon}'$, we obtain that $\int_0^\infty s^{z+1}g''(s)ds$ is in $L^2$ on the vertical lines $\sigma+i{{\mathbb R}}$ with $-1/2-{\varepsilon}'\leq\sigma\leq 1/2+{\varepsilon}'$. Taking into account the factor $1/(z+1)$ in front of the integral in , by Cauchy-Schwarz we arrive at the required bound .
\[lma.b5\] Let $h$ be as in Theorem \[thm.a5\], and let $H(h)f=Ef$ for some $f\in L^2({{\mathbb R}}_+)$ and some $0<E\leq\pi$. Then the Mellin transform ${\mathcal{M}}f(z)$ extends to an analytic function in the strip , satisfying the estimate $$\int_{-\infty}^\infty {\lvert (\sigma+i\tau){\mathcal{M}}f(\sigma+i\tau)\rvert}d\tau\leq C({\varepsilon}'),
\quad
-\frac12-{\varepsilon}'<\sigma<\frac12+{\varepsilon}',
\label{b9}$$ for every $0 < {\varepsilon}' < {\varepsilon}$.
Let us write the eigenvalue equation for $f$ as $$H(1/t)f-Ef=-H({\widetilde}h)f.$$ Applying the Mellin transform, letting $g=H({\widetilde}h)f$, and using , we obtain the equation $${\mathcal{M}}f(z)=-{\mathcal{M}}g(z)/u(z), \quad u(z)=\frac{\pi}{\sin(\pi z)}-E.
\label{b10}$$ Initially, this formula is valid for ${\hbox{{\rm Re}}\,}z=1/2$, but Lemma \[lma.b4\] ensures that the right side has a meromorphic extension to the strip .
Consider the poles of this extension. By Lemma \[lma.b4\], there may be a pole at $z=0$; however, this pole is cancelled out by the pole of $u(z)$ at $z=0$. There may also be poles arising due to the zeros of $u(z)$. Inspecting these, we find that the only zeros of $u(z)$ in the strip are given by $$z=\frac12\pm i\theta, \quad
\theta=\frac1\pi\log\left((\pi/E)-\sqrt{(\pi/E)^2-1}\right).$$ If $E=\pi$, these two zeros coalesce into one double zero at $z=1/2$. However, by the Plancherel identity for the Mellin transform, the integral of ${\lvert{\mathcal{M}}f(z)\rvert}^2$ over the vertical line ${\hbox{{\rm Re}}\,}z=1/2$ must be finite. This shows that ${\mathcal{M}}f(z)$ in fact cannot have poles on this line.
Summarizing, we see that defines an *analytic* extension of ${\mathcal{M}}f$ into the strip . The estimate follows from the estimate and from the fact that $u(\sigma+i\tau)\to -E$ as ${\lvert\tau\rvert}\to\infty$.
Proof of Lemma \[lma.b3\]
-------------------------
By , we have $$h'(t)=O(t^{-2}),\quad t\to\infty.$$ From here, applying the Cauchy-Schwarz inequality to , we obtain $$f'(t)=O(t^{-3/2}),\quad t\to\infty.$$ Thus, it remains to inspect the behaviour of $f(t)$ and $f'(t)$ for small $t$.
By the analyticity of ${\mathcal{M}}f(z)$ in the strip and by the estimate , we can shift the contour of integration in the Mellin inversion formula to ${\hbox{{\rm Re}}\,}z=\sigma$ for any $\sigma$ satisfying $-\frac12-{\varepsilon}<\sigma<\frac12$. This gives us that $$f(s)=\frac1{2\pi i}\int_{\sigma-i\infty}^{\sigma+i\infty} s^{-z}{\mathcal{M}}f(z) dz,
\quad
f'(s)=-\frac1{2\pi i}\int_{\sigma-i\infty}^{\sigma+i\infty}zs^{-z-1}{\mathcal{M}}f(z) dz.$$ Again using the estimate , we see that $${\lvertf(s)\rvert}\leq Cs^{-\sigma}, \quad {\lvertf'(s)\rvert}\leq Cs^{-\sigma-1}, \quad s>0.$$ It follows that $f(0)=0$ and that $f'\in L^2(0,1)$.
Proving Theorems \[thm.a1\] and \[thm.a2\] {#sec.aaa}
==========================================
By Theorem \[thm.a4\](i), it suffices to check that $H(h_0)|_{(\operatorname{Ker}H(h_0))^\perp}$ has a purely a.c. spectrum $[0,\pi]$ of multiplicity one. By the properties of the zeta function (see , ) the function $h_0(t)-1/t$ is analytic near $t=0$ and satisfies $$h_0^{(k)}(t)=O(2^{-t}), \quad t\to\infty, \quad k=0,1,2.$$ Thus, $h_0$ satisfies the hypotheses of both Theorem \[thm.a4a\] and Theorem \[thm.a5\]. It follows that $H(h_0)$ has the a.c. spectrum $[0,\pi]$ of multiplicity one, no singular continuous spectrum, and no eigenvalues in $(0,\pi]$. It remains to rule out the eigenvalues in $(\pi,\infty)$. But by the estimate and by Lemma \[lma.d1\], we have ${\lVertH(h_0)\rVert}\leq\pi$, and so $H(h_0)$ has no eigenvalues in $(\pi,\infty)$.
1\) Theorem \[thm.a4\](ii) reduces the question to analysing the spectrum of $H(h_a)|_{(\operatorname{Ker}H(h_a))^\perp}$. As in the proof of Theorem \[thm.a1\], by the properties of the zeta function it is straightforward to check that the kernel function $h_a$ satisfies the hypotheses of both Theorem \[thm.a4a\] and Theorem \[thm.a5\]. It follows that $H(h_a)$ has the a.c. spectrum $[0,\pi]$ of multiplicity one, no singular continuous spectrum, and no eigenvalues in $(0,\pi]$. So it remains to analyse the eigenvalues of $M(g_a)$ (which coincide with those of $H(h_a)$) in the interval $(\pi,\infty)$.
2\) For $E\geq\pi$, let us denote by $N((E,\infty);M(g_a))$ the total number of eigenvalues of $M(g_a)$ in the interval $(E,\infty)$, counting multiplicity. Let us prove that $N((E,\infty);M(g_a))$ is non-increasing in $a>0$. It suffices to prove that $M(g_a)$ is monotone non-increasing in the standard quadratic form sense, i.e. $$(M(g_a)x,x)\leq (M(g_b)x,x), \quad 0<b<a, \quad x=\{x_n\}_{n=1}^\infty\in \ell^2({{\mathbb N}}).
\label{a5a}$$ By the calculation in the proof of Theorem \[thm.a4\], we have that $$(M(g_a)x,x)=({\mathcal{N}}_a^* x,{\mathcal{N}}_a^* x)=\int_0^\infty e^{-at}{\left\lvertf(t)\right\rvert}^2dt,
\quad f(t)=\sum_{n=1}^\infty x_n n^{-\frac12-t}.$$ The required monotonicity obviously follows from this representation.
3\) Let us prove that $M(g_a)$ is continuous in $a>0$ in the operator norm. Taking $0<b<a$, we have, as in the previous step, and with $f(t)$ as above, $$\begin{gathered}
((M(g_b)-M(g_a))x,x)=\int_0^\infty (e^{-bt}-e^{-at}){\lvertf(t)\rvert}^2 dt
\\
=\int_0^\infty (1-e^{-(a-b)t})e^{-bt}{\lvertf(t)\rvert}^2 dt
\leq
\sup_{t>0}e^{-bt/2}(1-e^{-(a-b)t}){\lVertM(g_{b/2})x\rVert}^2.\end{gathered}$$ This supremum tends to zero as $a\to b$, and we therefore obtain the desired claim.
4\) By , we have $$h_a(t)=e^{-at/2}\sum_{n=1}^\infty n^{-t-1}
\leq
e^{-at/2}(1+1/t)=\frac{e^{-at/2}}{t}(1+t)\leq \frac{e^{-at/2}e^t}{t},$$ and so for $a\geq2$ we have $h_a(t)\leq 1/t$. Thus, by Lemma \[lma.d1\], $H(h_a)$ has norm less or equal to $\pi$ for $a\geq2$, and hence the same is true for $M(g_a)$. Thus, for $a\geq2$ the operator $M(g_a)$ has no eigenvalues in $(\pi,\infty)$.
5\) Let us check that for any $a>0$, the operator $H(h_a)$ has at most one eigenvalue in $(\pi,\infty)$. Let $$h_a(t)=e^{-at/2}+w_a(t), \quad
w_a(t)=e^{-at/2}\sum_{n=2}^\infty n^{-t-1}, \quad t>0.$$ By , we have $w_a(t)\leq 1/t$, and therefore, by Lemma \[lma.d1\], $H(w_a)$ has no eigenvalues in $(\pi,\infty)$. On the other hand, the Hankel operator corresponding to the kernel function $e^{-at/2}$ is the rank one operator with the integral kernel $e^{-a(t+s)/2}$, which we naturally denote by $(\cdot,e^{-at/2})e^{-at/2}$. It has the single non-zero eigenvalue $1/a$. Note that $$H(h_a)=(\cdot,e^{-at/2})e^{-at/2}+H(w_a).$$ From here our claim follows by a standard argument in perturbation theory. Indeed, for $E\geq\pi$, let us write the min-max principle in the form $$N((E,\infty);H(h_a))
=
\sup\{\dim L: (H(h_a)f,f)>E{\lVertf\rVert}^2 \quad \forall f\in L \setminus \{0\}\},
\label{a6a}$$ where the supremum is taken over all subspaces $L$ with the indicated property. We claim that $$N((\pi,\infty);H(h_a))\leq1.
\label{a6b}$$ Assume that this is false and take a subspace $L$ as in with $\dim L\geq2$. Then there is a non-zero element $f\in L\cap\{e^{-at/2}\}^\perp$ which satisfies $$(H(h_a)f,f)=(H(w_a)f,f)\leq\pi{\lVertf\rVert}^2,$$ in contradiction with the inequality in .
6\) Next, let us check for $0<a<1/\pi$ that the operator $H(h_a)$ has at least one eigenvalue in $(\pi,\infty)$. We claim that $$N((\pi,\infty);H(h_a))\geq1, \quad a<1/\pi.
\label{a6c}$$ To prove , let $f=e^{-at/2}$. Observe that $H(w_a)\geq0$; this follows by the same argument as in the proof of Theorem \[thm.a4\]. Thus, $$(H(h_a)f,f) \geq (f,e^{-at/2})(e^{-at/2},f)=1/a{\lVertf\rVert}^2>\pi{\lVertf\rVert}^2,$$ and so holds by choosing $L=\operatorname{span}\{e^{-at/2}\}$ in the min-max principle.
7\) Let us put together the above steps. By and , $N((\pi,\infty);M(g_a))=1$ for $0<a<1/\pi$, and, by step 4), $N((\pi,\infty);M(g_a))=0$ for $a\geq2$. By the monotonicity in step 2), there must exist a critical $a_*>0$ such that $$N((\pi,\infty); M(g_a))=
\begin{cases}
0 & a>a_*,
\\
1 & a<a_*.
\end{cases}$$ The norm continuity of $M(g_a)$ ensures that $N((\pi,\infty); M(g_a))$ is lower semi-continuous in $a$, and so $N((\pi,\infty); M(g_{a_*}))=0$.
Thus, we have exactly one eigenvalue $\lambda(a)$ for $0<a<a_*$ and no eigenvalues for $a\geq a_*$. The norm continuity and monotonicity of $M(g_a)$ ensures that $\lambda(a)$ is a continuous monotone function of $a$ with $\lambda(a) \to \pi$ as $a \to a_*^{-}$.
The above argument also gives the upper and lower bounds $1/\pi\leq a_*\leq 2$ for the critical value of $a$. It also shows that the eigenvalue $\lambda(a)$ satisfies $1/a<\lambda(a)\leq \pi+1/a$. Indeed, the argument in step 6) gives that $$N((1/a,\infty); H(h_a))\geq1,$$ i.e. $\lambda(a)>1/a$. On the other hand, the norm of $H(h_a)$ satisfies $${\lVertH(h_a)\rVert}\leq {\lVertH({\mathcal{L}}\nu_a)\rVert}+{\lVert(\cdot,e^{-at/2})e^{-at/2}\rVert}=\pi+1/a,$$ yielding the upper bound for $\lambda(a)$.
[1]{}
*Higher Transcendental Functions,* Bateman Manuscript Project, Vol. 2, McGraw–Hill, New York, 1953.
*The multiplicative Hilbert matrix,* Adv. Math. **302** (2016), 410–432.
*Hankel forms,* Studia Math. **198** (2010), 79–83.
*Spectral theory of operators of Hankel type, II,* Indiana University Mathematics Journal, **41** no. 2 (1992), 427–434.
*Some singular integral equations connected with integral representations of mathematical physics. (Russian)* Doklady Akad. Nauk SSSR (N.S.) **65** (1949), 621–624.
*The analogue of Parseval’s theorem for a certain integral transform.* (Russian) Doklady Akad. Nauk SSSR (N.S.) **68** (1949), 653–656.
*On the spectrum of Hilbert’s matrix,* American Journal of Mathematics, **72**, no. 4 (1950), 699–704.
*Hankel operators with discontinuous symbols,* Proc. Amer. Math. Soc. **65** (1977), 77–79.
*Spectral and scattering theory of self-adjoint Hankel operators with piecewise continuous symbols,* J. Operator Theory **74**, no.2 (2015) 417–455.
*Diophantine approximation and Dirichlet series.* Hindustan Book Agency, New Delhi, 2013.
*On the Hilbert matrix, I, II,* Proc. Amer. Math. Soc., **9** (1958), 137–140, 581–585.
*Some open questions in analysis for Dirichlet series.* Recent progress on operator theory and approximation in spaces of analytic functions, 179–191, Contemp. Math., **679**, Amer. Math. Soc., Providence, RI, 2016.
*Introduction to the theory of Fourier integrals.* Third edition. Chelsea Publishing Co., New York, 1986.
*A Commutator Method for the Diagonalization of Hankel Operators,* Functional Analysis and Its Applications, **44** (2010), no. 4, 295–306.
*Mathematical scattering theory. General theory.* Amer. Math. Soc., Providence, RI, 1992.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
author:
- |
M.A.Braun$^{a,b}$ and C.Pajares$^a$\
$^a$ Dep. of Elementary Particles,\
Univ. of Santiago de Compostela, 15706, Santiago de Compostela, Spain,\
$^b$ Dep. of High Energy physics, University of S.Petersburg,\
198504 S.Petersburg, Russia
title: '**Rapidity and centrality dependence in the percolating colour strings scenario**'
---
epsf
-30pt plus 1pt minus 1pt
In AA collisions fusion and percolation of colour strings is studied at fixed rapidity $y$. Distribution of strings in rapidity is obtained from the observed rapidity spectra in pp collisions. For $y$-dependence of multiplicities in Au-Au collisions good agreement is obtained with the existing experimental data. Predictions for LHC energies coincide with the extrapolation of the data. Agreement with the data of the transverse momentum spectra requires introduction of quenching into the model.
Introduction
============
The color string model with fusion [@Ref1], [@Ref2] and percolation [@Ref3]-[@Ref6] has produced results on multiplicities of secondaries which are in general agreement with the existing experimental data. The string fusion model predicted a strong reduction of multiplicities both at RHIC and LHC energies. At the time of the ALICE Technical proposal of 95 [@Ref7], most of model predictions, (including VENUS [@Ref8], HIJING [@Ref9], SHAKER [@Ref10] and DPM [@Ref11]) for LHC energies were more than 4000 charged particles at central rapidity region for central $(b\leq 3\ fm)$ Pb-Pb collisions and only the prediction of the string fusion model was much lower. Since then many of the models have lowered their predictions introducing several mechanisms, such as the triple pomeron coupling in DPM [@Ref12], stronger shadowing in HIJING [@Ref13] or other modifications in VENUS [@Ref14]. In the parton saturation picture, predictions for the central rapidity density of charged particles per participant for central Pb-Pb at the LHC energy range from around 15 [@Ref15] to a lower value around 9 [@Ref16]. Assuming that the observed geometrical scaling for the saturation momentum in lepton-hadron scattering is also valid for the nucleus-nucleus scattering, the value around 9 is also obtained [@Ref17]. In percolation of strings the obtained values 7.3 [@Ref18] and 8.6 [@Ref19] are not far from the above ones, as expected, given the similarities between percolation of strings and saturation of partons [@Ref20]. In any case, these values lie above 6.4 which is obtained by extrapolation to the LHC energy of the values experimentally found at $\sqrt{s}=19.4, 62.8, 130$ and 200 GeV [@Ref21]. The values obtained in percolation have some uncertainty due to simplifications done in the calculations, mainly related to the dependence of the multiplicity on the energy of a simple string and on the rapidities of fused strings. One of the goals of this paper is to take into account both these dependencies, using as an input the rapidity and energy dependence of multiplicities in pp collisions. We try to answer whether the charge particle density per participant is compatible with the one obtained in the mentioned extrapolation. We assume that strings occupy different regions in the available rapidity interval. Then at different rapidities one will see different number of overlapping strings, depending on the rapidities of the string ends. As a result the string density and the process of percolation become dependent on rapidity together with all the following predictions for observable quantities. The central part of our derivation is the calculation of the number of strings at a fixed rapidity which follows from the known parton distributions in the projectile and target. To this aim we shall use the simple parton distributions employed in the review [@Ref11]. Also, we pay attention to the fragmentation rapidity region. As an input we take the observed limiting fragmentation scaling in pp collisions. We then obtain a similar scaling for A-A collisions, in agreement with the experimental data [@Ref22].
One of the most interesting features of the RHIC data is the suppression of high transverse momenta. The nuclear modification factor defined as the ratio between inclusive A-A cross section normalized to the number of collisions and the inclusive proton-proton cross-section is found to lie below unity, in disagreement with the perturbative QCD expectations.
In the previous paper [@Ref6], in the framework of percolation of strings, a reasonable agreement with the data was obtained, describing A-A collisions as an exchange of clusters of overlapping strings. In percolation each cluster behaves like a new string with a larger tension, its value depending on the number of strings fused into the cluster and the cluster’s transverse area. Fragmentation of each cluster was assumed to give rise to an exponential distribution in $p^2_T$. Superposition of different exponential distributions then builds up a power-like distribution $\sim p_T^{-\kappa}$ with $\kappa$ inversely related to the magnitude of the dispersion in the number of different clusters. At low density, there is no overlapping of strings, thus no fluctuations and $\kappa$ is large. As the density increases, so does the string overlapping and more clusters are formed with different number of strings. So the dispersion increases and $\kappa$ decreases. Finally, at very high density, above the percolation threshold, there remains a single large cluster of nearly all the strings. Therefore, there are no fluctuations and $\kappa$ increases again becoming large. In this way, the suppression of high $p_T$ at large density follows as a result of formation of large clusters of color strings. In [@Ref6] we assumed that the spectrum of a simple string was a single exponential in $p^2_T$. However, at low string density, as in pp collisions, when fusion of strings is insignificant, the experimental data clearly show a power-like tail for the $p_T$ distribution.
In this paper we study this point with more attention. Instead of the exponential distribution we take the standard power-like parameterization for pp collisions [@Ref23]. The resulting $p_T$ distribution for central A-A collisions is again found suppressed at high $p_T$ but not enough to agree with the data. In order to describe the data we need additional suppression which would physically correspond to the fact that the produced particle, passing through a large cluster and interacting with the strong chromoelectric field, looses a part of its energy. This result is not unexpected. In fact a version of HIJING [@Ref24], with a string junction and doubling the string tension to simulate stronger color-fields, is able to explain the difference between baryons and mesons in the low and mid $p_T$ range but at high $p_T$ some jet quenching mechanism is needed. In our framework quenching at high $p_T$ may be introduced in a simple phenomenological manner by taking the average $p_T^2$ of a cluster of $n$ strings to grow with $n$ more slowly than $\sqrt{n}$, as predicted for a single small cluster in absence of others. Choosing an appropriate $n$ dependence of the average $p_T^2$ for clusters at a given string density allows to obtain a reasonable agreement with the experimental data on $p_T$ dependence in A-A collisions.
pp collisions
=============
Multiplicities and numbers of strings
-------------------------------------
The starting point for the calculation of fusion and percolation of strings in heavy -ion collisions is the distribution of strings in proton-proton collisions, where effects of fusion and percolation are very small. Our strategy will be to extensively use the existing experimental data for the multiplicity per unit rapidity $d\mu^{pp}(y)/dy\equiv \mu^{pp}(y)$ in pp collisions to extract the necessary distribution of strings in $y$ from them.
We recall that in the original DPM model without string fusion the multiplicity is given by a sum of contribution from the strings formed in the collision. In particular, in a configuration with $n=2k$ strings formed, which corresponds to the exchange of $k$ pomerons [@Ref11], the multiplicity is given by \^[pp]{}\_[n]{}(y)=\_[-Y/2]{}\^[Y/2]{} \_[i=1]{}\^ndu\_idw\_i p(u\_1,...,u\_n) t(w\_1,...,w\_n)\_[j=1]{}\^n\_j(y,u\_j,w\_j). We work in the c.m system of colliding protons; $Y$ is the overall rapidity admissible for nearly massless quarks. It is related to the beam rapidity as Y=Y\_[beam]{}+, where $m$ is the nucleon mass and $\mu$ the quark average transverse mass. The strings are enumerated according to their flavour content, that is according to which quark they are attached. Number 1 corresponds to the quark-diquark (qd) string, number 2 to diquark-quark (dq) string and all the rest correspond to sea quarks, which include $s\bar{s}$ and $\bar{s}s$ strings. Ends of strings in rapidity are denoted by $u_i$ in the projectile and $w_i$ in the target. Distributions $p(u_1,...,u_n)$ and $t(w_i,...w_n)$ give the probability to find the relevant quarks in the projectile and target proton respectively. Note that in the assumed notation $p(u_1,...u_n)$ gives the probability to find in the proton the valence quark at rapidity $u_1$ the diquark at rapidity $u_2$ and the sea quarks at rapidities $u_3,...u_n$. Distribution $t(w_1,...w_n)$, on the other hand, gives the probability to find the diquark at rapidity $w_1$, the quark at rapidity $w_2$ and sea quarks at rapidities $w_3,...w_n$. Function $\mu_j(y,u,w)$ gives the multiplicity per unit rapidity from the $j$th string at rapidity $y$ provided its ends are at $u$ in the projectile and $w$ in the target. Obviously this probability is zero if the string lies outside rapidity $y$. So $\mu_j(y,u,w)$ has a form \_j(y,u,w)=(y,u,w)\_j(y,u,w), where (y,u,w)=(u-w-y\_0)(u-y)(y-w) +(w-u-y\_0)(w-y)(y-u). Here the two terms correspond to the two possibilities of the higher rapidity end of the string to lie on the projectile or on the target parton. The rapidity interval $y_0$ corresponds to the minimal extension of the string in rapidity. We take $y_0=2$.
The string density $dN^{pp}_{n}(y)/dy\equiv
N^{pp}_{n}(y)$ at a given rapidity is given by an expression similar to (1) but without $\tilde{\mu}$: N\^[pp]{}\_[n]{}(y)=\_[-Y/2]{}\^[Y/2]{}\_[i=1]{}\^ndu\_idw\_i p(u\_1,...,u\_n) t(w\_1,...,w\_n)\_[j=1]{}\^n(y,u\_j,w\_j). To analyse fusion probabilities in nuclear collisions we need to know the latter quantity.
In principle, knowledge of the distributions $p(u_1,...,u_n)$ and $t(w_1,...,w_n)$ and of string luminosities $\tilde{\mu}_j(y,u,w)$ allows to calculate both $\mu_n^{pp}(y)$ and $N_n^{pp}(y)$. This was done in the extensive calculations within the original DPM model [@Ref11]. However these input quantities are in fact poorly known and our idea is to directly relate $\mu^{pp}(y)$ and $N^{pp}(y)$ using the experimental data for the former. To do this we assume that string luminosities are approximately $y$-independent and the same for all type of strings: \_j(y,u,w)=\_0. This approximation has been widely used in analytical studies of string fusion. It can be justified for relatively long strings far from their ends, when particle production can be well described by the Schwinger mechanism of pair creation in a strong field. With strings of finite dimension it may be considered as a sort of averaging over their length and rapidity of emission. With this approximation we obtain a simple and direct relation between the multiplicity and number of strings per unit rapidity: \^[pp]{}\_[n]{}(y)=\_0N\^[pp]{}\_[n]{}(y). In fact this relation for the central region was extensively used in earlier studies of string percolation.
Relation (7) allows to find only the total number of strings per unit rapidity from the experimental data on multiplicities. However we need something more. In nucleus-nucleus collisions separately enter multiplicities coming from the valence strings and sea strings. Obviously one cannot find each of them from the experimental data. So we choose to calculate the contribution of valence strings from the theoretical formulas (5) and (7) and then, subtracting this contribution from the experimental multiplicities, find the contribution from sea strings. This procedure can be justified by the fact that distributions of valence quarks are much better known and less dependent on the overall energy than the sea contribution.
Total and sea strings from the experimental data
------------------------------------------------
We calculate the number of quark-diquark strings from (5) as $$N^{qd}_{n}(y)=
\int_{-Y/2}^{Y/2}dudwq_n^{(p)}(u)d_n^{(t)}(w)\rho(y,u,w)$$= \_[y]{}\^[Y/2]{}du\_[-Y/2]{}\^[w\_1]{}dw (q\_n \^[(p)]{}(u)d\_n\^[(t)]{}(w)+d\_n\^[(t)]{}(u)q\_n\^[(p)]{}(w)), w\_1={y,u-y\_0}. Here $q_n ^{(p)}(u)$ and $d_n^{(t)}(w)$ are inclusive probabilities to find a valence quark in the projectile and a diquark in the target at rapidities $u$ and $w$ respectively in a configuration with $n$ strings. The second term in (8) corresponds to inverse strings whose upper ends lie on the target diquark. In our symmetric case the number of diquark-quark strings is obviously the same, so that the total number of valence strings is just twice the expression (8).
The final number of valence strings at given $y$ is obtained after averaging over the number of formed strings: N\^[v]{}(y)=2\_[k=1]{}\_kN\^[qd]{}\_[(2k)]{}(y)2N\^[qd]{}\_[(n)]{}(y). Here $\omega_k$ is the probability for the exchange of $k$ pomerons, given by [@Ref11] \_k=, where $\sigma_k$ is the cross-section for $k$ inelastic collisions. It is standardly taken in the K.A.Ter-Martirosyan model [@Ref25] \_n(s)=2\_0\^bdbe\^[-2]{}, where the eikonal $\chi$ corresponds to the single pomeron exchange (s,b)=C(s)e\^[-b\^2/b\_0\^2(s)]{}, with b\_0\^2(s)=4R\_N\^2+4’(s-i), C(s)=(se\^[-i/2]{})\^[-1]{}, $\alpha$ and $\alpha'$ are the pomeron intercept and slope, $g$ is its coupling to the proton and $R_N$ the proton radius. Some improvement of these $\sigma_k$ to include the triple pomeron interaction and diffractive states may be found in [@Ref11].
To calculate the number of valence strings per unit rapidity we have to know the inclusive distributions of quarks and diquarks. Following [@Ref11] we choose the exclusive distribution $p(u_1,u_2,...u_n)$ for a projectile in a factorized form p\_n(u\_1,u\_2,...u\_n)=c\_n(1-\_[i=1]{}\^nx\_i)\_[i=1]{}\^n x\_i\^[\_i]{}, where for the quark $\mu_1=1/2$ and for the diquark $\mu_2=5/2$ For the sea quarks and antiquarks we take $\mu=1/|\ln x_c|$, where in accordance with [@Ref11] $x_c=m_c/\sqrt{s}$ with $m_c=0.1$ GeV is a cutoff at small $x$. Scaling variables are related to rapidities as x=e\^[-Y/2+u]{}, Note that the distributions $p_n(u_1,u_2,...u_n)$ are defined and normalized in the interval $0<x<1$, that is for $-\infty<u<Y/2$. The actual strings are formed only in the part of this interval with $u>-Y/2$. This circumstance is inessential for valence quarks whose distributions rapidly vanish towards small values of $x$. For the distributions in the target one has only to invert the rapidities $u\to -w$ in (15). Integration over the scaling variables of unobserved partons gives the desired inclusive distributions. For the valence quark we find q\_p\^[(n)]{}(x)= c\_vx\^[1/2]{}(1-x)\^[3/2+(n-2)]{}, c\_v=. For the diquark d\_p\^[(n)]{}(x)=c\_d x\^[5/2]{}(1-x)\^[-1/2+(n-2)]{}, c\_d= .
After the averaged valence string number is found according to (9) we have to transform it into the valence multiplicity using (7). The value of $\mu_0$ can be found from the observed plateau hight assuming that at $y=0$ all strings contribute. Their average number can be found from (9) as $N^{pp}=2\langle k\rangle$. As a result we find values of $\mu_0$ slowly rising with energy and visibly saturating at TeV energies. In Fig. 1 we show these values extracted from the data [@Ref23] together with their extrapolation to the LHC energies in the assumption that the plateau in the pp multiplicity distribution rises linearly with $\ln s$.
The obtained multiplicities from valence strings vanish in the fragmentation region too slowly as compared to the experimental data, and at $y-Y_{beam}>0$ become greater than the latter. This is obviously related to our assumption of a constant string luminosity throughout the string length, whereas it should go to zero at its ends. Put in other words, in our approach the total energy is conserved in its division between different strings due to the $\delta$-function in (14) but it is not conserved inside each separate string, since near the string end its luminosity should vanish. To cure this defect in a simple manner we just assume that as soon as the calculated valence contribution becomes larger that the data we substitute the former by the latter, assuming that in this deep fragmentation region sea strings do not contribute at all.
With thus obtained valence contribution to the multiplicities we find the sea contribution just as the difference between the total and valence one. Dividing it by $\mu_0$ we find the number of sea strings per unit rapidity in pp collisions $N^{s,pp}(y)$. In fact we need not exactly this number but the one in the assumption that all strings are of the sea type, which is obtained from it by rescaling N\^[s]{}(y)= N\^[s,pp]{}(y). This quantity is shown in Fig. 2 for different values of the collision energy $\sqrt{s}$.
hA and AA collisions
====================
Generalization to hA and AA collisions is straightforward and follows [@Ref11]. At fixed impact parameter $b$ one introduces the average numbers of participants $2\nu_A(b)$ and collisions $\nu(b)$. Then the number of strings in AA collisions at given $b$ and $y$ is N\_[AA]{}(b,y)=\_[par]{}(b) N\^[pp]{}(y)+(\_[col]{}(b)-\_[par]{}(b)) N\^s(y). Here $N^{pp}(y)$ is obtained from the observed multiplicity in pp collisions according to (7) and $N^s(y)$ is given by (18).
One can easily further generalize (19) to collisions of different nuclei (see [@Ref11]).
In the Glauber approach the numbers $\nu_{par}$ and $\nu_{col}$ for AA collisions are obtained as follows \_[par]{}(b)=A, \_[col]{}(b)=A\^2, where $\sigma$ is the total pp-cross-section, $T_A(b)$ is the nuclear profile function normalized to unity and T\_[AA]{}(b)=d\^2b’T\_A(b’)T\_A(b-b’). For hA collision, as mentioned, $\nu_{par}=1$ and \_[col]{}(b)=A. Note that for AA collisions in the above formulas the denominator is written in the so-called optical approximation [@opt]. As is well-known, it works reasonably well except close to the nucleus boundary, where the collision numbers obtained from (20) may be quite deceptive.
It is customary to take the profile function $T_A(b)$ generated by the Woods-Saxon nuclear density. However for our purpose it is more convenient to assume the nucleus to have a well-defined radius, which allows to determine the interaction area in the transverse plane as just the area of the overlap. For this reason we take the nucleus as a sphere of radius $R_A= A^{1/3}\cdot 1.2$ fm, which gives T\_A(b)=(R\_A-b), where $V_A$ is the nuclear volume. With this choice the most peripheral collisions occur at $b=2R_A$ with $\nu_{col}=\nu_{par}=1$.
Multiplicities and $p_T$ distributions
======================================
In our previous studies of string fusion we always stressed that it can only occur in the common rapidity interval. However our attention was mostly centered on the central rapidity region where all (or nearly all) strings contribute, so that the requirement of common rapidity interval was of no relevance and strings could be considered as of practically infinite length in rapidity. Now we study the fusion process in more detail. At a fixed rapidity $y$ only strings which pass through this rapidity can fuse. Formulas of the previous sections allow to find the original number of strings $N(b,y)$ stretched between the projectile and target at fixed rapidity layer $y$ and impact parameter $b$. According to the percolation colour strings scenario these strings in fact fuse into strings with higher colour. The intensity of fusion is determined by the dimensionless percolation parameter $\eta$ proportional to the string density in the interaction area (b,y)=, where $s_0=\pi r_0^2$ is the transverse area of the string and $S(b)$ is the interaction area, that is, the overlap area in case of AB collisions. Obviously in our case the percolation parameter depends both on $b$ and $y$. At $\eta\sim 1.2\div 1.3$ fusion of strings leads to their percolation and formation of macroscopic string clusters. This phenomenon will take part only in restricted intervals of $b$ and $y$, predominantly at central collisions and rapidities, where the effects of string fusion and percolation will be most noticeable
Considering the case of AA collisions at reasonably high energies we shall assume the total number of strings high enough to allow use of the thermodynamic limit, in which the total areas of $n$-fold fused strings $S_n$ become distributed according to the Poisson law with $\langle n\rangle =\eta(b,y)$: S\_n(b,y)=S(b)e\^[-(b,)]{}. Due to averaging of the direction of colour, the $n$-fold fused string emits the number of particles which is only $\sqrt{n}$ times greater that the simple string. So the total production rate at fixed $b$ and $y$ will be given by (b,y)=\_0e\^[-(b,)]{}\_[n]{} , where $\mu_0$ is the production rate from the single string, which, as stated above, we assume to be independent of $y$ but dependent on energy.
As to the $p_T$ distribution, we use a slightly generalized model introduced and discussed in [@Ref26], in which the normalized probability $w_n(p)$ to find a particle with transverse momentum $p$ emitted from the $n$-fold fused string is given by w\_n(p)= ()\^. Here for $n=1$ the parameters are determined by the experimental data on pp collisions: p\_1=2 [GeV/c]{}, \_1=19.7-0.86E\_[cm]{} and $E_{cm}$ is the c.m. energy in GeV. With $n>1$ from the string fusion scenario it follows that the average transverse momentum squared of the particles emitted from the $n$-fold fused string is $n^{1/2}$ greater than for a single string. This gives a relation between $p_n$ and $\kappa_n$ p\_n\^2=n\^[1/2]{}p\_1\^2. So the distribution from $n$-fold string is fully determined by the $n$-dependence of $\kappa_n$. In [@Ref26] the simplest choice of $n$-independent $\kappa_n$ was used. However this simple choice does not allow to obtain the $p_T$ dependence in agreement with the data at RHIC. To improve our description, we introduce corrections to the original string picture which correspond to non-linear phenomena in string clustering and influence both $\kappa_n$ and the behaviour of $p_n$
In fact the value of $\kappa$ controls the difference of the distribution from the purely exponential one, passing into the latter at very large $\kappa$. For a single string a finite $\kappa_1$ may be thought of as a result of fluctuations in the string tension (or equivalently its transverse area) [@Ref27]. One may expect this fluctuations to grow as many string fuse and so the value of $\kappa_n$ should fall with $n$. However there is another effect acting in the opposite direction. As many strings fuse into clusters, multiple interactions of emitted particle inside the clusters should lead to thermalization of the particle spectra making it closer to an exponential. Thus eventually at large $n$ parameter $\kappa_n$ should grow to large values. Naturally we cannot determine the exact form of the $n$ dependence of $\kappa_n$ from purely theoretical reasoning. We can only think that the change from fall to growth should occur in the vicinity of the percolation threshold and that in any case $\kappa$ cannot be smaller than 4 to have a convergent $<p^2>$. In practice we take $\kappa_n$ as \_n=\_1+a(n-1)+b(n-1)\^2 and try to adjust $a$ and $b$ to get a better agreement with the experimental data. In fact the results are not very sensitive to the choice of $a$ and $b$ provided they are taken to have a behaviour of $\kappa_n$ in agreement with the above general theoretical observations. However the generalization (30) is not sufficient to bring our predictions in agreement with the observed quenching of the ratios $R_{AA}$ at RHIC. To this aim we have to introduce some quenching also in the string picture at large values of $\eta$. It corresponds to the fact that passing through a large cluster volume and interacting with the strong chromoelectric field the produced particles loose a part of their energy [@Ref28]. On our phenomenological level it would correspond to the behaviour of the average transverse momentum squared as \_n=n\^[\_n]{}<p\^2>\_1 with the exponent $\alpha_n$ less than 1/2 and diminishing with $n$. Similarly to (30) we parameterize \_n=+c(n-1)+d(n-1)\^2. The comparison with the experimental data determines the optimal fit for the parameters $a$, $b$, $c$ and $d$. The resulting values for $\kappa_n$ and $\alpha_n$ are shown in Figs. 3 and 4.
Averaging with the distribution in $n$ we find the final distribution in $p$ from the fusing strings at given $b$ and $y$ as w(p,b,y)=a(b,y)\_[n=1]{}w\_n(p), a\^[-1]{}(b,y)=e\^[(b,y)]{}- 1 (the change in the normalization is due to the restriction $n\geq 1$).
Numerical results
=================
We studied Au-Au collisions at energies 19,4, 62.8, 130, 200 and 6000 GeV corresponding to the existing experimental data and expected at LHC. Using our results on the string numbers in pp collisions we calculated their numbers in nucleus-nucleus collisions at a given $y$. Knowing these numbers and also numbers of participants and collisions we then determined values of the percolation parameter $\eta(y,b)$ at different rapidities and impact parameters. We have taken the transverse radius of the single string 0.3 fm. In Fig. 5 and 6 we illustrate values of $\eta(y,b)$ as a function of $y$ for central collisions and as a function of $b$ at mid-rapidity.
As one can observe, at RHIC and LHC energies these values are quite large, far beyond the percolation threshold. As a result one finds a very substantial reduction in multiplicity calculated according to Eq. (26) with luminosities determined from pp collisions. In Fig. 7 and 8 we show multiplicities as a function of $y$ for central collisions and function of $b$ at midrapidity. They agree rather well with the existing data both in form and absolute values. In Fig. 8 a rather sharp change is seen in the periphery of the nuclei, between $b/2R_A=0.8$ and unity. The RHIC data do not exhibit such a sharp saturation. This behaviour is a direct consequence of using in our calculations the nuclear profile function (23) corresponding to the step function for the nuclear density. A more realistic profile function would lead to a smoother transition from the periphery to center, although requiring a more complicated definition of the interaction area.
Our prediction for the plateau at LHC energy (divided by $\nu_{par}$) is around 9 for the sum of charged and neutral particles (that is 6 for charged), which is lower than derived in other publications [@Ref13]-[@Ref17]. The plateau height for charged particles in central Au-Au collisions as a function of energy is illustrated in Fig. 9. Of course our results are directly related to the chosen string radius and go upward if it is lowered. However then we loose the agreement with the existing data.
Note that our calculations also reproduce quite well the behaviour in the fragmentation region (limiting fragmentation). For this the dependence of the string luminosity on energy proved to be quite important. Without it the nice linear dependence of multiplicities in the fragmentation region is spoiled and the line is widened into a band.
To clearly see the effect of string fusion in Fig 10 we show the multiplicities at $b=0$ without fusion. Their values are several times greater than with fusion and do not agree with the experimental data at all.
Passing to the $p_T$ distributions in Fig. 11 we show the ratios $R_{AA}(p_T)$ for central collisions at midrapidity. The curve for 200 GeV served to determine our parameters $a,b,c$ and $d$ in (30) and (32). Our predictions for the LHC energy show a behaviour similar to the RHIC energy with a still more pronounced quenching effect. In the fragmentation region we prefer to show the ratios $R^{par}_{AA}$ with normalization respective to the number of participants, since in this region the multiplicities are roughly proportional to $\nu_{par}$ due to low string densities. Fig. 12 shows that these ratios are close to unity and may only fall a little below unity at LHC energies.
Conclusions
===========
In the framework of percolation of strings we have obtained a strong reduction of multiplicities at LHC, much larger than the rest of models but in agreement with the extrapolation from the SPS and RHIC experimental data. Due to similarities between percolation of strings and saturation of partons, it would be interesting to explore the possibility for further reduction of multiplicities in the saturation approach. In order to describe the energy dependence of multiplicities in AA collisions we need a rather large transverse size of the elementary string 0.3 fm. This enhances the interaction of strings and so cluster formation, which leads to stronger reduction of multiplicities. In our calculations we have used the standard optical approximation to compute the numbers of participants and collisions. This approximation enhances the number of collisions for peripheral collisions in comparison with Monte-Carlo evaluations and leads to some uncertainties also for central collisions. For this reason, our results must be regarded to have an uncertainty in the range of 10%-15% . So the string transverse size can be lower if the number of strings is in fact lower.
Taking limiting fragmentation scaling for pp collisions as an input, we have found the same behaviour for AA collisions, which is confirmed experimentally up to the RHIC energies. Our calculations predict that limiting fragmetation scaling also remains approximately valid at the LHC energy (with a 5% suppression compared with to SPS or RHIC, see Fig. 7).
We have been able to describe reasonably well the high transverse momentum spectrum at different energies ranging from SPS to RHIC. A large suppression is predicted for LHC. In order to obtain such an agreement, in addition to the usual effects of string clustering, such as reduction of the effective number of independent color sources and suppression of transverse momentum fluctuations, we need a shift of the $p_T$ spectrum due to energy loss. Considering string fusion as an initial state effect (before particle production), a final state effect is needed to account for the observed suppression, similarly to jet quenching in the QCD picture.
Acknowledgments
===============
This work has been partially done under contracts FPA2005-01963of Spain, and PGIDIT03PX1 of Galicia, and also supported by the NATO grant PST.CLG.980287 and Education Ministry of Russia grant RNP 2.1.1.1112.
[99]{} M.A. Braun and C. Pajares, Phys. Lett. B [**287**]{}, 154 (1992); Nucl. Phys. [**B390**]{}, 542 (1993).
N. Amelin, M.A. Braun and C. Pajares, Z. Phys. [**C63**]{}, 507 (1994);\
H.J. Mohring, J. Ranft, C. Merino and C. Pajares, Phys. Rev. [**D47**]{}, 4142 (1993);\
N. Armesto, C. Pajares and D. Sousa, Phys. Lett. [**B257**]{}, 92 (2002).
N. Armesto, M. A. Braun, E. G. Ferreiro and C. Pajares, Phys.Rev. Lett. [**77**]{}, 3736 (1996);\
M. Nardi and H. Satz, Phys. Lett. [**B442**]{}, 14 (1998).
J. Dias de Deus and R. Ugoccioni, Phys. Lett. [**B491**]{}, 253 (2000); Phys. Lett. [**B494**]{}, 53 (2000).
M.A. Braun, F. del Moral and C. Pajares, Phys. Rev. [**C65**]{}, 024907 (2002).
J. Dias de Deus, E.G. Ferreiro, C. Pajares and R. Ugoccioni, Eur. Phys. J. [**C40**]{}, 229 (2005).
ALICE Collaboration. CERN/LHCC/95-71 (1995).
K. Werner, Phys. Rep. [**232**]{}, 87 (1993).
X.N. Wang and M. Gyulassy, Phys. Rev. [**D44**]{}, 3501 (1991); Comput. Phys. Commun [**83**]{} 307 (1994).
N. van Eijndhoven et al. Internal Note ALICE 95-32 (1995).
A. Capella, U.P. Sukhatme, C.I. Tan and J. Tran Thanh Van,Phys. Rep. [**236**]{}, 225 (1994);\
A. Capella, C. Pajares and A.V. Ramallo,Nucl. Phys. [**B241**]{}, 75 (1984).
A. Capella, A.B. Kaidalov and J. Tran Thanh Van, Heavy Ion Physics [**9**]{}, 169 (1999).
X. Wang and M. Gyulassy, Phys. Rev. Lett. [**86**]{}, 3496 (2001).
H.J. Drescher, M. Hlakik, S. Ostapchenko, T. Pierog and K.Werner, Phys. Rep. [**350**]{}, 93 (2001).
A. Accardi, Phys. Rev. [**C64**]{}, 064905 (2001).
D. Kharzeev, E. Levin and M. Nardi, Phys. Rev. [**C72**]{}, 044904 (2005).
N. Armesto, C.A. Salgado, U.A. Wiedemann, Phys. Rev. Lett. [**94**]{}, 022002 (2005).
J. Dias de Deus and R. Ugoccioni, Phys. Lett. [**B494**]{}, 53 (2000).
M.A. Braun, F. del Moral and C. Pajares, Phys. Rev. [**C65**]{}, 024907 (2002)\
C. Pajares, Eur. Phys. J. [**C43**]{}, 9 (2005).
M.A. Braun and C. Pajares, Phys. Lett. [**B603**]{}, 21 (2004).
W. Busza, Acta Phys. Hungarica [**A24**]{}, 3 (2005).
PHOBOS Collaboration, B.B. Back et al., Nucl. Phys. [**A757**]{}, 28 (2005).
A. Drees, Nucl. Phys. [**A698**]{}, 331 (2002).
V. Topor Pop, M. Gyulassy, J. Barette and C. Gale, Phys. Rev., 054901 (2005).
K.A. Ter Martirosyan, Phys. Lett. [**B44**]{}, 377 (1973).
A. Bialas, M. Bleszynski and W. Czyz, Nucl. Phys. [**B111**]{}, 461 (1976)\
C. Pajares and A.V. Ramallo, Phys. Rev. [**C16**]{}, 2800 (1985).
M.A. Braun and C. Pajares, Phys. Rev. Lett. [**85**]{}, 4864 (2000).
A. Bialas, Phys. Lett. [**B466**]{}, 301 (1999).
M. Gyulassy, I. Vitev, X.N. Wang and B.W. Zhang. Published in Quark Gluon Plasma 3, editors: R.C. Hwa and X.N. Wang (World Scientific, Singapore, 2004);\
R. Baier, D. Schiff and B.G. Zakharov, Ann. Rev. Nucl.Part. Sci. 50, 37 (2000).
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'Experimental configuration for investigating the dynamics and the statistics of the phase locking level of coupled lasers that have no common frequency is presented. The results reveal that the probability distribution of the phase locking level of such coupled lasers fits a Gumbel distribution that describes the extreme value statistic of Gaussian processes. A simple model, based on the spectral response of the coupled lasers, is also described, and the calculated results are in good agreement with the experimental results.'
address: |
Weizmann Institute of Science, Dept. of Physics of Complex Systems, Rehovot 76100, Israel\
$^*$Corresponding author: [email protected]
author:
- 'Moti Fridman, Rami Pugatch, Micha Nixon, Asher A. Friesem, and Nir Davidson$^{*}$'
title: Extreme value statistics in coupled lasers
---
Phase locking of coupled oscillators was studied over the years in many different contexts including chemical oscillators with mutual coherence [@V1], arrays of Josephson junctions that are frequency locked [@V2] and arrays of coupled lasers that are phase locked [@V3; @V4; @V5]. In these, complete phase locking occurs when all the oscillators have at least one common frequency. When there is no common frequency, the oscillators group in several clusters, where each cluster oscillates at a different frequency [@Strogatz]. So far there is very little, if any, experimental investigations which deal with the dynamics and statistics of coupled oscillators that do not have a common frequency.
In this letter, we deal with the phase locking level in an array of lasers that have no common frequency, and show that the distribution of the phase locking level is in good agreement with the Gumbel distribution function which describes the extreme value statistics from Gaussian processes [@Gumbel1; @Gumbel2]. Specifically, we investigate the phase locking level of an array of 25 coupled fiber lasers. Although each fiber laser support 100,000 eigenfrequencies, the probability to find a common frequency for all the lasers in the array is very small ($<10^{-5}$) [@no_more_shirakawa; @no_more_rothenberg; @no_more_shakir], so the lasers group in several clusters, each with its own frequency [@Galvanauskas16; @Moti25]. Due to thermal and acoustic fluctuations, the length of each fiber laser and its corresponding eigenfrequencies changes rapidly and randomly. As we show below, phase locking minimizes loss in the array, so mode competition will favor frequencies that maximize the size of the phase locked clusters at each moment [@VarditPRL]. Since the distribution of the phase locking level for different frequencies is Gaussian, the statistics of the maximum phase locking level is described by the Gumbel distribution function.
The experimental configuration that we used for measuring the phase locking level of an array of fiber lasers is presented in Fig. \[setup\] and described in detail in [@Moti25]. Briefly, each fiber laser was comprised of a Ytterbium doped fiber, a rear high reflection ($>99\%$) fiber Bragg grating (FBG), and a front low reflecting ($5\%$) FBG both with a $10nm$ bandwidth. Each laser was pumped through the rear FBG with a $975nm$ diode laser at $200mW$, and after the front FBG we attached a collimator to obtain a $0.4mm$ diameter beam. The collimators of all the 25 lasers were accurately aligned in a $5X5$ square array of parallel beams with parallelism better than $0.1mrad$. The separation between adjacent beams was $3.6mm$. A representative near-field intensity distribution, measured close to the output coupler when all 25 fiber lasers are operating, is presented at the lower inset of Fig. \[setup\]. We determined the length of each fiber laser by measuring the longitudinal mode beating frequency at their output by means of a fast photodetector connected to a RF spectrum analyzer. We found that the distribution of the lasers lengths is Gaussian with a mean value of $3m$ and a width of $0.5m$. The intensity of each fiber laser was about $100mW$ which is much above threshold but still low enough so nonlinear effects where negligible [@Corcoran].
![\[setup\]Experimental configuration for phase locking an array of fiber lasers and for determining their phase locking level. OC - output coupler, PR - partial reflector, BS - 50% beam splitter. Inset - near field intensity distribution when all 25 fiber lasers are operating.](systemnopic2.pdf){width="8cm"}
The coupling between the fiber lasers was achieved by means of four coupling mirrors denoted as $r_1$, $r_2$, $r_3$ and $r_4$ with reflectivity of $40\%$ for $r_1$ and $r_3$ and reflectivity of $100\%$ for $r_2$ and $r_4$. All the coupling mirrors were located close to the focal plane of a focusing lens with $500mm$ focal length, forming a self imaging cavity with the array. Since there was only enough space for one pair of mirrors within the Rayleigh range of the focussing lens, we inserted a $50\%$ beam splitter to obtain another focal plane where we placed another pair of mirrors. By controlling the orientations of the coupling mirrors we could realize a variety of connectivities for the fiber lasers in the array, and in our experiments we concentrated on the two-dimensional nearest neighbors connectivity [@Moti25]. Finally, we directed about $10\%$ of the light with a partially reflecting mirror (PR), towards an output coupler (OC) of 99% reflectivity. The OC was placed at a distance of 2f from the collimator array, and reflected part of the light from each laser back onto itself with the same delay as the light that is coupled from the other lasers [@michaTime].
We measured the phase locking level as a function of time for different number of lasers in the array. This was done by continuously detecting the far-field intensity distribution of the interference pattern of all the light from the array with a CCD camera, determining the maxima and minima intensities, and calculating the average fringe visibility along the x and y directions. The fringe visibility provides a direct measure for the phase locking level that ranges from 0 to 1. The correlation time of the phase locking level is shorter than $100msec$, so over a 10 hours period we acquired about $370,000$ uncorrelated measurements of the fringe visibility. Representative experimental results of the fringe visibility as a function of time for 10 seconds interval are presented in Fig. \[bareData\]. The insets show two typical far field intensity distributions - one with low fringe visibility and the other with high fringe visibility.
![\[bareData\] Typical experimental results of the phase locking level as a function of time over a 10 seconds interval. The phase locking level was determined from the far field intensity distribution of the output. Insets show typical far field intensity distributions - (a) low fringe visibility where the phase locking level is low; (b) high fringe visibility were the phase locking level is high.](BareData2.pdf){width="8cm"}
Using the 370,000 measurements, we determined the probability distribution of the fringe visibility and fitted it to the distribution of extreme values in Gaussian processes, namely the BHP distribution [@BHP]. A simplified form of BHP distribution is given by $$P(x)=e^{c \left( \frac{x-\alpha}{\beta}-e^{\frac{x-\alpha}{\beta}} \right)}, \label{Eq1}$$ where $\alpha$ denote the mean value, $\beta$ the width and $c$ the measure for correlations in the Gaussian process, with $c=1.58$ indicating highly correlated process and $c=1$ an uncorrelated process. After fitting our probability distribution to the BHP distribution, we determined that $c=1.03$, indicating that the Gaussian process in our case is uncorrelated. Consequently, the BHP distribution of Eq. (\[Eq1\]), reduces to the generalized Gumbel distribution [@Gumbel1]. Representative experimental probability distributions for an array of 25 fiber lasers and an array of 12 fiber lasers with fits to a generalized Gumbel distribution functions are presented in Fig. \[gumbel\]. The inset shows the corresponding results in a linear scale. As evident, there is a very good agreement between the experimental probability distributions of the phase locking level and the Gumbel distribution function for the array of 25 fiber lasers, but not as good for the array of 12 fiber lasers.
![\[gumbel\] Probability distributions of the phase locking level for two arrays of fiber lasers. Asterisks (blue) - experimental probability distribution of the measured phase locking level of an array with 25 fiber lasers; circles (red) - experimental probability distribution of the measured phase locking level of an array with 12 fiber lasers; solid and dashed curves - Gumbel distributions fitted for both cases. Inset shows the probability distributions with the Gumbel fits in linear scale.](Fig3_Exp_res_gumbel2.pdf){width="8cm"}
First, we explain qualitatively the connection between the phase locking level and the Gumbel distribution. The number of lasers in each phase locked cluster changes rapidly and randomly, but always while maximizing the phase locking level in the array. This maximum phase locking level occurs at a specific frequency out of all the possible frequencies within the FBG bandwidth where the lasers losses are minimal. By considering the spectral response of the coupled lasers it was shown, that the distribution of the phase locking levels for all the available frequencies is an uncorrelated Gaussian distribution [@motiCommonLM; @no_more_shakir]. Since the distributions of maxima of uncorrelated Gaussian processes are described by the Gumbel distribution functions [@Gumbel1; @Gumbel2], the probability distribution of the phase locking level should be the same.
As evident from the results of Fig. \[gumbel\], for the array of 25 fiber lasers the average phase locking level is 0.28. Accordingly, rare events with more than triple the average phase locking level can occur, so there is good agreement between the experimental probability distribution and the Gumbel distribution for phase locking levels up to 0.85. This is not the case for smaller arrays of fiber lasers where the average phase locking level is higher and the effect of clipping of the distribution at unity phase locking level is significant. To illustrate this effect, we measured the phase locking level for smaller arrays, with 20, 16 and 12 fiber lasers, and found that as the number of fiber lasers in the array decreases, the fit for a Gumbel distribution is less exact. This is clearly evident for the array of 12 fiber lasers where the average phase locking level is 0.57 and the agreement to the Gumbel distribution is only good for phase locking level below 0.7.
Next, we present a simple quantitative model that relates the phase locking level to an extreme value in the spectral response of the array of coupled lasers. We start by assuming no gain and determining the spectral response of each laser cavity when we replaced all components that couple light into the laser cavity with an effective front mirror. The reflectivity of this effective mirror depends on the frequency [@motiCommonLM]. For example, the effective reflectivity of the $i$’th laser in a one-dimensional array of N coupled lasers, when each of which is coupled to its two nearest neighbors, is $$R_i=\frac{1}{1-r(1-2 \kappa)+\frac{\kappa^2}{1-R_{i-1}^{(u)} e^{2 \imath k l_{i-1}}}+\frac{\kappa^2}{1-R_{i+1}^{(d)} e^{2 \imath k l_{i+1}}}} \label{Eq2}$$ where $\kappa$ denote the coupling to the two neighbors, $r$ the reflectivity of the output coupler, $l_i$ the length of the $i$’th laser, $k$ the propagation vector, and $R_i^{(u)}$ and $R_i^{(d)}$ the effective reflectivities from all the lasers above and below the $i$’th laser, given by $$R_i^{(u)}=\frac{1}{1-r(1-2 \kappa)+\frac{\kappa^2}{1-R_{i-1}^{(u)} e^{2 \imath k l_{i-1}}}},$$ and $$R_i^{(d)}=\frac{1}{1-r(1-2 \kappa)+\frac{\kappa^2}{1-R_{i+1}^{(d)} e^{2 \imath k l_{i+1}}}}, \label{Eq4}$$ and $R_1^{(u)}=R_N^{(d)}=0$. Now, introducing the laser gain together with a mode competition results in amplifications only at frequencies with high effective reflectivity.
We solved Eqs. (\[Eq2\])-(\[Eq4\]) explicitly for an array of 25 fiber lasers and obtained the effective reflectivity of each laser as a function of the frequency. By counting the number of lasers with effective reflectivity above a certain threshold, we obtained the number of lasers in the main cluster as a function of frequency which provided a direct measure for the phase locking level of the array as a function of frequency [@BHP]. Representative results are shown in Fig. \[cluster\_size\]. It shows the probability distribution of the size of the main cluster together with a Gaussian fit for the tail of the distribution. The inset shows the size of the main cluster as a function of the frequency. As evident, the calculated probability distribution above the mean size of the main cluster is Gaussian. Then we selected the maximum size of the main cluster within the bandwidth of the FBG ($10nm$) and determined the resulting phase locking level as the ratio of the size of the main cluster over the size of the array.
![\[cluster\_size\] Calculated probability distribution of the size of the main cluster together with a Gaussian fit for clusters larger than the mean size. Inset shows representative calculated results of the main cluster size as a function of the laser frequency.](cluster_size_distribution3.pdf){width="8cm"}
We repeated these calculations for 50,000 different arrays each with a different random fibers lengths. The fiber length of the $i$’th laser in each array was chosen to be $l_i+\Delta l_i$, where $l_i$ is the measured length of the $i$’th fiber and $\Delta l_i$ a random length taken from a normal distribution with 0 mean and $10\mu m$ width. The results are presented in Fig. \[num\_eff\_ref\]. It shows the probability distribution of the calculated phase locking level together with a fit to a Gumbel distribution. As evident, there is a very good agreement between the distribution of the calculated results and the Gumbel distribution, indicating that the effective reflectivity is suitable for modeling arrays of coupled lasers, and that the underlaying Gaussian process is the number of lasers in the main cluster.
![\[num\_eff\_ref\] Calculated probability distribution of the phase locking level from 50,000 different arrays of fiber lasers as well as a fit to a Gumbel distribution.](DistOfClusterSize3.pdf){width="8cm"}
To conclude, we measured the probability distributions of the phase locking level for arrays of fiber lasers, and showed that they fit to Gumbel distributions. We also presented that the underlying Gaussian process is the number of lasers that have a common frequency. Our results can now be exploited to predict for arrays with an arbitrary number of fiber lasers what will be the probability to obtain a specific phase locking level. Finally, Finally, by operating the fiber lasers close to threshold we observe strong fluctuations also in the total power of the array which can be related to the statistics of extreme eigen values of random matrices [@TWM].
This research was supported by the Israeli Ministry of Science and Technology and by the USA-Israel Binational Science Foundation.
[99]{}
I. Z. Kiss, Y. Zhai, and J. L. Hudson, Science 296, 1676 (2002). K. Wiesenfeld, Physica (Amsterdam) 222B, 315 (1996). I. Kanter, N. Gross,E. Klein, E. Kopelowitz, P. Yoskovits, L. Khaykovich, W. Kinzel, and M. Rosenbluh, Phys. Rev. Lett. 98, 154101 (2007). G. D. VanWiggeren and R. Roy, Science 279, 1198 (1998). G.V. Osipov, B. Hu, C. Zhou, M.V. Ivanchenko, and J. Kurths, Phys. Rev. Lett. 91, 024101 (2003). S. H. Strogatz, , 268 (2001). E. J. Gumbel, [*Statistics of Extremes.*]{} Mineola, NY: Dover, 2004. N. Johnson, S. Kotz, and N. Balakrishnan, [*Continuous Univariate Distributions,*]{} Vol. 2, 2nd ed. New York: Wiley, 1995. E. J. Bochove, S. A. Shakir, “Analysis of a Spatial-Filtering Passive Fiber Laser Beam Combining System” [*IEEE J. of Selected Topics in Quantum Elec.*]{} [**15**]{}, 320 (2009). A. Shirakawa, K. Matsuo, and K. Ueda, “Fiber Laser Coherent Array for Power Scaling of Single-Mode Fiber Laser”, , [**5662**]{}, 482, (2004). J. E. Rothenberg, “Advances in Fiber Laser Beam Combination” in , 687315 (2008). M. Fridman, M. Nixon, N. Davidson, and A. A. Friesem “Passive phase locking of 25 fiber lasers” [*Opt. Lett.*]{} [**35,**]{} 1434 (2010). W. Chang, T. Wu, H. G. Winful, and A. Galvanauskas, “Array size scalability of passively coherently phased fiber laser arrays,” [*Opt. Express*]{} [**18,**]{} 9634, (2010). V. Eckhouse, M. Fridman, N. Davidson, and A. A. Friesem, *Phys. Rev. Lett.* [**100**]{}, 024102, (2008). C. J. Corcoran and F. Durville “Passive Phasing in a Coherent Laser Array” [*IEEE J. Quantum Electron.*]{} [**15,**]{} 294 (2009). M. Nixon, M. Fridman, E. Ronen, A. A. Friesem, and N. Davidson, , , 1864 (2009). S. T. Bramwell, K. Christensen, J. Y. Fortin, P. C. W. Holdsworth, H. J. Jensen, S. Lise, J. Lopez, M. Nicodemi, J. F. Pinton, and M. Sellitto,“Universal Fluctuations in Correlated Systems” [*Phys. Rev. Lett.*]{}, [**84**]{}, 3744, (2000). M. Fridman, M. Nixon, E. Ronen, A. A. Friesem, and N. Davidson “Phase locking of two coupled lasers with many longitudinal modes” , 526 (2010). M. Fridman, R. Pugatch, M. Nixon, A. A. Friesem, and N. Davidson, “Measuring maximal eigenvalue distribution of Wishart random matrices with coupled lasers” [*Submitted to PRL*]{}.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'Stochastic Gradient Descent (SGD) is the most popular algorithm for training deep neural networks (DNNs). As larger networks and datasets cause longer training times, training on distributed systems is common and distributed SGD variants, mainly asynchronous and synchronous SGD, are widely used. Asynchronous SGD is communication efficient but suffers from accuracy degradation due to delayed parameter updating. Synchronous SGD becomes communication intensive when the number of nodes increases regardless of its advantage. To address these issues, we introduce Layered SGD (LSGD), a new decentralized synchronous SGD algorithm. LSGD partitions computing resources into subgroups that each contain a communication layer (communicator) and a computation layer (worker). Each subgroup has centralized communication for parameter updates while communication between subgroups is handled by communicators. As a result, communication time is overlapped with I/O latency of workers. The efficiency of the algorithm is tested by training a deep network on the ImageNet classification task.'
author:
- Kwangmin Yu
- Thomas Flynn
- Shinjae Yoo
- 'Nicholas D’Imperio'
bibliography:
- 'references.bib'
title: 'Layered SGD: A Decentralized and Synchronous SGD Algorithm for Scalable Deep Neural Network Training'
---
<ccs2012> <concept> <concept\_id>10010147.10010919.10010172</concept\_id> <concept\_desc>Computing methodologies Distributed algorithms</concept\_desc> <concept\_significance>500</concept\_significance> </concept> <concept> <concept\_id>10010147.10010257.10010321</concept\_id> <concept\_desc>Computing methodologies Machine learning algorithms</concept\_desc> <concept\_significance>500</concept\_significance> </concept> </ccs2012>
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
address: |
Department of Biomedical Engineering, Duke University, Durham, NC 27708\
[email protected]\
[email protected]
author:
- 'Kevin C. Zhou and Roarke Horstmeyer'
bibliography:
- 'sample.bib'
title: Diffraction tomography with a deep image prior
---
We present a tomographic imaging technique, termed Deep Prior Diffraction Tomography (DP-DT), to reconstruct the 3D refractive index (RI) of thick biological samples at high resolution from a sequence of low-resolution images collected under angularly varying illumination. DP-DT processes the multi-angle data using a phase retrieval algorithm that is extended by a deep image prior (DIP), which reparameterizes the 3D sample reconstruction with an untrained, deep generative 3D convolutional neural network (CNN). We show that DP-DT effectively addresses the missing cone problem, which otherwise degrades the resolution and quality of standard 3D reconstruction algorithms. As DP-DT does not require pre-captured data or pre-training, it is not biased towards any particular dataset. Hence, it is a general technique that can be applied to a wide variety of 3D samples, including scenarios in which large datasets for supervised training would be infeasible or expensive. We applied DP-DT to obtain 3D RI maps of bead phantoms and complex biological specimens, both in simulation and experiment, and show that DP-DT produces higher-quality results than standard regularization techniques. We further demonstrate the generality of DP-DT, using two different scattering models, the first Born and multi-slice models. Our results point to the potential benefits of DP-DT for other 3D imaging modalities, including X-ray computed tomography, magnetic resonance imaging, and electron microscopy.
Introduction
============
There are a variety of microscopes that can obtain high-resolution images in three dimensions, including scanning confocal microscopes, two-photon microscopes, and light sheet microscopes, to name a few [@mertz2019strategies]. Most of these methods are geared towards incoherent fluorescent imaging, and cannot produce a quantitative estimation of the 3D refractive index (RI) distributions of thick biological samples. Quantitative RI is useful for a number of reasons - it does not require labeling with fluorescent proteins or dyes, it can directly yield useful measures of cell mass and protein concentration, and it can provide useful information about how light is scattered in thick samples for subsequent imaging system correction, for example [@park2018quantitative].
Currently, the primary technique for quantitative 3D RI measurement is diffraction tomography (DT) [@Wolf_1969; @Sung_2009; @Fiolka_2009; @Lauer_2002; @chowdhury2017refractive]. The first implementations of DT relied on holography to measure the complex field scattered from an object under illumination from a variety of angles. Since DT is a phase-sensitive technique, it requires a highly coherent beam with interferometric stability and some sort of angular scanning mechanism to steer the incident beam through a range of angles. These requirements make its practical implementation relatively complicated and challenging. Several recent techniques have demonstrated DT without a reference beam, instead using intensity-only images and a suitable phase retrieval algorithm for 3D sample reconstruction [@Horstmeyer_2016; @chowdhury2019high; @tian20153d; @li2019high; @ling2018high; @pham2018versatile; @jiang2016inverse]. These methods, which effectively extend Fourier ptychography (FP) techniques into the third dimension [@Zheng_2013], remove the need for a highly coherent beam and interferometric stability. Instead, they rely on a programmable LED array to provide angularly-varying illumination, which leads to a simple and compact device that requires no moving parts [@Konda2019]. While a direct extension of FP into 3D relies on the first Born approximation, other related methods have also used the multi-slice (MS) model (also known as the beam propagation method, BPM) which can incorporate the effects of multiple-scattering [@kamilov2015learning; @kamilov2016optical; @tian20153d; @chowdhury2019high; @jiang2016inverse; @pham2018versatile].
Since all of the above techniques illuminate a stationary sample from a finite angular extent, and use a single objective lens to collect the scattered light, they all suffer from the missing cone problem (also referred to as the missing wedge) [@Tam_1981]. In both phase-sensitive and intensity-only diffraction tomography, the missing cone manifests itself as a bandlimited transfer function that is zero within a cone surrounding the $k_z$ axis. This limited transfer function produces axial artifacts and underestimates the RI, thus presenting challenges to the accurate 3D reconstruction of thick samples at high resolution.
![Summary of Deep Prior Diffraction Tomography (DP-DT). Microscope captures variably-illuminated image $k$-stack. Standard tomographic reconstruction methods suffer from artifacts caused by a missing cone in $k$-space. In this work, we propose the use of a deep image prior to help account for missing cone artifacts to improve 3D image reconstruction. \[teaser\]](figures/teaser.pdf){width="0.8\columnwidth"}
As we will show, the missing cone problem becomes significantly worse when using low numerical aperture (NA) objective lenses, and thus high-NA objectives are desirable for DT. On the other hand, low-NA objective lenses typically exhibit larger space-bandwidth products (SBPs) than high-NA lenses, since they can image over larger fields of view (FOVs). They also exhibit greater depths of field (and therefore high signal-to-noise ratio (SNR) over a larger axial range), longer working distances, and often fewer aberrations. Low-NA lenses are thus highly desirable for DT, if we hope to achieve multi-gigavoxel 3D reconstructions in the future. There is thus a tradeoff between the NA of the objective and two desirable properties: multi-gigavoxel SBPs and reduced missing cone artifacts. While rotating the sample [@M_ller_2015], imaging from multiple angles [@konda2018parallelized], or at multiple focal planes [@he2018design] can potentially help fill in missing spatial frequencies, these experimental modifications significantly complicate a standard microscope setup. Thus, unsurprisingly, there is extensive prior work on computational means to fill in the missing cone, such as using positivity constraints [@choi2007tomographic; @Lim_2015; @sung2012stain; @sung2011deterministic] and imposing penalties on the spatial image gradient, such as total variation (TV) regularization [@Lim_2015; @sung2011deterministic; @sung2012stain; @krauze2016generalized; @delaney1998globally; @goris2012electron]. More recently, data-driven deep-learning-based approaches have also been proposed to fill in the analogous missing wedge in X-ray computed tomography (CT) [@anirudh2018lose; @zhang2016image; @ding2019joint]. It has also been shown that accounting for multiple scattering can mitigate the effects of the missing cone (however, this depends on the sample itself exhibiting multiple scattering) [@kamilov2015learning; @lim2018learning]. Regardless of the forward model employed (e.g., ray-based [@choi2007tomographic], Born[@Horstmeyer_2016], Rytov [@Sung_2009], MS [@tian20153d; @kamilov2015learning; @kamilov2016optical]), or whether or not phase is detected, or the imaging modality itself (e.g., CT, magnetic resonance imaging (MRI), electron microscopy (EM), or standard fluorescence imaging), the missing cone problem is a ubiquitous one and thus addressing it would have far-reaching implications.
In this work, we propose a new approach to address the missing cone problem, termed Deep Prior Diffraction Tomography (DP-DT), which uses a deep image prior (DIP [@ulyanov2018deep]) as an *untrained* deep 3D convolutional neural network (CNN) to generate 3D object reconstructions (Fig. \[teaser\]). Unlike other recent works that propose to use supervised deep learning to aid in computational image reconstruction problems [@dave2018solving; @lucas2018using; @jo2018quantitative; @mccann2017convolutional; @barbastathis2019use], including a number of works that rely on multi-angle illumination [@goy2018high; @nguyen2018computational; @jin2017deep; @anirudh2018lose; @zhang2016image; @ding2019joint; @kellman2019data], the technique proposed here does [*not*]{} require any pre-training or dataset-specific assumptions. Instead, during iterative object reconstruction, DP-DT simply performs its optimization updates with respect to the parameters of a CNN, as opposed to directly updating the object voxels. The authors of the original DIP paper [@ulyanov2018deep] found that the structure of CNNs alone has an inherent bias towards natural images. We thus hypothesized that the artifacts induced by the missing cone problem in the spatial domain would be outside of the domain of natural images that are representable by a DIP.
Here, we empirically confirm this hypothesis by applying DP-DT to reconstruct 3D images of beads and a biological sample, from their associated variably-illuminated intensity-only images, with higher fidelity than alternative regularization techniques like TV regularization and positivity constraints. We test DP-DT under several different conditions and light propagation models, including the first Born and MS models, and find that the DIP consistently improves 3D reconstruction quality. We also suggest that DP-DT is more general than alternative regularization strategies, as the assumptions of TV regularization and positivity are not always valid, and when they are, they can easily be added into the DP-DT framework. Furthermore, DP-DT does not rely on preexisting datasets, which may be difficult to acquire to learn representations or extract features for object reconstruction. DP-DT thus does not inherit any generalization errors or biases when it is applied to a new types of sample, instead offering a general strategy to improve 3D reconstruction quality.
![The missing cone problem and its effect on the 3D SBP. (a) Simulations of the effects of transfer functions containing missing cones for a variety of illumination (vertical) and collection (horizontal) NAs. The odd columns are the $k$-space supports of the transfer functions and the even columns are the spatial domain representations of 0.8-$\mu m$-diameter bead ($n$=1.35) immersed in water ($n$=1.33), which have been filtered by these transfer functions. (b) 3D SBPs in gigavoxels; 2D fields of view, collection NAs, and magnifications taken from [@Zheng14]; arbitrarily assumes a 20-$\mu m$ axial range; please note the semilog scale. [\[missing\_cone\]]{}](figures/Fig2-SVP_horizontal.pdf){width="0.8\columnwidth"}
Missing cone problem
====================
To model 3D image formation, we will begin with the first Born approximation, which offers a clear description of the missing cone problem. In general, we can represent the 3D scattering potential of a sample of interest with $$V(\mathbf{r})=\frac{k^2}{4\pi}\left(n(\textbf{r})^2-n_0^2\right),
\label{V}$$ where $\textbf{r}=(x,y,z)$ represents the 3D spatial coordinates of the sample, $k=2\pi/\lambda$ is the vacuum wavenumber, $n_0$ is the surrounding medium’s RI, $\lambda$ is the wavelength of light, and $n(\textbf{r})$ is the sample’s RI distribution, which is the unknown quantity of interest that we aim to reconstruct. We note that $n(\textbf{r})$ is complex-valued, where $\mathrm{Re}\{n(\textbf{r})\}$ is associated with the index of refraction and $\mathrm{Im}\{n(\textbf{r})\}$ is associated with absorption.
If we take the 3D Fourier transform of $V(\textbf{r})$, we arrive at the sample’s scattering potential spectrum: $$\widetilde{V}(\textbf{k}) = \mathcal{F}_{3D}\left(V(\textbf{r})\right),
\label{V_hat}$$ where $\textbf{k}=(k_x, k_y, k_z)$ is the 3D wavevector. Under the first Born approximation, a DT system can only measure a finite range of sample wavevectors, bounded by the angular span of incident and observable light [@Haeberle]. We can represent this limited range of spatial frequencies with a transfer function $H(\textbf{k})$, which defines the observable information at the imaging plane as $\mathcal{F}^{-1}_{3D}(\widetilde{V}(\textbf{k})H(\textbf{k}))$. In practice, $H(\textbf{k})$ can be synthesized by taking a superposition of partial spherical shells (i.e., Ewald spheres), with radii set by the wavelength of light, radial extent set by the imaging NA, and whose positions shift along an arc defined by the illumination k-vector (for mathematical details, see Appendix B). This transfer function’s extent in $k$-space shrinks with smaller imaging NAs, even if a large illumination NA is used, resulting in significant blurring in the spatial domain.
To illustrate this point, we modeled the DT transfer function with variable illumination and imaging NAs in Fig. \[missing\_cone\](a). Next to each DT transfer function, we also simulated its effect on a 0.8-$\mu m$-diameter bead. In these plots, we show 2D $xz$ and $k_xk_z$ slices of the 3D transfer functions. Here, we can see that for a fixed imaging NA, increasing the illumination NA only modestly reduces the axial blurring and RI underestimation induced by the missing cone. Unlike 2D synthetic aperture techniques, such as FP, 3D diffraction tomography methods are thus highly dependent on large imaging NAs for high-fidelity tomographic reconstructions.
As mentioned above, the use of a high imaging NA unfortunately leads to a much smaller lateral imaging FOV. To explore this point in detail, we numerically computed the 3D SBP for each modeled transfer function, which is equivalent to the total number of resolvable voxels in a DT system for a given imaging lens and illumination configuration (Fig. \[missing\_cone\](b)). In particular, we defined the 3D SBP as the product of the 3D $k$-space volume and the 3D spatial reconstruction volume, basing the SBP computation on the FOV of standard microscope objective lenses [@Zheng14]. Here, we assumed a fixed axial imaging range of 20 $\mu$m. We note that although in theory, even under the first Born approximation, the axial range of the reconstruction volume can be unbounded, in practice the axial range is often limited by factors such as the SNR. In any case, selecting a different axial range only re-scales Fig. \[missing\_cone\](b)’s y-axis and does not affect the relative comparison between different objective lenses. From this simple analysis, it is clear that lower-NA objectives are desirable for high-throughput tomographic imaging, as they yield significantly larger 3D SBPs. However, such low-NA lenses generate a large missing cone, thus pointing to a critical limitation that must be addressed before it is possible to rapidly acquire high-resolution 3D images.
![[The physical forward model and reconstruction algorithm for DP-DT, under the first Born or Rytov approximation. [\[forward\_figure\]]{}]{}](figures/Fig3_forwardmodel.pdf){width=".8\columnwidth"}
Deep prior diffraction tomography (DP-DT) {#igw}
=========================================
To address the above challenge of creating a high-fidelity 3D image reconstruction in the presence of a large missing cone, we propose DP-DT, a forward-model-agnostic framework that merges a deep image prior (DIP) into an iterative tomographic reconstruction process. To demonstrate its versatility, we use DP-DT to improve the quality of 3D sample reconstruction under the assumptions of both the first Born [@Horstmeyer_2016] and MS approximations [@tian20153d; @kamilov2015learning; @kamilov2016optical]. While the rest of this section assumes that we are reconstructing with phaseless measurements, we note that the DIP pipeline can easily be applied with different assumed forward models (e.g., [@liu2017seagle; @lim2019high; @kamilov2016recursive]), and can be also applied to phase-sensitive measurements from traditional DT setups, which we include examples of in Appendix C.
Inverse problem formulation
---------------------------
Fig. \[forward\_figure\] shows a high-level summary of the forward image formation process and the inverse problem formulation, with mathematical details presented in Appendix B. In particular, let $S_{model}[\cdot, \cdot, \cdot]$ be the 3D reconstruction target for a particular $model$. For the first Born and Rytov approximations, $S_{Born}$ and $S_{Rytov}$ are a discretized, complex-valued 3D scattering potential spectrum tensor, $\widetilde{V}[\cdot, \cdot, \cdot]$ (Eq. \[V\_hat\]), while for the MS approximation, $S_{MS}$ is a complex-valued 3D tensor, $\delta_{obj}[\cdot, \cdot, \cdot]$ that describes the 3D RI distribution of the sample relative to the medium RI. Next, let the forward predictions for the multi-angle 2D images based on these reconstruction targets be given by $I_{pred}^{model}[\cdot, \cdot, p]$, where $p$ indexes the $p^{th}$ image under illumination from the $p^{th}$ LED and $model$ specifies the employed forward model. Finally, let $I_{data}[\cdot, \cdot, p]$ be a 3D tensor of experimental intensity measurements (i.e., an LED image stack).
The error metric we seek to minimize is the mean square error loss, $$\label{mse}
E(S_{model}, \phi_{pupil},u_0)=\frac{1}{M}\sum_{i,j,p}\left(\sqrt{I_{pred}^{model}[i,j,p]}-\sqrt{I_{data}[i,j,p]}\right)^2,$$ where $\phi_{pupil}$ and $u_0$ are the optionally optimizable pupil phase and per-LED input field amplitudes (see Appendix B), and $M$ is the total number of summed elements. We use the amplitudes rather than the intensities in the loss because it was previously found that the former is more robust to noise and experimental deviations [@Yeh_2015]. We may also regularize the reconstruction target to form our final loss function, $$L(S_{model}, \phi_{pupil},u_0)= E(S_{model}, \phi_{pupil},u_0) + \mathrm{Reg}(S_{model}).
\label{reg_eq}$$ We discuss the specific regularization techniques in the next two sections. Thus, the optimization problem for 3D image reconstruction is $$\operatorname*{arg\,min}_{S_{model},\phi_{pupil},u_0} L(S_{model},\phi_{pupil},u_0).
\label{argmin}$$ For the first Born and Rytov approximations, once $S_{model}=\widetilde{V}$ is optimized, we can use Eq. \[V\] to infer the complex 3D RI distribution.
While we have described intensity-based DT, which is afforded by our LED array setup on a standard microscope, it is straightforward to extend our framework to describe holographic DT setups that also measure phase. The equations to achieve this are presented in Appendix C.
Deep image prior (DIP)
----------------------
Directly minimizing $L$ can be problematic, given the effects of the missing cone. The DIP is a recently-presented, data-independent method to assist with a large variety of inverse image reconstruction problems without supervised training [@ulyanov2018deep; @Liu_2019; @heckel2018deep; @Gong_2018; @van2018compressed; @mataev2019deepred]. It is an untrained regularization technique that reparameterizes the reconstruction target in the spatial domain as the output of a deep generative CNN that uses pseudorandom noise as input (see Fig. \[DIP\] in Appendix A for the architecture we used). After initializing the CNN with pseudorandom noise, DIP optimization then proceeds to update CNN weights to minimize loss, as opposed to directly optimizing the reconstruction target ($S_{model}$). Here, we hypothesize that the DIP’s resistance to unnatural images extends to the third dimension and can help eliminate missing cone artifacts in diffraction tomography.
Compared to the DIP, other recently proposed techniques based on deep neural networks for FP reconstruction [@shamshad2019deep; @Nguyen_2018; @Kappeler_2017; @icsil2019deep; @metzler2018prdeep] all rely on pretraining, and hence are inherently biased towards a particular set of examples found in the training data set. Furthermore, these techniques were primarily applied to 2D reconstruction. However, while the 2D FP inverse problem is well-posed if the LEDs are sufficiently densely packed (i.e., to obtain $>50\%$ overlap [@bunk2008influence]), the 3D FP inverse problem is always ill-posed due to the missing cone, no matter how densely packed the LEDs are. Hence, the 3D inverse problem presents a more significant challenge that could benefit more significantly from the DIP.
To incorporate the DIP reparameterization into our framework, we modify Eq. \[argmin\], for the first Born (or Rytov) and MS forward models, respectively:
$$\operatorname*{arg\,min}_{\theta,\phi_{pupil},u_0} L_{Born}(\widetilde{V} = \mathcal{F}_{3D}\left(G(\theta)\right),\phi_{pupil},u_0),
\label{DIP_final_born}$$
$$\operatorname*{arg\,min}_{\theta,\phi_{pupil},u_0} L_{MS}(\delta_{obj} = G(\theta),\phi_{pupil},u_0),
\label{DIP_final_MS}$$
where $G$ is deep generative CNN parameterized by $\theta$. Eqs. \[DIP\_final\_born\] and \[DIP\_final\_MS\] are the final optimization problems for DP-DT. Note that $G$ outputs a prediction for the 3D object in the spatial domain, and thus we take its Fourier transform to get $\widetilde{V}$ in Eq. \[DIP\_final\_born\]. However, regardless of the form of $S_{model}$, it is critical to note that the optimization now is performed with respect to $\theta$, which are the weights within the CNN.
Other regularization
--------------------
It is straightforward to also include other well-known regularization methods in Eq. \[reg\_eq\], in particular TV and positivity regularization, to which we compare our proposed approach. We used the isotropic TV regularization, given by $$R_{TV}(S(\mathbf{r}))=\sum_{\mathbf{r}}\sqrt{|\nabla_xS(\mathbf{r})|^2+|\nabla_yS(\mathbf{r})|^2+|\nabla_zS(\mathbf{r})|^2},$$ where spatial gradients are approximated by finite differences. TV regularization promotes piecewise smoothness by encouraging sparsity in reconstruction gradients.
The positivity constraint that we test is applied to the real part of the RI under the assumption that the sample index does not fall below that of the immersion medium, noting in Fig. \[missing\_cone\] that one manifestation of the missing cone problem is negative index value artifacts: $$R_+(n(\mathbf{r}))=\sum_{\mathbf{r}}\text{min}(\text{Re}\{n(\mathbf{r})\}-n_0, 0)^2.$$ This expression returns 0 when the real part of the RI is above the immersion RI, and a quadratic penalty otherwise.
Note that both of these regularization terms are differentiable almost everywhere and thus are suitable for gradient-based optimizers. The modified loss function to be minimized is thus $$L=E+\lambda_{TV}R_{TV}+\lambda_+R_+,$$ where $\lambda_{TV}$ and $\lambda_+$ are regularization coefficients tuning the respective relative contributions. Unless otherwise specified, the DIP-based reconstructions do not include these extra positivity or TV regularization terms.
Results
=======
Setup for experiments and simulations
-------------------------------------
In this section, we demonstrate the effectiveness of DP-DT in both simulation and experiment. Our experimental setup (previously described in Ref. [@Horstmeyer_2016]) consists of a standard microscope equipped with an infinity-corrected 20$\times$ objective lens (NA=0.4, Olympus MPLN), an 8-bit camera with 1920$\times$1456 4.54-$\mu m$ pixels (Prosilica GX 1920), and a 31$\times$31 LED array as the illumination source (SMD3528, center wavelength = 632 nm, 4-mm LED pitch), positioned below the sample to give an illumination NA of approximately 0.4. In our simulations, we used the same setup, but varied the distance between the LED board and the sample to tune the illumination NA, and varied the diameter of the aperture function ($A[\cdot,\cdot]$, see Appendix B) to tune the imaging NA. For experimental results, we ignored the dark-field images from the corners of the LED array due to a low SNR, resulting in a total of 641 multi-LED images. However, for simulations, we used all 31$\times$31 LEDs in a centered square grid, computationally increasing the exposure (or equivalently, the illumination intensity) for the dark-field LEDs to boost the detected image SNR. To make simulations more realistic, we added Poisson-distributed noise to the forward intensity predictions, assuming a pixel well depth of 50,000 photoelectrons, and discretized the result into 8 bits. To ensure that the regularizers are only accounting for the missing cone, we simulated samples that approximately followed the first Born approximation and used this for both simulation and reconstruction. We also note that jointly optimizing the pupil function did not make much of a difference in terms of the reconstruction quality, and we thus did not end up doing so for either the simulated or experimental results.
We specified the forward model and performed gradient descent using the Adam optimizer [@kingma2014adam] in TensorFlow [@abadi2016tensorflow]. The code will be available at deepimaging.io. To accelerate optimization, we used a NVIDIA Tesla T4 GPU on the Google Cloud Platform. By default, we use the entire dataset per optimization step if it fit within the GPU’s 16 GB of memory. Otherwise, we split the dataset into roughly evenly-sized batches along the LED dimension such that the batches each fit in memory. After every full pass through the dataset (i.e., an epoch), we reshuffled the data along the LED dimension before splitting into a new set of batches. Optimization times depend on the FOV, the batch size, whether DIP is used (DIP requires several times more iterations), and which scattering model is used (the MS model is more computationally expensive than the first Born model). For our experimental implementation, each iteration was on the order of seconds, resulting in optimization times on the order of minutes to an hour (100s to 1000s of iterations) without DIP, and several hours (10s of 1000s of iterations) with DIP.
![Comparison of 3D reconstruction quality for phaseless DT using several regularizers. (a) 1D traces through simulations of two beads spaced axially. Rows show different bead axial separations, while columns show different bead sizes (imaging NA = 0.4, illumination NA = 0.5). Each curve corresponds to a different regularization technique (i.e., none, DIP, positivity (+), and TV) and ground truth. Scale bar corresponds to the Nyquist period. (b) The RI RMSEs from the ground truths for each regularizer. Each of the four plots corresponds to a different bead size, and each curve corresponds to a different edge-to-edge bead separation, where $z_0=$ 0.75 $\mu m$.[]{data-label="bead_sim"}](figures/{bead1D_fullfig_0.5}.pdf)
![2D cross-sections of select simulated bead pairs. (a) The first row is the through-origin $k_xk_y$ cross-sections of the scattering potential spectra of the reconstructions containing all the bead pairs, under illumination NA = 0.4, imaging NA = 0.5. The second and third rows show through-center $xz$ cross sections at two different separations. (b) The same information as (a), but under imaging NA = 0.3. (c) The same information as (a), but under imaging NA = 0.1.[]{data-label="bead_sim2D"}](figures/bead2D_fullfig.pdf)
Bead simulation results
-----------------------
First, to test the axial resolution of DP-DT, we simulated 31$\times$31 intensity-only images under the first Born approximation from pairs of 3D beads ($n$=1.525, on a background of $n_0$=1.515) of various sizes and various axial spacings under illumination from a spatially coherent LED array. We tested multiple imaging NAs (0.1, 0.3, 0.5) and a fixed illumination NA of 0.4. We reconstructed the 3D RI map of the bead phantom using the following priors: none, DIP, positivity, TV ($\lambda_{TV}$=1e-8), and TV ($\lambda_{TV}$=1e-9). Fig. \[bead\_sim\] summarizes these results for a 0.5 imaging NA, showing 1D axial RI cross-sections through the center of the 3D bead reconstructions (Fig. \[bead\_sim\](a)), as well as the root mean square error (RMSE) of these 1D axial profiles with respect to the ground truth (Fig. \[bead\_sim\](b)). We can see that DP-DT performs as well as or better than the other regularizers. In particular, in all four columns of Fig.\[bead\_sim\](a), there are separations for which the dip in between the two beads is deeper for DP-DT than the other regularization strategies, and in almost all cases DP-DT has a smaller RMSE than the other regularized results (Fig. \[bead\_sim\](b)). Similar observations apply to the imaging NAs of 0.1 and 0.3 (Supplementary Figs. \[bead\_sim\_0\_1\] and \[bead\_sim\_0\_3\] in Appendix C). Furthermore, we obtained very similar results for imaging NAs of 0.1, 0.3, and 0.5 in the case of reconstructing from phase-sensitive measurements (Supplementary Figs. \[bead\_sim\_phase\_0\_1\], \[bead\_sim\_phase\_0\_3\], and \[bead\_sim\_phase\_0\_5\] in Appendix C).
Fig. \[bead\_sim2D\] shows select 2D $xz$ RI cross-sections for the largest particular bead size and the two closest separations, from which we can see that the gaps between the two beads for DP-DT is more pronounced and faithful to the ground truth, compared to the other regularization techniques. Note that even though this sample is ideally suited for a TV prior, as it is piecewise smooth, DP-DT still produces superior results. We obtained similar results for reconstructions from phase-sensitive measurements (see details in Appendix C). We emphasize that for DP-DT the network was not pretrained to prefer such reconstructions; rather, these results were innately preferred by the DIP’s CNN structure.
{width=".8\columnwidth"}
. \[EM\]
Biological simulation results
-----------------------------
It is expected that TV-regularized reconstructions would perform well for bead samples, because such samples contain regions of uniform RIs. Thus, we also simulated a more complicated and realistic biological sample, based upon 3D isotropic EM images of hippocampal cells [@Lucchi_2012], for which the smoothness imposed by TV regularization may not be as appropriate an assumption. To convert the 3D EM data into a ground-truth 3D RI map, we renormalized its voxelized measurements to extend from 1.515 (i.e., $n_0$) to 1.517. With this rescaled dataset as the 3D RI, we used the first Born approximation to simulate 2D multi-LED intensity images, using an imaging NA of 0.2 and illumination NA of 0.4. The reconstruction results, assuming the same priors as used for the bead simulation (none, DIP+TV 1e-10, positivity, TV 1e-8, and TV 1e-9) are shown in Fig. \[EM\]. Note that the DIP reconstruction contains a small amount of TV regularization, which we found produced better results than DIP alone [@Liu_2019]. While in the $xy$ dimensions (Fig. \[EM\], first row), the reconstructions look similar across the different priors, in the $xz$ dimensions (Fig. \[EM\], second row), TV regularization only axially blurs the reconstruction. Furthermore, positivity regularization does not offer much improvement over the non-regularized reconstruction. However, DP-DT is able to estimate a significant amount of information in the missing cone, which is also apparent when considering the $k_xk_z$ cross-sections of the scattering potential spectra (Fig. \[EM\], third row).
![Experimental 2-layer, 2-$\mu$m bead results, using the first-Born model. (a) Example raw LED images. (b) Comparison of performance of different regularizers (columns). The first two rows are $xy$ slices through the two bead layers, and the last two rows are two axial slices whose positions are indicated by the red and magenta lines in the upper left plot. Scale bars, 5 $\mu m$. (c) 1D traces through regions indicated in the first panel of the $yz$ slice row. The expected RI of the beads is 1.59.[]{data-label="2layer"}](figures/2layer_figure.pdf){width=".8\columnwidth"}
To quantify the comparisons of the different techniques, we computed the RMSE and the structural similarity (SSIM) index [@wang2004image], which are displayed at the bottom of Fig. \[EM\]. While DP-DT did not produce the best statistics, its reconstruction visually exhibits less blurring and fewer artifacts. We also found that DP-DT produces less biased estimates of RI compared to all the other techniques (Fig. \[EM\], last row). In particular, while the non-DIP approaches exhibit reconstructions that overestimate RI for low RI values, and underestimate RI for high RI values (as one would expect for spatial blurring), DP-DT produces unbiased estimates for low RI values and less biased results for high RI values (i.e., its flatter red histogram indicates a more consistent bias-variance tradeoff with respect to RI). This is a unique property of DIP, as most conventional regularization techniques trade off unbiasedness for lower variance in their estimations.
Experimental results, first Born approximation
----------------------------------------------
To experimentally test DP-DT with phaseless 3D imaging data, we first examined its imaging performance using two bead phantom samples. We first created a bead phantom consisting of two layers of 2-$\mu$m-diameter beads ($n$=1.59), separated by 3.9 $\mu$m and embedded in oil ($n_0$=1.515). We reconstructed the 3D RI map of this two-layer sample under the first Born approximation, using no prior, DIP, positivity, TV ($\lambda$=1e-7), and TV ($\lambda$=1e-8). The top two rows of Fig. \[2layer\](b) shows two $xy$ cross-sections at the two layers of interest. Without regularization, there is leakage of information between layers, preventing clean separation, due to the missing cone artifact. This is also apparent in the $yz$ and $xz$ slices (Fig. \[2layer\](b), bottom two rows). While all of the regularized reconstructions were effective to some extent in reducing this artifact, reconstructions with no regularization and with TV show severely underestimated bead RIs. DIP- and positivity-regularized reconstructions were less underestimated. RI underestimation is seen more clearly in Fig. \[2layer\](c), which show 1D traces in the $yz$ plane, among with DP-DT shows the least underestimation. The remaining RI underestimation may be explained by the small bead sizes, noting that there is also RI underestimation in the reconstructions of our simulated beads in the first column of Fig. \[bead\_sim\], even for DP-DT (also note that for the closest separations, DP-DT has the lowest degree of RI underestimation). Also, the beads in Fig. \[2layer\] are not axially resolved fully, consistent with the incomplete separation in the upper left panel of Fig. \[bead\_sim\]. Differences between our simulated and experimental results (i.e., better axial separation in simulation) may be attributed to imperfections in our experimental setup, most notably the challenge of establishing an exact correspondence between the estimated and true LED positions, and perhaps the assumed scattering model. As shown in section \[results\_MS\] below, DP-DT using the MS model both reduces the degree of RI underestimation and improves the separation between the two bead layers.
![Experimental 800-nm bead sample. (a) Example raw LED images. (b) Comparison of different regularizers, excluding positivity because the RI is below that of the medium. The first row is $xy$ cross-sections and the second row is $xz$ cross-sections indicated by the red line. The third row shows a close-up view of the first row, indicated by the purple box. Scale bars, 5 $\mu m$ (top two rows of (b)), 1 $\mu m$ (bottom row of (b)).[]{data-label="small_bead"}](figures/micron_bead_figure.pdf){width=".8\columnwidth"}
![Experimental stacked starfish embryo results. (a) Example raw LED images. (b) $xy$ cross-sections at various depths. (c) $xz$ cross-sections at positions indicated by the horizontal red lines in the upper left panel of (b). Scale bars, 10 $\mu m$.[]{data-label="starfish"}](figures/starfish_figure.pdf){width=".8\columnwidth"}
Our second experimental target consisted of a single layer of 800-nm-diameter beads with an unspecified RI below that of the embedding oil ($n_0$=1.515). We reconstructed the 3D RI map using the same priors as for the 2-layer bead sample (Fig. \[small\_bead\]), except for positivity, because the 800-nm bead samples exhibited RI values lower than that of the medium. Note the enhanced lateral resolution in all reconstructions with respect to the raw data, as expected via DT aperture synthesis. The more heavily TV-regularized result ($\lambda_{TV}$=1e-7), while exhibiting reduced axial missing cone artifacts, has a reduced lateral resolution compared to the other reconstructions (Fig. \[small\_bead\](b), bottom row). Furthermore, as with the 2-layer sample, we found that the RI difference estimate for the unregularized and the TV-regularized reconstructions were underestimated, unlike for the DP-DT result (Fig. \[small\_bead\]). Finally, we note that the DP-DT reconstruction contains more energy concentrated at the center of the beads along the axial dimension, demonstrating a successful counter to the effects of the missing cone.
As the final sample, we imaged fixed early-stage starfish embryo cells with an LED-outfitted microscope (same configuration as for the bead experiments, Fig. \[starfish\]). We used the same regularization settings as for the 800-nm bead sample (also without positivity regularization, as we were not certain that the RI was strictly greater than that of the medium). $xy$ and $xz$ slices of the reconstructions are shown in Fig. \[starfish\], respectively. Here, we can see that TV removes axial artifacts due to the missing cone problem, but at the cost of blurring features in the lateral and axial dimensions and even erasing many of the cellular features within each embryo. This is because the piecewise smoothness assumption of TV may not be appropriate for this biological sample with a relatively highly varying spatial RI profile. On the other hand, the DP-DT reconstruction not only has a higher axial resolution, but also produces cells with a rounder appearance in the axial direction, while the other reconstructions exhibit characteristic missing cone artifacts that cause the cells to taper in the axial direction. It is worth noting that DP-DT produces higher RI estimates, which is further evidence that DP-DT is filling in the missing cone, which would otherwise cause RI underestimation.
![Experimental 2-layer, 2-$\mu$m bead results, using the multi-slice model. (a) $xy$ slices at the two bead layers. (b) Axial slices indicated by the lines in the upper left panel of (a). (c) 1D axial traces through the dotted lines in (b) (expected RI=1.59). Scale bars, 5 $\mu m$.[]{data-label="2layer_MS"}](figures/2layer_MS.pdf){width=".8\columnwidth"}
Experimental results, multi-slice (MS) approximation {#results_MS}
----------------------------------------------------
We also tested DP-DT under the MS forward model with the experimental 2-layer, 2-$\mu$m bead phantom. Because the MS forward model is more computationally intensive, we used what we call “spatial patching”, whereby at each iteration we select a random, apodized spatial crop within the reconstruction over which to optimize (for more details, see Appendix B). We also note that for the cases of positivity, weaker TV (1e-8), and no regularization, we had to terminate the optimization early to prevent the reconstruction algorithm from diverging.
The results are shown in Fig. \[2layer\_MS\]. Without regularization, or with positivity or weak TV regularization, the axial resolution is very poor, with poor discrimination of the two layers. However, DP-DT here has superior axial resolution, even resolving the beads from the two layers, which was not possible with DP-DT with the first Born model. Furthermore, the RI values reconstructed using DP-DT exhibit the least underestimation, approaching the expected value of 1.59. These results may be attributed to the fact that the MS model is able to model multiple forward scattering events; however, note that only DP-DT attains this RI value and axial separation.
Discussion and conclusion
=========================
In summary, we have presented DP-DT, a flexible and general framework that augments existing 3D diffraction tomography techniques with a DIP, which we have shown to alleviate the effects of the missing cone problem. Specifically, we have applied DP-DT to two scattering models, the first Born and MS approximations, and demonstrated its effectiveness in simulated and experimental data with intensity measurements, as well as with simulated phase-sensitive measurements. DP-DT differs from other deep-learning-based approaches in that it does not require pre-training on and hence is not biased towards a pre-existing dataset. DP-DT can thus be applied in situations where it is expensive or otherwise infeasible to collect large datasets for supervised training. Instead, DP-DT relies on the inherent preference of CNN structures for “natural” images, a class of images which we have empirically shown to exclude images with missing cone artifacts. These results open the door to 3D DT with multi-gigavoxel-scale SBPs.
We used a single common architecture for all the reconstructions in this paper (see Appendix A), based on the recently reported encoder-decoder DIP architecture [@ulyanov2018deep], which we did not have to tune for specific samples. Note that the number of parameters in this architecture is fixed, regardless of the reconstruction size, because it is a fully convolutional network that adapts to the output size. Unlike with conventional priors, whose impact may be tuned by a coefficient, there is no obvious way to adjust the “degree of naturalness” conferred to the reconstruction by the DIP, except perhaps indirectly by the CNN’s architecture. Thus, future work may include exploring the impact of alternative architectures [@heckel2018deep].
Furthermore, we did not have to employ early stopping to avoid overfitting for DP-DT, unlike in the original paper [@ulyanov2018deep], which found that running the optimization for too many iterations resulted in recapitulation of image artifacts. We hypothesize that this may be because DP-DT indirectly inpaints in the Fourier domain, such that the missing information is not spatially localized. However, it is possible that the MS model may experience overfitting, as it is not explicitly a Fourier inpainting approach (outside of the weak-scattering limit, in which it coincides with the first Born model). We did not, however, run the DIP-regularized MS optimization loops long enough to observe such effects.
We found that during optimization of DIP-regularized reconstructions under the first-Born approximation, the optimization would sometimes diverge rapidly, effectively resetting the reconstruction. The authors of the original DIP paper also observed a similar phenomenon. To counteract this divergence in an automated fashion, we periodically checkpointed the parameters and monitored the ratio of the current loss versus the mean loss over the last few iterations. If the ratio exceeded a certain threshold, we would restore the parameters to the previous checkpoint and anneal the learning rate by a factor of 0.9. We note that we did not observe this phenomenon for DIP-regularized MS reconstructions, though it is unclear whether that was due to the scattering model or the fact that we used spatial patching.
Future work includes further investigation of DP-DT under the MS forward model, especially when scaling to larger FOVs. Due to memory constraints, we would have to use smaller LED batch sizes and smaller spatial patch sizes relative to the full FOV. The DP-DT reconstruction under the MS model also showed artifactually high RI values near the axial edges of the reconstruction volume. However, these are not of concern because they occurred outside of the object support. These high values are not seen in non-DIP reconstructions, perhaps because they were initialized at 0, while DIP does not default to 0 (otherwise, it wouldn’t fill in the missing cone). Other future work includes investigating whether the DIP can account for artifacts that arise from using a scattering model in situations where its assumptions are not met, or artifacts from experimental uncertainties. Finally, although our primary goal is to achieve high fidelity multi-gigavoxel-scale 3D image reconstructions, currently, the 3D FOVs that we reconstructed here are on the order of 10s of $\mu m$ in the lateral dimensions. To achieve such SBPs, we would thus need to reconstruct the full FOV afforded by low-NA objectives, which can be on the order of several mm. In other words, we are currently about two orders of magnitude away in terms of a full demonstration. The major challenge of scaling up to multi-gigavoxel imaging using DP-DT is computation time, whether we use spatial patching or just divide up the reconstruction into patches and reconstruct them sequentially. Thus, we will explore more memory-compact DIP architectures and will investigate the effect of batch size and density of illumination angles in order to make the reconstruction process more computationally tractable.
While we have so far demonstrated DP-DT for coherent diffraction tomography, utilizing two popular scattering models and various samples, we expect our results to be more generally applicable to other imaging modalities that exhibit artifacts due to the missing cone problem. This includes ill-posed problems [@van2018compressed] and other domain-specific problems (e.g., anisotropy, speckle, coherent ringing artifacts, noise, etc.) from the fields of 3D X-ray CT, MRI, EM and fluorescence microscopy. Furthermore, DP-DT may certainly be applied to other scattering models, such as higher-order Born approximations and other multiple scattering models that may exhibit unstable convergence during optimization that would invariably produce unnatural-looking reconstructions, against which DP-DT may safeguard. We already observed this optimization stabilizing phenomenon for the MS-model-based reconstructions, where without DIP or heavy TV regularization the reconstruction acquired extreme artifacts, similar to the unregularzied reconstructions in Fig. S3(e),(f) in Chowdhury et al. 2019 [@chowdhury2019high]. Finally, we also hope to apply DP-DT to setups that include reflection geometries [@roberts1985reflection; @zhou2019optical], which invariably contain gaps in $k$-space between disjoint high and low frequency bands. In conclusion, we are hopeful that DP-DT will open up options for using wider-FOV, lower-NA imaging lenses for 3D imaging without axial reconstruction artifacts, thus paving the way for multi-gigavoxel tomographic imaging in the future.
Funding {#funding .unnumbered}
=======
KCZ was supported by the National Science Foundation (DGF-1106401) and RH was supported by funding from the Erlangen Graduate School in Advanced Optical Technologies (SAOT) by the German Research Foundation (DFG) in the framework of the German excellence initiative.
Acknowledgments {#acknowledgments .unnumbered}
===============
The authors would like to thank the lab of Prof. Changhuei Yang for assistance with experimental data capture, as well as Ruobing Qian, Shiqi Xu, and Mykola Kadobianskyi for helpful comments during preparation of this manuscript.
Disclosures {#disclosures .unnumbered}
===========
The authors declare no conflicts of interest.
Appendix A: CNN architecture {#architecture .unnumbered}
============================
For our DIP’s convolutional neural network structure, we adopted a slightly modified version of the symmetric encoder-decoder architecture used in the original paper [@ulyanov2018deep], detailed in Fig. \[DIP\]. The input is sampled from a uniform random distribution between 0 and 0.1, and is fixed throughout optimization. The downsampling blocks used strided convolutions as the downsampling operation, while the upsampling blocks used nearest neighbor upsampling. Unlike the original paper, we did not use skip connections. The original network made heavy used of batch normalization [@ioffe2015batch] and leaky ReLU activations [@xu2015empirical]. To generate the real and imaginary components of $V$, we split the real-valued output of $G$ into two equally sized tensors and summed across the feature dimension. Since leaky ReLU has a preference for positive numbers, we used a linear activation in the final upsampling block if we do not necessarily expect the sample scattering potential to be strictly positive.
![[The DIP architecture used for the DP-DT reconstructions in this paper. [\[DIP\]]{} ]{}](figures/DIP_figure.pdf){width=".8\columnwidth"}
Appendix B: Forward models {#forward_models_detailed .unnumbered}
==========================
First Born and Rytov approximations {#born_rytov .unnumbered}
-----------------------------------
To provide a mathematical description of our forward model that is amenable to computational reconstruction, we assume discretized coordinates where appropriate to emphasize practical implementation. First, we define the following collection of wavevectors, $$\mathbf{k^{cap}}[i,j]=(k^{cap}_x[i,j], k^{cap}_y[i,j], k^{cap}_z[i,j]),$$ where the $[i,j]^{th}$ wavevector is of a partial spherical “cap,” which is a segment of a discretized sphere in $k$-space (i.e., an Ewald sphere) with radius $k$. We initially place the center of this spherical cap at the origin of $k$-space at $[i,j]=[0,0]$. The $[i,j]$ indexing emphasizes that the spherical cap, although defined in 3D $k$-space, is indexed on a 2D Cartesian grid when orthographically projected onto the $k_xk_y$ plane. The maximum lateral extent of the spherical cap is defined by the NA of the imaging lens such that $k_{xy}^{max}=kN\!A$, where $N\!A$ is the imaging numerical aperture. Next, let $$\mathbf{k^{ill}}[p]=(k^{ill}_x[p], k^{ill}_y[p], k^{ill}_z[p])$$ be the $p^{th}$ illumination wavevector corresponding to the position of the $p^{th}$ LED relative to the sample. Subtracting the illumination vector from the spherical cap coordinate vector places the spherical cap coordinate in the correct location in 3D $k$-space for the $p^{th}$ LED. Following the Fourier diffraction theorem [@Wolf_1969], the $[i,j]^{th}$ coordinate of the 2D discrete Fourier transform (DFT) of the field scattered off of the object from the $p^{th}$ illumination, measured at the detector plane, is provisionally, $$\widetilde{u}[i,j,p]=\frac{\widetilde{V}(\mathbf{k^{cap}}[i,j]-\mathbf{k^{ill}}[p])}{4\pi1ik_z^{cap}[i,j]}
\label{FDT}$$ where $(1i=\sqrt{-1})$ is the imaginary unit, and the DFT scaling constants and other proportionality constants are assumed to be properly handled. In short, the $p^{th}$ scattered field in $k$-space is defined by the scattering potential spectrum, $\widetilde{V}$, sampled along the $p^{th}$ offset spherical cap and properly scaled. The scattered field in the spatial domain for the $p^{th}$ illumination is thus given by, $$\label{forward}
u[\cdot,\cdot,p]=\mathcal{F}_2^{-1}(\widetilde{u}[\cdot,\cdot,p]A[\cdot,\cdot]exp(1i\phi_{pupil}[\cdot,\cdot]))$$ where $\mathcal{F}_2^{-1}$ is the 2D inverse DFT, $\phi_{pupil}[\cdot,\cdot]$ is an unknown pupil phase function that accounts for aberrations in the imaging lens (to be optimized later), and $A[\cdot,\cdot]$ is a circ function with radius $kN\!A$ that specifies the imaging lens bandpass.
Eq. \[forward\] represents the scattered field from the sample at the microscope’s image plane. As we are primarily concerned with standard microscopes that only record total intensity, we must also consider the unscattered field, which we model as a plane wave: $$\label{unscattered}
u_{back}[\cdot,\cdot,p]=u_0[p]exp(1i(k_x^{ill}[p]x[\cdot,\cdot]+k_y^{ill}[p]y[\cdot,\cdot]))$$ where $u_0[p]$ is the optimizable amplitude of the $p^{th}$ illumination wavevector and $x[\cdot,\cdot]$ and $y[\cdot,\cdot]$ are 2D meshgrids sampling the lateral plane, consistent with the lateral spatial sampling of the scattered field (Eq. \[forward\]). The third factor of Eq. \[unscattered\] accounts for the direction of the $p^{th}$ illumination. Note that we do not include an optimizable phase of the unscattered field because we desire the relative phase between the scattered and unscattered fields; furthermore, any constant phase shift common to both fields will drop when we take the absolute value below.
Finally, we arrive at the forward prediction of the image formed from the $p^{th}$ illumination LED at the image sensor under the first Born approximation: $$\label{born}
I_{pred}^{Born}[\cdot,\cdot,p]=|u_{back}[\cdot,\cdot,p]+u[\cdot,\cdot,p]|^2.$$ Fig. \[forward\_appendix\] graphically summarizes this forward model. Alternatively, under the first Rytov approximation, $$\label{rytov}
I_{pred}^{Rytov}[\cdot,\cdot,p]=|u_{back}[\cdot,\cdot,p]exp(u[\cdot,\cdot,p]/u_{back}[\cdot,\cdot,p])|^2.$$
![[(a) Graphical representation of the forward model under the first Born approximation. (b) $k$-space vectorial picture, depicting the interaction between $\mathbf{k^{cap}}$ and $\mathbf{k^{ill}}$ in the Fourier diffraction theorem. [\[forward\_appendix\]]{}]{}](figures/Fig3_forwardmodel_appendix.pdf){width="0.6\columnwidth"}
Multi-slice model {#multislice .unnumbered}
-----------------
In principle, any valid forward scattering model can be used with our DP-DT procedure (replacing the “forward model” box in Fig. \[forward\_figure\]). Here we describe another popular scattering model, known variously as the multi-slice (MS) approximation and the beam propagation method (BPM), whereby the sample RI is parameterized as multiple thin discrete layers, within which the thin-sample approximation is assumed to apply. An incident field then propagates layer by layer, with the field emerging from the other side serving as the forward prediction. Note that this model can account for multiple scattering, but only in the forward direction.
In particular, let $\delta n_{obj}[\cdot,\cdot,\cdot]$ be a complex-valued tensor representing the sample RI deviation from the background medium RI, $n_0$. Again, the square brackets denote discrete indexing, where the dimensions correspond to the $x$, $y$, and $z$ dimensions. Let $\delta z$ be the axial sampling over a total axial sample thickness of $\Delta z$. Then, the sample can be modeled as a stack of $\Delta z/\delta z$ thin slices, separated by $\delta z$ and with phase $$\theta_{obj}[\cdot,\cdot,r]=k\delta n_{obj}[\cdot,\cdot,r]\delta z$$ for slices $r=0, 1, ..., \Delta z/\delta z-1$. Then, given a field incident on the $r^{th}$ slice from the $p^{th}$ LED, $u_\mathbf{k_p}[\cdot,\cdot,r]$, the field exiting that slice follows the recursive relationship: $$\label{base_case}
u_\mathbf{k_p}[\cdot,\cdot,0]=u_{back}[\cdot,\cdot,p]$$ $$u_\mathbf{k_p}[\cdot,\cdot,r+1]=\mathcal{F}_2^{-1}\left(\mathcal{F}_2(u_\mathbf{k_p}[\cdot,\cdot,r])\mathcal{D}(\delta z)[\cdot,\cdot]\right)exp(1i\theta_{obj}[\cdot,\cdot,r])$$ where $u_{back}[\cdot,\cdot,p]$ is given by Eq. \[unscattered\], and $$\mathcal{D}(z)[\cdot,\cdot]=exp\left(-1i\frac{(k_x^2+k_y^2)z}{kn_0+\sqrt{k^2n_0^2-k_x^2-k_y^2}}\right)$$ is the Fresnel diffraction kernel over a distance $z$. In other words, the output field is given by the input field convolved with the diffraction kernel, followed by a phase shift due to the sample.
We then propagate the output field to the axial position of in-focus image plane in sample space, which is then imaged by the imaging optics after the sample to the detector plane: $$\begin{aligned}
&u^{det}_\mathbf{k_p}[\cdot,\cdot]=\mathcal{F}_2^{-1}(\mathcal{F}_2(u_\mathbf{k_p}[\cdot,\cdot,\Delta z/\delta z-1]) \\
&\times\mathcal{D}(\Delta z_f-\Delta z/2)[\cdot,\cdot]A[\cdot,\cdot]exp(1i\phi_{pupil}[\cdot,\cdot]))
\end{aligned}$$ where $\Delta z_f$ is the position of the focus relative to the center of the sample, and $A[\cdot,\cdot]$ and $\phi_{pupil}[\cdot,\cdot]$ are the imaging optics’ pupil amplitude and phase, as defined in the first Born and Rytov cases.
To avoid the edge effects that may arise from the implied circular convolutions when using DFTs, we apodize the input fields in Eq. \[base\_case\] with a Gaussian envelope $$u_\mathbf{k_p}^{apod}[\cdot,\cdot,0]=u_{back}[\cdot,\cdot,p]exp\left(-\frac{(x-x_p)^2+(y-y_p)^2}{2\sigma^2}\right)$$ $$x_p=(\Delta z_f+\Delta z/2)\frac{k_x^{ill}[p]}{k_z^{ill}[p]}$$ $$y_p=(\Delta z_f+\Delta z/2)\frac{k_y^{ill}[p]}{k_z^{ill}[p]}$$ where the position of the Gaussian window is chosen such that after propagation to the camera conjugate plane, the resulting Gaussian-windowed field is centered within the field of view. We used a $\sigma$ value of 0.4 of the width of the FOV.
Thus, using these apodized input fields, the MS forward prediction is $$I_{pred}^{MS}[\cdot,\cdot,p]=|u^{det}_\mathbf{k_p}[\cdot,\cdot]|^2.$$ Furthermore, the intensity data is also apodized with a centered Gaussian window of the same width to match the forward prediction: $$I_{data}[\cdot,\cdot,p]\leftarrow I_{data}[\cdot,\cdot,p]exp\left(-\frac{(x-x_p)^2+(y-y_p)^2}{2\sigma^2}\right).$$
Spatial patching with the MS model {#spatial_patching .unnumbered}
----------------------------------
As the MS model is more computationally intensive and requires more memory than the first Born model, we found that batching along the LED dimension was insufficient to alleviate the memory constraints of the GPU. To circumvent this issue, we used what we refer to as “spatial patching”, whereby at each iteration a uniformly random spatial patch is selected from the reconstruction (or the random noise input of DIP) along the $xy$ plane, as well as the spatially corresponding patch in the data, and the loss is computed only over that patch. In the case of DIP, as the same network is used for all patches, the spatial locations are encoded by the fixed random input to the CNN. After optimization, to create the final reconstruction, we use a stochastic stitching algorithm whereby m = 1,000 patches are randomly chosen and reconstructed, depadded to avoid edge artifacts, and superimposed.
Spatial patching, along with regular batching along the LED dimension, allows us to reconstruct arbitrarily large fields of view with limited GPU memory. Spatial patching was especially necessary when using DIP, which adds further computational overhead due to the use of multiple layers of 3D convolutions.
![Comparison of the quality of 3D reconstruction from **intensity** measurements, using several regularizers for various bead sizes and separations (illumination NA = 0.4, **imaging NA = 0.1**). Figure layout is analogous to that of Fig. \[bead\_sim\]. (a) 1D traces through simulations of two beads spaced axially. Rows show different bead separations, while columns show different bead sizes. Each curve corresponds to a different regularization technique (i.e., none, DIP, positivity (+), and TV) as well as the ground truth. Scale bar corresponds to the Nyquist period. (b) The RI RMSEs from the ground truths for each regularizer. Each of the four plots corresponds to a different bead size, and each curve corresponds to a different edge-to-edge bead separation, where $z_0=$ 0.75 $\mu m$.[]{data-label="bead_sim_0_1"}](figures/{bead1D_fullfig_0.1}.pdf){width=".6\columnwidth"}
![Comparison of the quality of 3D reconstruction from **intensity** measurements, using several regularizers for various bead sizes and separations (illumination NA = 0.4, **imaging NA = 0.3**). See caption for Fig. \[bead\_sim\_0\_1\].[]{data-label="bead_sim_0_3"}](figures/{bead1D_fullfig_0.3}.pdf){width=".6\columnwidth"}
![Comparison of the quality of 3D reconstruction from **phase-sensitive** measurements, using several regularizers for various bead sizes and separations (illumination NA = 0.4, **imaging NA = 0.1**). See caption for Fig. \[bead\_sim\_0\_1\].[]{data-label="bead_sim_phase_0_1"}](figures/{bead1D_fullfig_0.1_phase}.pdf){width=".6\columnwidth"}
![Comparison of the quality of 3D reconstruction from **phase-sensitive** measurements, using several regularizers for various bead sizes and separations (illumination NA = 0.4, **imaging NA = 0.3**). See caption for Fig. \[bead\_sim\_0\_1\].[]{data-label="bead_sim_phase_0_3"}](figures/{bead1D_fullfig_0.3_phase}.pdf){width=".6\columnwidth"}
![Comparison of the quality of 3D reconstruction from **phase-sensitive** measurements, using several regularizers for various bead sizes and separations (illumination NA = 0.4, **imaging NA = 0.5**). See caption for Fig. \[bead\_sim\_0\_1\].[]{data-label="bead_sim_phase_0_5"}](figures/{bead1D_fullfig_0.5_phase}.pdf){width=".6\columnwidth"}
Appendix C: Additional bead simulation results {#additional_results .unnumbered}
==============================================
Intensity measurements (imaging NA = 0.1 and 0.3) {#intensity-measurements-imaging-na-0.1-and-0.3 .unnumbered}
-------------------------------------------------
Here, we provide the figures analogous to Fig. \[bead\_sim\], which shows results for an imaging NA of 0.5, for imaging NAs of 0.1 and 0.3 (Figs. \[bead\_sim\_0\_1\] and \[bead\_sim\_0\_3\], respectively). For these imaging NAs, DIP also outperforms other regularizers, with particularly striking results for the second column of Fig. \[bead\_sim\_0\_1\].
Phase-sensitive measurements (imaging NA = 0.1, 0.3, and 0.5) {#phase_sensitive_equations .unnumbered}
-------------------------------------------------------------
In the case of a phase-sensitive DT setup, we can replace the measured intensities in Eqs. \[born\], \[rytov\], and \[mse\], respectively, with measured optical field quantities, as follows: $$f_{pred}^{Born}[\cdot,\cdot,p]=(u_{back}[\cdot,\cdot,p]+u[\cdot,\cdot,p])exp(1i\theta_{back}),
\label{born_field}$$ $$f_{pred}^{Rytov}[\cdot,\cdot,p]=u_{back}exp(1i\theta_{back})[\cdot,\cdot,p]exp(u[\cdot,\cdot,p]/u_{back}[\cdot,\cdot,p]),
\label{rytov_field}$$ $$E_{field}(F, \phi_{pupil},u_0)=\frac{1}{M}\sum_{i,j,p}|f_{pred}[i,j,p]-f_{data}[i,j,p]|^2,
\label{mse_field}$$ where $\theta_{back}$ accounts for a constant phase shift due to a difference in path length between the sample and reference fields. Note that in the intensity-only case, this factor does not contribute. Thus, since this additional phase shift is a nuisance parameter, for simplicity we do not include it in our simulations.
We repeated the bead simulation experiments in Figs. \[bead\_sim\], \[bead\_sim\_0\_1\] and \[bead\_sim\_0\_3\], this time simulating phase-sensitive detection under similar noise conditions (see below). These results are shown in Figs. \[bead\_sim\_phase\_0\_1\], \[bead\_sim\_phase\_0\_3\], and \[bead\_sim\_phase\_0\_5\], which correspond to imaging NAs of 0.1, 0.3, and 0.5, respectively. Fig. \[bead\_sim2D\_phase\] shows select $xz$ cross-sections. The results are very similar to those from intensity measurements, showing that DP-DT is applicable to interferometric and non-interferometric DT. Even though the data for phase-sensitive DT in principle has more information than for intensity DT, we hypothesize that the similarity of the results is due to the dense angular spacing of the illumination sources, resulting in redundancy of the $k$-space sampling. Furthermore, the $k$-space representation of relatively thin samples has redundancy in the $k_z$ direction, further increasing the redundancy of the measurements.
![2D cross-sections of select simulated bead pairs, reconstructed from **phase-sensitive** measurements. Figure layout is identical to that of Fig. \[bead\_sim2D\]. (a) The first row is the through-origin $k_xk_y$ cross-sections of the scattering potential spectra of the reconstructions containing all the bead pairs, under illumination NA = 0.4, imaging NA = 0.5. The second and third rows show through-center $xz$ cross sections largest simulated bead pairs at two different separations. (b) The same information as (a), but under imaging NA = 0.3. (c) The same information as (a), but under imaging NA = 0.1.[]{data-label="bead_sim2D_phase"}](figures/{bead2D_fullfig_phase}.pdf){width="0.6\columnwidth"}
Simulating noise for the phase-sensitive measurement {#noise_sim .unnumbered}
----------------------------------------------------
To simulate noise in the complex-valued datasets in order to allow fair comparison with the intensity-only scenarios, we added pixel-independent, circularly-symmetric complex Gaussian noise, such that when the intensity values are computed, the resulting quantities exhibit the same variance as predicted by photon shot noise. That is, let $I$, $E_{re}$, and $E_{im}$ be random variables representing the intensity, real part of the field, and imaginary part of the field, respectively, and let $\mu_I$, $\mu_{E, re}$, and $\mu_{E, im}$ be their respective expected value. Further, assume that all of these quantities have been properly normalized to photon counts by accounting for pixel area and integration time within a single proportionality constant, $C$; that is, $$\mu_I\leftarrow\ C^2\mu_I,\ \mu_{E, re}\leftarrow\ C\mu_{E, re},\ \mu_{E, im}\leftarrow\ C\mu_{E, im}$$ $$I\leftarrow C^2I,\ E_{re}\leftarrow CE_{re},\ E_{im}\leftarrow CE_{im}$$ Then, we have the following: $$\mu_I=\mu_{E, re}^2+\mu_{E, im}^2$$ $$I\sim Poisson(\mu_I)$$ $$E_{re}\sim Normal(\mu_{E, re}, \sigma^2)$$ $$E_{im}\sim Normal(\mu_{E, im}, \sigma^2)$$ We desire $\sigma$ such that $$\label{condition}
Var(E_{re}^2+E_{im}^2)=Var(I)=\mu_I$$ where $Var(\cdot)$ denotes variance. We note that $\frac{E_{re}^2+E_{im}^2}{\sigma^2}$ follows a noncentral chi-squared distribution as follows: $$\frac{E_{re}^2+E_{im}^2}{\sigma^2} \sim \chi^2_{noncentral}(k=2, \lambda=\frac{\mu_{E, re}^2+\mu_{E, im}^2}{\sigma^2})$$ with the following statistics: $$\label{mean}
E(E_{re}^2+E_{im}^2)=\sigma^2E\left(\frac{E_{re}^2+E_{im}^2}{\sigma^2}\right)=\mu_{E, re}^2+\mu_{E, im}^2+2\sigma^2=\mu_I+2\sigma^2$$ $$\label{variance}
Var(E_{re}^2+E_{im}^2)=\sigma^4Var\left(\frac{E_{re}^2+E_{im}^2}{\sigma^2}\right)=4\sigma^2(\sigma^2+\mu_I)$$ where $E(\cdot)$ denotes expectation (not to be confused with field). Setting Eq. \[variance\] equal to $\mu_I$, we arrive at $$\label{result}
\sigma=\sqrt{\frac{\sqrt{\mu_I^2+\mu_I}-\mu_I}{2}}\approx\frac{1}{2}$$ Thus, for our field simulation experiments, we add circularly-symmetric Gaussian noise with this standard deviation to our forward predictions.
We point out a couple interesting observations about this result. According to Eq. \[mean\], the computed intensity is biased by $2\sigma^2$. However, perhaps more interestingly, $$\lim_{\mu_I \to \infty} \sigma = \frac{1}{2}$$ In fact, even for $\mu_I=$ 1 photon, $\sigma\approx0.4551$, and for $\mu_I=$ 10 photons, $\sigma\approx0.4940$. Since in practice we detect far more than 1 photon, $\mu_I\gg1$, $\sigma=\frac{1}{2}$ is a very good approximation, regardless of the number of photons detected. We can see this by substituting this approximation in Eq. \[variance\], we have $$Var(E_{re}^2+E_{im}^2)=\mu_I+\frac{1}{4}\approx\mu_I$$
Incidentally, since $\sigma<\frac{1}{2}$, the aforementioned bias in the expected value is in practice negligible. It’s also worth noting that for large photon counts, the noncentral chi-squared distribution, like the Poisson distribution, approaches a Gaussian distribution, such that not only do the variances aysmptotically match (Eq. \[condition\]), but also the distributions.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'Temporal logic has been widely used to express complex task specifications for cyber-physical systems (CPSs). One way to synthesize a controller for CPS under temporal logic constraints is to first abstract the CPS as a discrete transition system, and then apply formal methods. This approach, however, is computationally demanding and its scalability suffers due to the curse of dimensionality. In this paper, we propose a control barrier function (CBF) approach to abstraction-free control synthesis under a linear temporal logic (LTL) constraint. We first construct the deterministic Rabin automaton of the specification and compute an accepting run. We then compute a sequence of LTL formulae, each of which must be satisfied during a particular time interval, and prove that satisfying the sequence of formulae is sufficient to satisfy the LTL specification. Finally, we compute a control policy for satisfying each formula by constructing an appropriate CBF. We present a quadratic program to compute the controllers, and show the controllers synthesized using the proposed approach guarantees the system to satisfy the LTL specification, provided the quadratic program is feasible at each time step. A numerical case study is presented to demonstrate the proposed approach.'
author:
- 'Luyao Niu and Andrew Clark [^1] [^2]'
bibliography:
- 'IEEEabrv.bib'
- 'MyBib.bib'
title: '**Control Barrier Functions for Abstraction-Free Control Synthesis under Temporal Logic Constraints**'
---
[^1]: L. Niu and A. Clark are with the Department of Electrical and Computer Engineering, Worcester Polytechnic Institute, Worcester, MA 01609 USA. [{lniu,aclark}@wpi.edu]{}
[^2]: This work was supported by the National Science Foundation and the Office of Naval Research via grants CNS-1941670 and N00014-17-S-B001.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
author:
- |
John Ehrlinger\
Microsoft
bibliography:
- 'ggRandomForests.bib'
title: ': Exploring Random Forest Survival'
---
Introduction
============
Random forest [@Breiman:2001] (RF) is a non-parametric statistical method which requires no distributional assumptions on covariate relation to the response. RF is a robust, nonlinear technique that optimizes predictive accuracy by fitting an ensemble of trees to stabilize model estimates. Random Survival Forest (RSF) [@Ishwaran:2007a; @Ishwaran:2008] is an extension of Breiman’s RF techniques to survival settings, allowing efficient non-parametric analysis of time to event data. The package (<http://CRAN.R-project.org/package=randomForestSRC>) [@Ishwaran:RFSRC:2014] is a unified treatment of Breiman’s random forest for survival, regression and classification problems.
Predictive accuracy make RF an attractive alternative to parametric models, though complexity and interpretability of the forest hinder wider application of the method. We introduce the package (<http://CRAN.R-project.org/package=ggRandomForests>) for visually exploring random forest models. The package is structured to extract intermediate data objects from objects and generate figures using the graphics package (<http://CRAN.R-project.org/package=ggplot2>) [@Wickham:2009].
Many of the figures created by the package are also available directly from within the package. However offers the following advantages:
- Separation of data and figures: contains functions that operate on either the `rfsrc` forest object directly, or on the output from post processing functions (i.e., `plot.variable`, `var.select`) to generate intermediate data objects. functions are provide to further process these objects and plot results using the graphics package. Alternatively, users can use these data objects for their own custom plotting or analysis operations.
- Each data object/figure is a single, self contained unit. This allows simple modification and manipulation of the data or `ggplot` objects to meet users specific needs and requirements.
- We chose to use the package for our figures for flexibility in modifying the output. Each plot function returns either a single `ggplot` object, or a `list` of `ggplot` objects, allowing the use of additional functions to modify and customize the final figures.
This document is structured as a tutorial for using the package for building and post-processing random survival forest models and using the package for understanding how the forest is constructed. In this tutorial, we will build a random survival forest for the primary biliary cirrhosis (PBC) of the liver data set [@fleming:1991], available in the package.
In we introduce the `pbc` data set and summarize the proportional hazards analysis of this data from Chapter 4 of [@fleming:1991]. In , we describe how to grow a random survival forest with the package. Random forest is not a parsimonious method, but uses all variables available in the data set to construct the response predictor. We demonstrate random forest variable selection techniques () using Variable Importance (VIMP) [@Breiman:2001] in and Minimal Depth [@Ishwaran:2010] in . We then compare both methods with variables used in the [@fleming:1991] model.
Once we have an idea of which variables we are most interested in, we use dependence plots [@Friedman:2000] () to understand how these variables are related to the response. Variable dependence () plots give us an idea of the overall trend of a variable/response relation, while partial dependence plots () show us the risk adjusted relation by averaging out the effects of other variables. Dependence plots often show strongly non-linear variable/response relations that are not easily obtained through parametric modeling.
We then graphically examine forest variable interactions with the use of variable and partial dependence conditioning plots (coplots) [@chambers:1992; @cleveland:1993] () and the analogouse partial dependence surfaces () before adding concluding remarks in .
Data summary: primary biliary cirrhosis (PBC) data set
======================================================
The *primary biliary cirrhosis* of the liver (PBC) study consists of 424 PBC patients referred to Mayo Clinic between 1974 and 1984 who met eligibility criteria for a randomized placebo controlled trial of the drug D-penicillamine (DPCA). The data is described in [@fleming:1991 Chapter 0.2] and a partial likelihood model (Cox proportional hazards) is developed in Chapter 4.4. The `pbc` data set, included in the package, contains 418 observations, of which 312 patients participated in the randomized trial [@fleming:1991 Appendix D].
R> data(“pbc”, package = “randomForestSRC”)
For this analysis, we modify some of the data for better formatting of our results. Since the data contains about 12 years of follow up, we prefer using `years` instead of `days` to describe survival. We also convert the `age` variable to years, and the `treatment` variable to a factor containing levels of `c(DPCA, placebo)`. The variable names, type and description are given in .
Variable name Description Type
--------------- -------------------------------- ---------
years Time (years) numeric
status Event (F = censor, T = death) logical
treatment Treament (DPCA, Placebo) factor
age Age (years) numeric
sex Female = T logical
ascites Presence of Asictes logical
hepatom Presence of Hepatomegaly logical
spiders Presence of Spiders logical
edema Edema (0, 0.5, 1) factor
bili Serum Bilirubin (mg/dl) numeric
chol Serum Cholesterol (mg/dl) integer
albumin Albumin (gm/dl) numeric
copper Urine Copper (ug/day) integer
alk Alkaline Phosphatase (U/liter) numeric
sgot SGOT (U/ml) numeric
trig Triglicerides (mg/dl) integer
platelet Platelets per cubic ml/1000 integer
prothrombin Prothrombin time (sec) numeric
stage Histologic Stage factor
: \[tab:dta-table\]‘pbc‘ data set variable dictionary.\[T:dataLabs\]
Exploratory data analysis
-------------------------
It is good practice to view your data before beginning analysis. Exploratory Data Analysis (EDA) [@Tukey:1977] will help you to understand the data, and find outliers, missing values and other data anomalies within each variable before getting deep into the analysis. To this end, we use figures with the `facet_wrap` function to create two sets of panel plots, one of histograms for categorical variables (), and another of scatter plots for continuous variables (). Variables are plotted along a continuous variable on the X-axis to separate the individual observations.
![EDA plots for categorical variables (logicals and factors). Bars indicate number of patients within 1 year of followup interval for each categorical variable. Colors correspond to class membership within each variable. Missing values are included in white.[]{data-label="fig:categoricalEDA"}](rfs-categoricalEDA-1)
In categorical EDA plots (), we are looking for patterns of missing data (white portion of bars). We often use surgical date for our X-axis variable to look for possible periods of low enrollment. There is not a comparable variable available in the `pbc` data set, so instead we used follow up time (`years`). Another reasonable choice may have been to use the patient `age` variable for the X-axis. The important quality of the selected variable is to spread the observations out to aid in finding data anomalies.
![EDA plots for continuous variables. Symbols indicate observations with variable value on Y-axis against follow up time in years. Symbols are colored and shaped according to the death event (‘status‘ variable). Missing values are indicated by rug marks along the X-axis[]{data-label="fig:continuousEDA"}](rfs-continuousEDA-1)
In continuous data EDA plots (), we are looking for missingness (rug marks) and extreme or non-physical values. For survival settings, we color and shape the points as red ‘x’s to indicate events, and blue circles to indicate censored observation.
Extreme value examples are evident in a few of the variables in . We are typically looking for values that are outside of the biological range. This is often caused by measurements recorded in differing units, which can sometimes be corrected algorithmically. Since we can not ask the original investigator to clarify these values in this particular study, we will continue without modifying the data.
pbc pbc.trial
------------- ----- -----------
treatment 106 0
ascites 106 0
hepatom 106 0
spiders 106 0
chol 134 28
copper 108 2
alk 106 0
sgot 106 0
trig 136 30
platelet 11 4
prothrombin 2 0
stage 6 0
: \[tab:missing\]Missing value counts in ‘pbc‘ data set and pbc clinical trial observations (‘pbc.trial‘).\[T:missing\]
Both EDA figures indicate the `pbc` data set contains quite a bit of missing data. shows the number of missing values in each variable of the `pbc` data set. Of the 19 variables in the data, 12 have missing values. The `pbc` column details variables with missing data in the full `pbc` data set, though there are patients that were not randomized into the trial. If we restrict the data to the trial only, most of the missing values are also removed, leaving only 4 variables with missing values. Therefore, we will focus on the 312 observations from the clinical trial for the remainder of this document. We will discuss how handles missing values in .
PBC Model Summary
-----------------
We conclude the data set investigation with a summary of[@fleming:1991] model results from Chapter 4.4. We start by generating Kaplan–Meier (KM) survival estimates comparing the treatment groups of DPCA and placebo. We use the `gg_survival` function to generate these estimates from the data set as follows.
R> \# Create the trial and test data sets. R> pbc.trial <- pbc R> pbc.test <- pbc R> R> \# Create the gg\_survival object R> gg\_dta <- gg\_survival(interval = “years”, R+ censor = “status”, R+ by = “treatment”, R+ data = pbc.trial, R+ conf.int = 0.95)
The code block reduces the `pbc` data set to the `pbc.trial` which only include observations from the clinical trial. The remaining observations are stored in the `pbc.test` data set for later use. The package is designed to use a two step process in figure generation. The first step is data generation, where we store a `gg_survival` data object in the `gg_dta` object. The `gg_survival` function uses the `data` set, follow up `interval`, `censor` indicator and an optional grouping argument (`by`). By default `gg_survival` also calculates $95\%$ confidence band, which we can control with the `conf.int` argument.
In the figure generation step, we use the plot routine `plot.gg_survival` as shown in the following code block. The `plot.gg_survival` function uses the `gg_dta` data object to plot the survival estimate curves for each group and corresponding confidence interval ribbons. We have used additional commands to modify the axis and legend labels (`labs`), the legend location (`theme`) and control the plot range of the y-axis (`coord_cartesian`) for this figure.
R> plot(gg\_dta) + R+ labs(y = “Survival Probability”, x = “Observation Time (years)”, R+ color = “Treatment”, fill = “Treatment”) + R+ theme(legend.position = c(0.2, 0.2)) + R+ coord\_cartesian(y = c(0, 1.01))
![Kaplan–Meier survival estimates comparing the DPCA treatment (red) with placebo (blue) groups for the pbc.trail data set. Median survival with shaded 95% confidence band.[]{data-label="fig:plot_gg_survival"}](rfs-plot_gg_survival-1)
The `gg_survival` plot of is analogous to[@fleming:1991] Figure 0.2.3 and Figure 4.4.1, showing there is little difference between the treatment and control groups.
The `gg_survival` function generates a variety of time-to-event estimates, including the cumulative hazard. The follow code block creates a cumulative hazard plot [@fleming:1991 Figure 0.2.1] in using the same data object generated by the original `gg_survival` function call. The red `DPCA` line is directly comparable to Figure 0.2.1, we’ve add the cumulative hazard estimates for the `placebo` population in blue.
R> plot(gg\_dta, type = “cum\_haz”) + R+ labs(y = “Cumulative Hazard”, x = “Observation Time (years)”, R+ color = “Treatment”, fill = “Treatment”) + R+ theme(legend.position = c(0.2, 0.8)) + R+ coord\_cartesian(ylim = c(-0.02, 1.22))
![Kaplan–Meier cumulative hazard estimates comparing the DPCA treatment (red) with placebo (blue) groups for the pbc data set.[]{data-label="fig:plot_gg_cum_hazard"}](rfs-plot_gg_cum_hazard-1)
In , we demonstrated grouping on the categorical variable (`treatment`). To demonstrate plotting grouped survival on a continuous variable, we examine KM estimates of survival within stratified groups of bilirubin measures. The groupings are obtained directly from[@fleming:1991] Figure 4.4.2, where they presented univariate model results of predicting survival on a function of bilirubin.
We set up the `bili` groups on a temporary data set (`pbc.bili`) using the `cut` function with intervals matching the reference figure. For this example we combine the data generation and plot steps into a single line of code. The `error` argument of the `plot.gg_survival` function is used to control display of the confidence bands. We suppress the intervals for this figure with `error = none` and again modify the plot display with commands to generate .
R> pbc.bili <- pbc.trial R> pbc.bili$bili_grp <- cut(pbc.bili$bili, breaks = c(0, 0.8, 1.3, 3.4, 29)) R> R> plot(gg\_survival(interval = “years”, censor = “status”, by = “bili\_grp”, R+ data = pbc.bili), error = “none”) + R+ labs(y = “Survival Probability”, x = “Observation Time (years)”, R+ color = “Bilirubin”)
![Kaplan–Meier survival estimates comparing different groups of Bilirubin measures (bili) for the pbc data set. Groups defined in Chapter 4 of \[@fleming:1991\].[]{data-label="fig:gg_survival-bili"}](rfs-gg_survival-bili-1)
In Chapter 4,[@fleming:1991] use partial likelihood methods to build a linear model with log transformations on some variables. We summarize the final, biologically reasonable model in for later comparison with our random forest results.
Coef. Std. Err. Z stat.
----------------------- -------- ----------- ---------
Age 0.033 0.009 3.84
log(Albumin) -3.055 0.724 -4.22
log(Bilirubin) 0.879 0.099 8.90
Edema 0.785 0.299 2.62
log(Prothrombin Time) 3.016 1.024 2.95
: \[tab:xtab\]‘pbc‘ proportional hazards model summary of 312 randomized cases in ‘pbc.trial‘ data set. (Table 4.4.3c \[@fleming:1991\])\[T:FHmodel\]
Random survival forest
======================
A Random Forest [@Breiman:2001] is grown by *bagging* [@Breiman:1996] a collection of *classification and regression trees* (CART) [@cart:1984]. The method uses a set of $B$ *bootstrap* [@bootstrap:1994] samples, growing an independent tree model on each sub-sample of the population. Each tree is grown by recursively partitioning the population based on optimization of a *split rule* over the $p$-dimensional covariate space. At each split, a subset of $m \le p$ candidate variables are tested for the split rule optimization, dividing each node into two daughter nodes. Each daughter node is then split again until the process reaches the *stopping criteria* of either *node purity* or *node member size*, which defines the set of *terminal (unsplit) nodes* for the tree. In regression trees, node impurity is measured by mean squared error, whereas in classification problems, the Gini index is used [@Friedman:2000].
Random forest sorts each training set observation into one unique terminal node per tree. Tree estimates for each observation are constructed at each terminal node, among the terminal node members. The Random Forest estimate for each observation is then calculated by aggregating, averaging (regression) or votes (classification), the terminal node results across the collection of $B$ trees.
Random Survival Forest [@Ishwaran:2007; @Ishwaran:2008] (RSF) are an extension of Random Forest to analyze right censored, time to event data. A forest of survival trees is grown using a log-rank splitting rule to select the optimal candidate variables. Survival estimate for each observation are constructed with a Kaplan–Meier (KM) estimator within each terminal node, at each event time.
Random Survival Forests adaptively discover nonlinear effects and interactions and are fully nonparametric. Averaging over many trees enables RSF to approximate complex survival functions, including non-proportional hazards, while maintaining low prediction error. [@Ishwaran:2010a] showed that RSF is uniformly consistent and that survival forests have a uniform approximating property in finite-sample settings, a property not possessed by individual survival trees.
The `rfsrc` function call grows the forest, determining the type of forest by the response supplied in the `formula` argument. In the following code block, we grow a random forest for survival, by passing a survival (`Surv`) object to the forest. The forest uses all remaining variables in the `pbc.trial` data set to generate the RSF survival model.
R> rfsrc\_pbc <- rfsrc(Surv(years, status) ., data = pbc.trial, R+ nsplit = 10, na.action = “na.impute”, R+ tree.err = TRUE,importance = TRUE)
The `print.rfsrc` function returns information on how the random forest was grown. Here the `family = surv` forest has `ntree = 1000` trees (the default `ntree` argument). The forest selected from `ceil`$(\sqrt{p=17}) = 5$ randomly selected candidate variables for splitting at each node, stopping when a terminal node contained three or fewer observations. For continuous variables, we used a random logrank split rule, which randomly selects from `nsplit = 10` split point values, instead of optimizing over all possible values.
Generalization error
--------------------
One advantage of random forest is a built in generalization error estimate. Each bootstrap sample selects approximately $63.2\%$ of the population on average. The remaining $36.8\%$ of observations, the Out-of-Bag [@BreimanOOB:1996e] (OOB) sample, can be used as a hold out test set for each tree. An OOB prediction error estimate can be calculated for each observation by predicting the response over the set of trees which were not trained with that particular observation. Out-of-Bag prediction error estimates have been shown to be nearly identical to $n$–fold cross validation estimates [@StatisticalLearning:2009]. This feature of random forest allows us to obtain both model fit and validation in one pass of the algorithm.
The `gg_error` function operates on the random forest (`rfsrc_pbc`) object to extract the error estimates as a function of the number of trees in the forest. The following code block first creates a `gg_error` data object, then uses the `plot.gg_error` function to create a `ggplot` object for display in a single line of code.
R> plot(gg\_error(rfsrc\_pbc))
![Random forest OOB prediction error estimates as a function of the number of trees in the forest.[]{data-label="fig:errorPlot"}](rfs-errorPlot-1)
The `gg_error` plot of demonstrates that it does not take a large number of trees to stabilize the forest prediction error estimate. However, to ensure that each variable has enough of a chance to be included in the forest prediction process, we do want to create a rather large random forest of trees.
Training Set Prediction
-----------------------
The `gg_rfsrc` function extracts the OOB prediction estimates from the random forest. This code block executes the data extraction and plotting in one line, since we are not interested in holding the prediction estimates for later reuse. Each of the plot commands return `ggplot` objects, which we can also store for modification or reuse later in the analysis (`ggRFsrc` object). Note that we again use additional commands to modify the display of the plot object.
R> ggRFsrc <- plot(gg\_rfsrc(rfsrc\_pbc), alpha = 0.2) + R+ scale\_color\_manual(values = strCol) + R+ theme(legend.position = “none”) + R+ labs(y = “Survival Probability”, x = “Time (years)”) + R+ coord\_cartesian(ylim = c(-0.01, 1.01)) R> show(ggRFsrc)
![Random forest OOB predicted survival. Blue curves correspond to censored observations, red curves correspond to observations experiencing death events.[]{data-label="fig:rfsrc-plot"}](rfs-rfsrc-plot-1)
The `gg_rfsrc` plot of shows the predicted survival from our RSF model. Each line represents a single patient in the training data set, where censored patients are colored blue, and patients who have experienced the event (death) are colored in red. We extend all predicted survival curves to the longest follow up time (12 years), regardless of the actual length of a patient’s follow up time.
Interpretation of general survival properties from is difficult because of the number of curves displayed. To get more interpretable results, it is preferable to plot a summary of the survival results. The following code block compares the predicted survival between treatment groups, as we did in .
R> plot(gg\_rfsrc(rfsrc\_pbc, by = “treatment”)) + R+ theme(legend.position = c(0.2, 0.2)) + R+ labs(y = “Survival Probability”, x = “Time (years)”) + R+ coord\_cartesian(ylim = c(-0.01, 1.01))
![Random forest predicted survival stratified by treatment groups. DPCA group in red, placebo in blue with shaded 95% confidence bands.[]{data-label="fig:rfsrc-mean2"}](rfs-rfsrc-mean2-1)
The `gg_rfsrc` plot of shows the median survival with a $95\%$ shaded confidence band for the `DPCA` group in red, and the `placebo` group in blue. When calling `gg_rfsrc` with either a `by` argument or a `conf.int` argument, the function calculates a bootstrap confidence interval around the median survival line. By default, the function will calculate the `conf.int=0.95` confidence interval, with the number of `bs.samples` equal to the number of observations.
Random forest imputation
------------------------
There are two modeling issues when dealing with missing data values: `How does the algorithm build a model when values are missing from the training data?`, and `How does the algorithm predict a response when values are missing from the test data?`. The standard procedure for linear models is to either remove or impute the missing data values before modelling. Removing the missingness is done by either removing the variable with missing values (column wise) or removing the observations (row wise). Removal is a simple solution, but may bias results when either observations or variables are scarce.
The package imputes missing values using *adaptive tree imputation* [@Ishwaran:2008]. Rather than impute missing values before growing the forest, the algorithm takes a `just-in-time` approach. At each node split, the set of `mtry` candidate variables is checked for missing values. Missing values are then imputed by randomly drawing values from non-missing data within the node. The split-statistic is then calculated on observations that were not missing values. The imputed values are used to sort observations into the subsequent daughter nodes and then discarded before the next split occurs. The process is repeated until the stopping criteria is reached and all observations are sorted into terminal nodes.
A final imputation step can be used to fill in missing values from within the terminal nodes. This step uses a process similar to the previous imputation but uses the OOB non-missing terminal node data for the random draws. These values are aggregated (averaging for continuous variables, voting for categorical variables) over the `ntree` trees in the forest to estimate an imputed data set. By default, the missing values are not filled into the training data, but are available within the forest object for later use if desired.
Adaptive tree imputation still requires the missing at random assumptions [@Rubin:1976]. At each imputation step, the random forest assumes that similar observations are grouped together within each node. The random draws used to fill in missing data do not bias the split rule, but only sort observations similar in non-missing data into like nodes. An additional feature of this approach is the ability of predicting on test set observations with missing values.
Test set predictions
--------------------
The strength of adaptive tree imputation becomes clear when doing prediction on test set observations. If we want to predict survival for patients that did not participate in the trial using the model we created in Section \[random-survival-forest\], we need to somehow account for the missing values detailed in .
The `predict.rfsrc` call takes the forest object (`rfsrc_pbc`), and the test data set (`pbc_test`) and returns a predicted survival using the same forest imputation method for missing values within the test data set (`na.action=na.impute`).
R> rfsrc\_pbc\_test <- predict(rfsrc\_pbc, newdata = pbc.test, R+ na.action = “na.impute”, R+ importance = TRUE)
The forest summary indicates there are 106 test set observations with 36 deaths and the predicted error rate is $19.1\%$. We plot the predicted survival just as we did the training set estimates.
R> plot(gg\_rfsrc(rfsrc\_pbc\_test), alpha=.2) + R+ scale\_color\_manual(values = strCol) + R+ theme(legend.position = “none”) + R+ labs(y = “Survival Probability”, x = “Time (years)”) + R+ coord\_cartesian(ylim = c(-0.01, 1.01))
![Random forest survival estimates for patients in the pbc.test data set. Blue curves correspond to censored patients, red curves correspond to patients experiencing a death event.[]{data-label="fig:predictPlot"}](rfs-predictPlot-1)
The `gg_rfsrc` plot of shows the test set predictions, similar to the training set predictions in , though with fewer patients the survival curves do not cover the same area of the figure. It is important to note that because is constructed with OOB estimates, the survival results are comparable as estimates from unseen observations in .
Variable selection
==================
Random forest is not a parsimonious method, but uses all variables available in the data set to construct the response predictor. Also, unlike parametric models, random forest does not require the explicit specification of the functional form of covariates to the response. Therefore there is no explicit $p$-value/significance test for variable selection with a random forest model. Instead, RF ascertains which variables contribute to the prediction through the split rule optimization, optimally choosing variables which separate observations.
The typical goal of a random forest analysis is to build a *prediction* model, in contrast to extracting *information* regarding the underlying process [@Breiman:twoCultures:2001]. There is not usually much care given in how variables are included into the training data set. Since the goal is prediction, investigators often include the “kitchen sink” if it can help.
In contrast, in survival settings we are typically also interested in how we can possibly improve the the outcome of interest. To achieve this, for understandable inference, it is important to avoid both duplication and transformations of variables whenever possible when building our data sets. Duplication of variables, including multiple measures of a similar covariate, can reduce or mask the importance of the covariate. Transformations can also mask importance as well as make interpretation of the inference results difficult to impossible.
In this Section, We explore two separate approaches to investigate the RF variable selection process. Variable Importance (), a property related to variable misspecification, and Minimal Depth (), a property derived from the construction of the trees within the forest.
Variable Importance
-------------------
*Variable importance* (VIMP) was originally defined in CART using a measure involving surrogate variables (see Chapter 5 of [@cart:1984]). The most popular VIMP method uses a prediction error approach involving “noising-up” each variable in turn. VIMP for a variable $x_v$ is the difference between prediction error when $x_v$ is randomly permuted, compared to prediction error under the observed values [@Breiman:2001; @Liaw:2002; @Ishwaran:2007; @Ishwaran:2008].
Since VIMP is the difference in OOB prediction error before and after permutation, a large VIMP value indicates that misspecification detracts from the predictive accuracy in the forest. VIMP close to zero indicates the variable contributes nothing to predictive accuracy, and negative values indicate the predictive accuracy *improves* when the variable is misspecified. In the later case, we assume noise is more informative than the true variable. As such, we ignore variables with negative and near zero values of VIMP, relying on large positive values to indicate that the predictive power of the forest is dependent on those variables.
The `gg_vimp` function extracts VIMP measures for each of the variables used to grow the forest. The `plot.gg_vimp` function shows the variables, in VIMP rank order, labeled with the named vector in the `lbls` argument.
R> plot(gg\_vimp(rfsrc\_pbc), lbls = st.labs) + R+ theme(legend.position = c(0.8, 0.2)) + R+ labs(fill = “VIMP > 0”)
![Random forest Variable Importance (VIMP). Blue bars indicates positive VIMP, red indicates negative VIMP. Importance is relative to positive length of bars.[]{data-label="fig:rf-vimp"}](rfs-rf-vimp-1)
The `gg_vimp` plot of details VIMP ranking for the `pbc.trial` baseline variables, from the largest (Serum Bilirubin) at the top, to smallest (Treament (DPCA, Placebo)) at the bottom. VIMP measures are shown using bars to compare the scale of the error increase under permutation and colored by the sign of the measure (red for negative values). Note that four of the five highest ranking variables by VIMP match those selected by the[@fleming:1991] model listed in , with urine copper (2) ranking higher than age (8). We will return to this in .
Minimal Depth
-------------
In VIMP, prognostic risk factors are determined by testing the forest prediction under alternative data settings, ranking the most important variables according to their impact on predictive ability of the forest. An alternative method uses inspection of the forest construction to rank variables. *Minimal depth* [@Ishwaran:2010; @Ishwaran:2011] assumes that variables with high impact on the prediction are those that most frequently split nodes nearest to the root node, where they partition the largest samples of the population.
Within each tree, node levels are numbered based on their relative distance to the root of the tree (with the root at 0). Minimal depth measures important risk factors by averaging the depth of the first split for each variable over all trees within the forest. The assumption in the metric is that smaller minimal depth values indicate the variable separates large groups of observations, and therefore has a large impact on the forest prediction.
In general, to select variables according to VIMP, we examine the VIMP values, looking for some point along the ranking where there is a large difference in VIMP measures. Given minimal depth is a quantitative property of the forest construction, [@Ishwaran:2010] also derive an analytic threshold for evidence of variable impact. A simple optimistic threshold rule uses the mean of the minimal depth distribution, classifying variables with minimal depth lower than this threshold as important in forest prediction.
The `var.select` function uses the minimal depth methodology for variable selection, returning an object with both minimal depth and vimp measures. The `gg_minimal_depth` function is analogous to the `gg_vimp` function. Variables are ranked from most important at the top (minimal depth measure), to least at the bottom (maximal minimal depth).
R> varsel\_pbc <- var.select(rfsrc\_pbc)
minimal depth variable selection ...
———————————————————– family : surv var. selection : Minimal Depth conservativeness : medium x-weighting used? : TRUE dimension : 17 sample size : 312 ntree : 1000 nsplit : 10 mtry : 5 nodesize : 3 refitted forest : FALSE model size : 14 depth threshold : 6.7549 PE (true OOB) : 16.502
Top variables: depth vimp bili 1.708 0.067 albumin 2.483 0.012 copper 2.904 0.015 prothrombin 2.931 0.014 chol 3.227 0.006 platelet 3.329 0.000 edema 3.333 0.016 sgot 3.677 0.007 age 3.702 0.009 alk 4.039 0.001 trig 4.514 0.000 ascites 5.194 0.013 stage 5.247 0.007 hepatom 6.476 0.003 ———————————————————–
R> gg\_md <- gg\_minimal\_depth(varsel\_pbc, lbls = st.labs) R> \# print(gg\_md)
The `gg_minimal_depth` summary mostly reproduces the output from the `var.select` function from the package. We report the minimal depth threshold (`threshold` 6.755) and the number of variables with depth below that threshold (`model size` 14). We also list a table of the top (14) selected variables, in minimal depth rank order with the associated VIMP measures. The minimal depth numbers indicate that `bili` tends to split between the first and second node level, and the next three variables (`albumin`, `copper`, `prothrombin`) split between the second and third levels on average.
R> plot(gg\_md, lbls = st.labs)
![Minimal Depth variable selection. Low minimal depth indicates important variables. The dashed line is the threshold of maximum value for variable selection.[]{data-label="fig:mindepth-plot"}](rfs-mindepth-plot-1)
The `gg_minimal_depth` plot of is similar to the `gg_vimp` plot in , ranking variables from most important at the top (minimal depth measure), to least at the bottom (maximal minimal depth). The vertical dashed line indicates the minimal depth threshold where smaller minimal depth values indicate higher importance and larger values indicate lower importance.
Variable selection comparison
-----------------------------
Since the VIMP and Minimal Depth measures use different criteria, we expect the variable ranking to be somewhat different. We use `gg_minimal_vimp` function to compare rankings between minimal depth and VIMP in .
R> plot(gg\_minimal\_vimp(gg\_md), lbls = st.labs) + R+ theme(legend.position=c(0.8, 0.2))
![Comparing Minimal Depth and Vimp rankings. Points on the red dashed line are ranked equivalently, points above have higher VIMP ranking, those below have higher minimal depth ranking.[]{data-label="fig:depthVimp"}](rfs-depthVimp-1)
The points along the red dashed line indicate where the measures are in agreement. Points above the red dashed line are ranked higher by VIMP than by minimal depth, indicating the variables are more sensitive to misspecification. Those below the line have a higher minimal depth ranking, indicating they are better at dividing large portions of the population. The further the points are from the line, the more the discrepancy between measures.
Variable FH Min depth VIMP
------------- ---- ----------- ------
bili 1 1 1
albumin 2 2 6
copper NA 3 3
prothrombin 4 4 4
chol NA 5 10
platelet NA 6 16
edema 5 7 2
sgot NA 8 8
age 3 9 7
alk NA 10 13
trig NA 11 15
ascites NA 12 5
stage NA 13 9
hepatom NA 14 11
: \[tab:models\]Comparison of variable selection criteria. Minimal depth ranking, VIMP ranking and \[@fleming:1991\] (FH) proportional hazards model ranked according to ‘abs(Z stat)‘ from Table \[T:FHmodel\].\[T:modelComp\]
We examine the ranking of the different variable selection methods further in . We can use the Z statistic from to rank variables selected in the[@fleming:1991] model to compare with variables selected by minimal depth and VIMP. The table is constructed by taking the top ranked minimal depth variables (below the selection threshold) and matching the VIMP ranking and[@fleming:1991] model transforms. We see all three methods indicate a strong relation of serum bilirubin to survival, and overall, the minimal depth and VIMP rankings agree reasonably well with the[@fleming:1991] model.
The minimal depth selection process reduced the number of variables of interest from\~17 to 14, which is still a rather large subset of interest. An obvious selection set is to examine the five variables selected by[@fleming:1991]. Combining the Minimal Depth and[@fleming:1991] model, there may be evidence to keep the top 7 variables. Though minimal depth does not indicate the `edema` variable is very interesting, VIMP ranking does agree with the proportional hazards model, indicating we might not want to remove the `edema` variable. Both minimal depth and VIMP suggest including `copper`, a measure associated with liver disease.
Regarding the `chol` variable, recall missing data summary of . In in the trial data set, there were 28 observations missing `chol` values. The forest imputation randomly sorts observations with missing values into daughter nodes when using the `chol` variable, which is also how calculates VIMP. We therefore expect low values for VIMP when a variable has a reasonable number of missing values.
Restricting our remaining analysis to the five[@fleming:1991] variables, plus the `copper` retains the biological sense of these analysis. We will now examine how these six variables are related to survival using variable dependence methods to determine the direction of the effect and verify that the log transforms used by[@fleming:1991] are appropriate.
Variable/Response dependence
============================
As random forest is not parsimonious, we have used minimal depth and VIMP to reduce the number of variables to a manageable subset. Once we have an idea of which variables contribute most to the predictive accuracy of the forest, we would like to know how the response depends on these variables.
Although often characterized as a *black box* method, the forest predictor is a function of the predictor variables $\hat{f}_{RF} = f(x).$ We use graphical methods to examine the forest predicted response dependency on covariates. We again have two options, variable dependence plots () are quick and easy to generate, and partial dependence plots () are more computationally intensive but give us a risk adjusted look at variable dependence.
Variable dependence
-------------------
*Variable dependence* plots show the predicted response relative to a covariate of interest, with each training set observation represented by a point on the plot. Interpretation of variable dependence plots can only be in general terms, as point predictions are a function of all covariates in that particular observation.
Variable dependence is straight forward to calculate, involving only the getting the predicted response for each observation. In survival settings, we must account for the additional dimension of time. We plot the response at specific time points of interest, for example survival at 1 or 3 years.
R> ggRFsrc + geom\_vline(aes(xintercept = 1), linetype = “dashed”) + R+ geom\_vline(aes(xintercept = 3), linetype = “dashed”) + R+ coord\_cartesian(xlim = c(0, 5))
![Random forest predicted survival (Figure \[fig:rfsrc-plot\]) with vertical dashed lines indicate the 1 and 3 year survival estimates.[]{data-label="fig:rfsrc-plot3Mnth"}](rfs-rfsrc-plot3Mnth-1)
The `gg_rfsrc` of identical to (stored in the `ggRFsrc` variable) with the addition of a vertical dashed line at the 1 and 3 year survival time. A variable dependence plot is generated from the predicted response value of each survival curve at the intersecting time line plotted against covariate value for that observation. This can be visualized as taking a slice of the predicted response at each time line, and spreading the resulting points out along the variable of interest.
The `gg_variable` function extracts the training set variables and the predicted OOB response from `rfsrc` and `predict` objects. In the following code block, we store the `gg_variable` data object for later use (`gg_v`), as all remaining variable dependence plots can be constructed from this object.
R> gg\_v <- gg\_variable(rfsrc\_pbc, time = c(1, 3), R+ time.labels = c(“1 Year”, “3 Years”)) R> R> plot(gg\_v, xvar = “bili”, alpha = 0.4) + \#, se=FALSE R+ labs(y = “Survival”, x = st.labs\[“bili”\]) + R+ theme(legend.position = “none”) + R+ scale\_color\_manual(values = strCol, labels = event.labels) + R+ scale\_shape\_manual(values = event.marks, labels = event.labels) + R+ coord\_cartesian(ylim = c(-0.01, 1.01))
![Variable dependence of survival at 1 and 3 years on bili variable. Individual cases are marked with blue circles (alive or censored) and red x (dead). Loess smooth curve with shaded 95% confidence band indicates decreasing survival with increasing bilirubin.[]{data-label="fig:variable-plotbili"}](rfs-variable-plotbili-1)
The `gg_variable` plot of shows variable dependence for the Serum Bilirubin (`bili`) variable. Again censored cases are shown as blue circles, events are indicated by the red `x` symbols. Each predicted point is dependent on the full combination of all other covariates, not only on the covariate displayed in the dependence plot. The smooth loess line [@cleveland:1981; @cleveland:1988] indicates the trend of the prediction over the change in the variable.
Examination of indicates most of the cases are grouped in the lower end of `bili` values. We also see that most of the higher values experienced an event. The “normal” range of Bilirubin is from 0.3 to 1.9 mg/dL, indicating the distribution from our population is well outside the normal range. These values make biological sense considering Bilirubin is a pigment created in the liver, the organ effected by the PBC disease. The figure also shows that the risk of death increases as time progresses. The risk at 3 years is much greater than that at 1 year for patients with high Bilirubin values compared to those with values closer to the normal range.
The `plot.gg_variable` function call operates on the `gg_variable` object controlled by the list of variables of interest in the `xvar` argument. By default, the `plot.gg_variable` function returns a list of `ggplot` objects, one figure for each variable named in `xvar`. The remaining arguments are passed to internal functions controlling the display of the figure. The `se` argument is passed to the internal call to `geom_smooth` for fitting smooth lines to the data. The `alpha` argument lightens the coloring points in the `geom_point` call, making it easier to see point over plotting. We also demonstrate modification of the plot labels using the `labs` function and point attributes with the `scale_` functions.
An additional `plot.gg_variable` argument (`panel = TRUE`) can be used to combine multiple variable dependence plots into a single figure. In the following code block, we plot the remaining continuous variables of interest found in .
R> xvar <- c(“bili”, “albumin”, “copper”, “prothrombin”, “age”) R> xvar.cat <- c(“edema”) R> R> plot(gg\_v, xvar = xvar\[-1\], panel = TRUE, alpha = 0.4) + R+ labs(y = “Survival”) + R+ theme(legend.position = “none”) + R+ scale\_color\_manual(values = strCol, labels = event.labels) + R+ scale\_shape\_manual(values = event.marks, labels = event.labels) + R+ coord\_cartesian(ylim = c(-0.05, 1.05))
![Variable dependence of predicted survival at 1 and 3 years on continuous variables of interest. Individual cases are marked with blue circles for censored cases and red x for death events. Loess smooth curve indicates the survival trend with increasing values.[]{data-label="fig:variable-plot"}](rfs-variable-plot-1)
The `gg_variable` plot in displays a panel of the remaining continuous variable dependence plots. The panels are sorted in the order of variables in the `xvar` argument and include a smooth loess line [@cleveland:1981; @cleveland:1988] to indicate the trend of the prediction dependence over the covariate values. The `se=FALSE` argument turns off the loess confidence band, and the `span=1` argument controls the degree of smoothing.
The figures indicate that survival increases with `albumin` level, and decreases with `bili`, `copper`, `prothrombin` and `age`. Note the extreme value of `prothrombin` (> 16) influences the loess curve more than other points, which would make it a candidate for further investigation.
We expect survival at 3 years to be lower than at 1 year. However, comparing the two time plots for each variable does indicate a difference in response relation for `bili`, `copper` and `prothrombine`. The added risk for high levels of these variables at 3 years indicates a non-proportional hazards response. The similarity between the time curves for `albumin` and `age` indicates the effect of these variables is constant over the disease progression.
There is not a convenient method to panel scatter plots and boxplots together, so we recommend creating panel plots for each variable type separately. We plot the categorical variable (`edema`) in separately from the continuous variables in .
R> plot(gg\_v, xvar = xvar.cat, alpha = 0.4) + labs(y = “Survival”) + R+ theme(legend.position = “none”) + R+ scale\_color\_manual(values = strCol, labels = event.labels) + R+ scale\_shape\_manual(values = event.marks, labels = event.labels) + R+ coord\_cartesian(ylim = c(-0.01, 1.02))
![Variable dependence of survival 1 and 3 years on edema categorical variable. Symbols with blue circles indicate censored cases and red x indicate death events. Boxplots indicate distribution of predicted survival for all observations within each edema group.[]{data-label="fig:variable-plotCat"}](rfs-variable-plotCat-1)
The `gg_variable` plot of for categorical variable dependence displays boxplots to examine the distribution of predicted values within each level of the variable. The points are plotted with a jitter to see the censored and event markers more clearly. The boxes are shown with horizontal bars indicating the median, 75th (top) and 25th (bottom) percentiles. Whiskers extend to 1.5 times the interquartile range. Points plotted beyond the whiskers are considered outliers.
When using categorical variables with linear models, we use boolean dummy variables to indicate class membership. In the case of `edema`, we would probably create two logical variables for `edema = 0.5` (complex Edema presence indicator) and `edema = 1.0` (Edema with diuretics) contrasted with the `edema = 0` variable (no Edema). Random Forest can use factor variables directly, separating the populations into homogeneous groups of `edema` at nodes that split on that variable. indicates similar survival response distribution between 1 and 3 year when `edema = 1.0`. The distribution of predicted survival does seem to spread out more than for the other values, again indicating a possible non-proportional hazards response.
Partial dependence
------------------
*Partial dependence* plots are a risk adjusted alternative to variable dependence. Partial plots are generated by integrating out the effects of variables beside the covariate of interest. The figures are constructed by selecting points evenly spaced along the distribution of the variable of interest. For each of these points ($X = x$), we calculate the average RF prediction over all remaining covariates in the training set by
$$\tilde{f}(x) = \frac{1}{n} \sum_{i = 1}^n \hat{f}(x, x_{i, o}),
\label{E:partial}$$
where $\hat{f}$ is the predicted response from the random forest and $x_{i, o}$ is the value for all other covariates other than $X = x$ for observation $i$ [@Friedman:2000].
Generating partial dependence data is effectively averaging the response for a series of nomograms constructed for each observation by varying the variable of interest. The operation is computationally intensive, especially when there are a large number of observations. The default parameters for the `plot.variable` function generate partial dependence estimates at `pts = 25` points along the variable of interest. For each point of interest, the `plot.variable` function averages the `n` response predictions. This process is repeated for each of the variables of interest.
For time to event data, we also have to deal with the additional time dimension, as with variable dependence. The following code block uses the `mclapply` function from the package to run the `plot.variable` function for three time points (`time`=1, 3 and 5 years) in parallel. For RSF models, we calculate a risk adjusted survival estimates (`surv.type=surv`), suppressing the internal base graphs (`show.plots = FALSE`) and store the point estimates in the `partial_pbc` `list`.
R> xvar <- c(xvar, xvar.cat) R> R> time\_index <- c(which(rfsrc\_pbc$time.interest > 1)[1]-1,
R+ which(rfsrc_pbc$time.interest > 3)\[1\]-1, R+ which(rfsrc\_pbc$time.interest > 5)[1]-1)
R> partial_pbc <- mclapply(rfsrc_pbc$time.interest\[time\_index\], R+ function(tm)[ R+ plot.variable(rfsrc\_pbc, surv.type = “surv”, R+ time = tm, xvar.names = xvar, R+ partial = TRUE , R+ show.plots = FALSE) R+ ]{})
Because partial dependence data is collapsed onto the risk adjusted response, we can show multiple time curves on a single panel. The following code block converts the `plot.variable` output into a list of `gg_partial` objects, and then combines these data objects, with descriptive labels, along each variable of interest using the `combine.gg_partial` function.
R> gg\_dta <- mclapply(partial\_pbc, gg\_partial) R> pbc\_ggpart <- combine.gg\_partial(gg\_dta\[\[1\]\], gg\_dta\[\[2\]\], R+ lbls = c(“1 Year”, “3 Years”))
We then segregate the continuous and categorical variables, and generate a panel plot of all continuous variables in the `gg_partial` plot of . The panels are ordered by minimal depth ranking. Since all variables are plotted on the same Y-axis scale, those that are strongly related to survival make other variables look flatter. The figures also confirm the strong non-linear contribution of these variables. Non-proportional hazard response is also evident in at least the `bili` and `copper` variables by noting the divergence of curves as time progresses.
![Partial dependence of predicted survival at 1 year (red circle) and 3 years (blue triangle) as a function continuous variables of interest. Symbols are partial dependence point estimates with loess smooth line to indicate trends.[]{data-label="fig:pbc-partial-panel"}](rfs-pbc-partial-panel-1)
Categorical partial dependence is displayed as boxplots, similar to categorical variable dependence. Risk adjustment greatly reduces the spread of the response as expected, and may also move the mean response compared to the unadjusted results. The categorical `gg_partial` plot of indicates that, adjusting for other variables, survival decreases with rising `edema` values. We also note that the risk adjusted distribution does spread out as we move further out in time.
R> ggplot(pbc\_ggpart\[\[“edema”\]\], aes(y=yhat, x=edema, col=group))+ R+ geom\_boxplot(notch = TRUE, R+ outlier.shape = NA) + \# panel=TRUE, R+ labs(x = “Edema”, y = "Survival (R+ theme(legend.position = c(0.1, 0.2))
![Partial dependence plot of predicted survival at 1 year (red) and 3 years (blue) as a function of edema groups (categorical variable). Boxplots indicate distribution within each group.[]{data-label="fig:pbc-partial-edema"}](rfs-pbc-partial-edema-1)
Partial dependence is an extrapolation operation. By averaging over a series of nomograms, the algorithm constructs observations for all values of the variable of interest, regardless of the relation with other variables. In contrast, variable dependence only uses observations from within the training set. A simple example would be for a model including BMI, weight and height. When examining partial dependence of BMI, the algorithm only manipulates BMI values, height or weight values. The averaging operation is then confounded in two directions. First, dependence on height and weight is shared with BMI, making it difficult to see the true response dependence. Second, partial dependence is calculated over nomograms that can not physically occur. For simple variable combinations, like BMI, it is not difficult to recognize this and modify the independent variable list to avoid these issues. However, care must be taken when interpreting more complex biological variables.
Temporal partial dependence
---------------------------
In the previous section, we calculated risk adjusted (partial) dependence at two time points (1 and 3 years). The selection of these points can be driven by biological times of interest (i.e., 1 year and 5 year survival in cancer studies) or by investigating time points of interest from a `gg_rfsrc` prediction plot. We typically restrict generating `gg_partial` plots to the variables of interest at two or three time points of interest due to computational constraints.
It is instructive to see a more detailed map of the risk adjusted response to get a feel for interpreting partial and variable dependence plots. In , we can visualize the two curves as extending into the plane of the page along a time axis. Filling in more partial dependence curves, it is possible to create a partial dependence surface.
For this exercise, we will generate a series of 50 `gg_partial` plot curves for the `bili` variable. To fill the surface in, we also increased the number of points along the distribution of `bili` to `npts=50` to create a grid of $50 \times 50$ risk adjusted estimates of survival along time in one dimension and the `bili` variable in the second.
![Partial dependence surface. Partial dependence of predicted survival (0 to 5 years) as a function of bili. Blue lines indicate partial dependence at 1 and 3 years, as in bili panel of .[]{data-label="fig:timeSurface3d"}](rfs-timeSurface3d-1)
The `gg_partial` surface of was constructed using the `surf3D` function from the package [@plot3D:2014 <http://CRAN.R-project.org/package=plot3D>].
The figure shows partial dependence of survival (Z-axis) as a function of `bili` over a five year follow up time period. Lines perpendicular to the Bilirubin axis are distributed along the `bili` variable. Lines parallel to the Bilirubin axis are taken at 50 training set event times, the first event after $t=0$ at the back to last event before $t=5$ years at the front. The distribution of the time lines is also evenly selected using the same procedure as selecting points for partial dependence curves.
The 2500 estimated partial dependence points are joined together with a simple straight line interpolation to create the surface, colored according to the survival estimates (yellow close to 1, red for lower values) to aid the visualization of 3 dimensions on a 2 dimensional page. The blue lines in correspond to the 1 and 3 year partial dependence, as shown in the `bili` panel of .
Viewed as a surface, we see how the partial dependence changes with time. For low values of `bili`, survival decreases at a constant rate. For higher values, the rate seems constant until somewhere near 2 years, where it increases rapidly before slowing again as we approach the 5 year point.
Variable interactions
=====================
We could stop with the results that our RF analysis has found these six variables to be important in predicting survival. Where the survival response is decreasing with increasing `bili`, `copper`, `prothrombin`, `age` and `edema` and increasing with increasing `albumin`. These results agree with the sign of the [@fleming:1991] model coefficients shown in . The `gg_partial` plot in supports the `log` transform of `bili`, `albumin` and `prothrombin` and suggest a similar transform for including the `copper` variable in a proportional hazards model. The `age` variable does seem to have a more linear response than the other continuous variables, and using dummy variables for `edema` would preclude the need for a transformation.
Using minimal depth, it is also possible to calculate measures of pairwise interactions among variables. Recall that minimal depth measure is defined by averaging the tree depth of variable $i$ relative to the root node. To detect interactions, this calculation can be modified to measure the minimal depth of a variable $j$ with respect to the maximal subtree for variable $i$ [@Ishwaran:2010; @Ishwaran:2011].
The `randomForestSRC::find.interaction` function traverses the forest, calculating all pairwise minimal depth interactions, and returns a $p \times p$ matrix of interaction measures. The diagonal terms are normalized to the root node, and off diagonal terms are normalized measures of pairwise variable interaction.
R> ggint <- gg\_interaction(rfsrc\_pbc)
Method: maxsubtree No. of variables: 17 Variables sorted by minimal depth?: TRUE
bili albumin copper prothrombin chol edema platelet sgot age bili 0.11 0.26 0.29 0.30 0.29 0.50 0.29 0.31 0.28 albumin 0.33 0.17 0.37 0.39 0.38 0.66 0.37 0.39 0.37 copper 0.36 0.38 0.19 0.40 0.41 0.66 0.41 0.40 0.39 prothrombin 0.40 0.41 0.43 0.20 0.45 0.69 0.46 0.45 0.44 chol 0.42 0.42 0.46 0.48 0.22 0.76 0.47 0.46 0.43 edema 0.45 0.47 0.50 0.51 0.50 0.22 0.49 0.51 0.48 platelet 0.49 0.50 0.53 0.54 0.52 0.80 0.22 0.53 0.51 sgot 0.49 0.47 0.50 0.53 0.52 0.80 0.52 0.25 0.49 age 0.41 0.44 0.47 0.50 0.48 0.79 0.48 0.46 0.25 alk 0.55 0.55 0.57 0.60 0.57 0.83 0.59 0.59 0.55 trig 0.56 0.55 0.57 0.59 0.57 0.84 0.58 0.57 0.55 ascites 0.53 0.54 0.59 0.58 0.58 0.76 0.58 0.60 0.57 stage 0.59 0.60 0.61 0.64 0.61 0.83 0.63 0.62 0.60 hepatom 0.67 0.68 0.70 0.71 0.69 0.86 0.72 0.70 0.69 spiders 0.81 0.82 0.84 0.84 0.84 0.94 0.83 0.82 0.81 treatment 0.87 0.87 0.88 0.87 0.88 0.96 0.89 0.88 0.87 sex 0.84 0.84 0.86 0.86 0.85 0.94 0.86 0.86 0.84 alk trig ascites stage hepatom spiders treatment sex bili 0.31 0.35 0.65 0.45 0.60 0.62 0.61 0.66 albumin 0.39 0.44 0.78 0.55 0.67 0.70 0.66 0.72 copper 0.42 0.45 0.77 0.57 0.69 0.71 0.69 0.79 prothrombin 0.45 0.48 0.81 0.61 0.72 0.76 0.72 0.79 chol 0.46 0.50 0.86 0.63 0.77 0.76 0.72 0.81 edema 0.50 0.55 0.80 0.64 0.73 0.75 0.75 0.81 platelet 0.53 0.55 0.90 0.70 0.81 0.81 0.77 0.86 sgot 0.51 0.53 0.90 0.67 0.79 0.79 0.76 0.84 age 0.46 0.48 0.88 0.62 0.77 0.76 0.72 0.80 alk 0.27 0.61 0.93 0.74 0.85 0.83 0.81 0.88 trig 0.56 0.30 0.92 0.74 0.85 0.84 0.80 0.90 ascites 0.59 0.62 0.34 0.69 0.77 0.80 0.80 0.83 stage 0.62 0.64 0.91 0.35 0.84 0.83 0.81 0.87 hepatom 0.71 0.71 0.91 0.80 0.43 0.88 0.85 0.89 spiders 0.83 0.84 0.96 0.90 0.92 0.53 0.90 0.95 treatment 0.87 0.88 0.99 0.94 0.96 0.95 0.55 0.97 sex 0.86 0.86 0.98 0.91 0.94 0.95 0.92 0.60
The `gg_interaction` function wraps the `find.interaction` matrix for use with the plot and print functions. The `xvar` argument is used to restrict the variables of interest and the `panel = TRUE` argument displays the results in a single figure.
R> plot(ggint, xvar = xvar)
![Minimal depth variable interaction plot for six variables of interest. Higher values indicate lower interactivity with target variable marked in red.[]{data-label="fig:interactionPanel"}](rfs-interactionPanel-1)
The `gg_interaction` plots in show interactions for the target variable (shown with the red cross) with interaction scores for all remaining variables. We expect the covariate with lowest minimal depth (`bili`) to be associated with almost all other variables, as it typically splits close to the root node, so viewed alone it may not be as informative as looking at a collection of interactive depth plots. Scanning across the panels, we see each successive target depth increasing, as expected. We also see the interactive variables increasing with increasing target depth.
Conditional dependence plots
============================
Conditioning plots (coplots) [@chambers:1992; @cleveland:1993] are a powerful visualization tool to efficiently study how a response depends on two or more variables [@cleveland:1993]. The method allows us to view data by grouping observations on some conditional membership. The simplest example involves a categorical variable, where we plot our data conditional on class membership, for instance on groups of the `edema` variable. We can view a coplot as a stratified variable dependence plot, indicating trends in the RF prediction results within panels of group membership.
Interactions with categorical data can be generated directly from variable dependence plots. Recall the variable dependence for bilirubin shown in . We recreated the `gg_variable` plot in , modified by adding a linear smooth as we intend on segregating the data along conditional class membership.
R> \# Get variable dependence at 1 year R> ggvar <- gg\_variable(rfsrc\_pbc, time = 1) R> R> \# For labeling coplot membership R> ggvar$edema <- paste("edema = ", ggvar$edema, sep = “”) R> R> \# Plot with linear smooth (method argument) R> var\_dep <- plot(ggvar, xvar = “bili”, R+ alpha = 0.5) + R+ \# geom\_smooth(method = “glm”,se = FALSE) + R+ labs(y = “Survival”, R+ x = st.labs\[“bili”\]) + R+ theme(legend.position = “none”) + R+ scale\_color\_manual(values = strCol, labels = event.labels) + R+ scale\_shape\_manual(values = event.marks, labels = event.labels) + R+ coord\_cartesian(y = c(-.01,1.01)) R> R> var\_dep
![Variable dependence of survival at 1 year against bili variable. Reproduction of top panel of with a linear smooth to indicate trend.[]{data-label="fig:var_dep"}](rfs-var_dep-1)
We can view the conditional dependence of survival against bilirubin, conditional on `edema` group membership (categorical variable) in by reusing the saved `ggplot` object (`var_dep`) and adding a call to the `facet_grid` function.
R> var\_dep + facet\_grid( edema)
![Variable dependence coplot of survival at 1 year against bili, conditional on edema group membership. Linear smooth indicates trend of variable dependence.[]{data-label="fig:coplot_bilirubin"}](rfs-coplot_bilirubin-1)
Comparing with conditional panels of , we see the overall response is similar to the `edema=0` response. The survival for `edema=0.5` is slightly lower, though the slope of the smooth indicates a similar relation to `bili`. The `edema=1` panel shows that the survival for this (smaller) group of patients is worse, but still follows the trend of decreasing with increasing `bili`.
Conditional membership within a continuous variable requires stratification at some level. We can sometimes make these stratification along some feature of the variable, for instance a variable with integer values, or 5 or 10 year age group cohorts. However with our variables of interest, there are no logical stratification indications. Therefore we arbitrarily stratify our variables into 6 groups of roughly equal population size using the `quantile_cuts` function. We pass the break points located by `quantile_cuts` to the `cut` function to create grouping intervals, which we can then add to the `gg_variable` object before plotting with the `plot.gg_variable` function. This time we use the `facet_wrap` function to generate the panels grouping interval, which automatically sorts the six panels into two rows of three panels each.
![Variable dependence coplot of survival at 1 year against bili, conditional on albumin interval group membership.[]{data-label="fig:albumin-coplot"}](rfs-albumin-coplot-1)
The `gg_variable` coplot of indicates that the effect of `bili` decreases conditional on membership within increasing `albumin` groups. To get a better feel for how the response depends on both these variables together, it is instructive to look at the compliment coplot of `albumin` conditional on membership in `bili` groups. We repeat the previous coplot process, predicted survival as a function of the `albumin` variable, conditional on membership within 6 groups `bili` intervals. As the code to create the coplot of is nearly identical to the code for creating .
R> \# Find intervals with similar number of observations. R> bili\_cts <-quantile\_pts(ggvar$bili, groups = 6, intervals = TRUE)
R>
R> # We need to move the minimal value so we include that observation
R> bili_cts[1] <- bili_cts[1] - 1.e-7
R>
R> # Create the conditional groups and add to the gg_variable object
R> bili_grp <- cut(ggvar$bili, breaks = bili\_cts) R> ggvar$bili_grp <- bili_grp
R>
R> # Adjust naming for facets
R> levels(ggvar$bili\_grp) <- paste(“bilirubin =”, levels(bili\_grp)) R> R> \# plot.gg\_variable R> plot(ggvar, xvar = “albumin”, alpha = 0.5) + R+ \# method = “glm”, se = FALSE) + R+ labs(y = “Survival”, x = st.labs\[“albumin”\]) + R+ theme(legend.position = “none”) + R+ scale\_color\_manual(values = strCol, labels = event.labels) + R+ scale\_shape\_manual(values = event.marks, labels = event.labels) + R+ facet\_wrap( bili\_grp) + R+ coord\_cartesian(ylim = c(-0.01,1.01))
![Variable dependence coplot of survival at 1 year against albumin, conditonal on bili interval group membership.[]{data-label="fig:bili-coplot"}](rfs-bili-coplot-1)
The `gg_variable` coplot of indicates the probability of survival increases with increasing `albumin` and increases within groups of increasing `bili`.
Typically, conditional plots for continuous variables include overlapping intervals along the grouped variable [@cleveland:1993]. We chose to use mutually exclusive continuous variable intervals for the following reasons:
- Simplicity - We can create the coplot figures directly from the `gg_variable` object by adding a conditional group column directly to the object.
- Interpretability - We find it easier to interpret and compare the panels if each observation is only in a single panel.
- Clarity - We prefer using more space for the data portion of the figures than typically displayed in the `coplot` function which requires the bar plot to present the overlapping segments.
It is still possible to augment the `gg_variable` to include overlapping conditional membership with continuous variables by duplicating rows of the training set data within the `rfsrc$xvar` object, and then setting the conditional group membership as described. The `plot.gg_variable` function recipe above could be used to generate the panel plot, with panels ordered according to the factor levels of the grouping variable. We leave this as an exercise for the reader.
Partial dependence coplots
--------------------------
By characterizing conditional plots as stratified variable dependence plots, the next logical step would be to generate an analogous conditional partial dependence plot. The process is similar to variable dependence coplots, first determine conditional group membership, then calculate the partial dependence estimates on each subgroup using the function with a argument for each grouped interval. The function is a wrapper for generating conditional partial dependence data objects. Given a random forest () object and a vector for conditioning the training data set observations, calls the function the training set observations conditional on membership. The function returns a object, a subclass of the `gg_partial` object, which can be plotted with the function.
The following code block will generate the data object for creating partial dependence coplot of 1 year survival as a function of `bili` conditional on membership within the 6 groups of `albumin` intervals that we examined in the .
R> partial\_coplot\_pbc <- gg\_partial\_coplot(rfsrc\_pbc, xvar = “bili”, R+ groups = ggvar$albumin_grp,
R+ surv_type = "surv",
R+ time = rfsrc_pbc$time.interest\[time\_index\[1\]\], R+ show.plots = FALSE) R> R> ggplot(partial\_coplot\_pbc, aes(x=bili, y=yhat, col=group, shape=group)) + R+ geom\_smooth(se = FALSE) + R+ labs(x = st.labs\[“bili”\], y = “Survival at 1 year (R+ color = ”albumin“, shape = ”albumin")
![Partial dependence coplot of survival at 1 year against bili, conditional on albumin interval group membership. Points estimates with loess smooth to indicate trend within each group.[]{data-label="fig:bili-albumin"}](rfs-bili-albumin-1)
The of shows point estimates of the risk adjusted survival as a function of `bili` conditional on group membership defined by `albumin` intervals. The figure is slightly different than the `gg_partial` plot of as each set of partial dependence estimates is calculated over a subset of the training data. We again connect the point estimates with a Loess curve.
For completeness, we construct the compliment coplot view of one year survival as a function of `albumin` conditional on `bili` interval group membership in .
![Partial dependence coplot of survival at 1 year against albumin, conditional on bili interval group membership. Points estimates with loess smooth to indicate trend within each group.[]{data-label="fig:albumin-bili"}](rfs-albumin-bili-1)
Partial plot surfaces
=====================
Just as in partial dependence, we can view the partial coplot curves as slices along a surface that could extend along an axis into the page. This visualization is made a bit difficult by our choice to select groups of similar population size, as the curves are not evenly spaced along the grouping variables. So, similar to the partial dependence surface we created along time in , we can examine the relation of these two variables using a partial dependence surface.
A difficulty with conditional dependence for this exercise is the reduction of the sample sizes for calculating a coplot surface. So instead, we calculate the full partial dependence surface by generating 50 `albumin` values spaced evenly along the data distribution. For each value of `albumin`, we calculate the partial dependence on `bili` at `npts = 50` points with the `plot.variable` function. We generate the surface again using the `surf3D` function.
![Partial dependence surface of survival at 1 year as a funtion of bili and albumin. Blue lines indicate partial coplot cut points for albumin () and bili ().[]{data-label="fig:surface3d"}](rfs-surface3d-1)
The partial dependence surface of shows partial dependence of 1 year survival on the Z-axis against values of Bilirubin and Albumin. We again use linear interpolation between the 2500 estimates, and color the surface by the response. Here blue corresponds to lower and yellow to higher risk adjusted survival. The blue lines are placed at the cut points between groups of `albumin` and `bili` used in the partial coplots of Figures \[fig:bili-albumin\] and \[fig:albumin-bili\] respectively.
To construct the partial coplot for groups of `albumin` in , we arbitrarily segmented the training set into 6 groups of equal membership size. The segments between blue lines parallel to the Bilirubin axis indicate where on the surface these observations are located. Similarly, the blues lines perpendicular to the Bilirubin axis segment observations into the 6 groups of `bili` intervals. indicates the arbitrary grouping for groups of `bili` in .
The figure indicates that partial dependence of higher `albumin` levels are similar, which results in the over plotting seen in . The distribution is sparser at lower `albumin` levels, creating the larger area in lowest `albumin` values, where the partial dependence changes the most.
Conclusion
==========
In this vignette, we have demonstrated the use of Random Survival Forest methods with the \~(<http://CRAN.R-project.org/package=ggRandomForests>) package. We have shown how to grow a random forest model and determine which variables contribute to the forest prediction accuracy using both VIMP and Minimal Depth measures. We outlined how to investigate variable associations with the response variable using variable dependence and the risk adjusted partial dependence plots. We’ve also explored variable interactions by using pairwise minimal depth interactions and directly viewed these interactions using variable dependence coplots and partial dependence coplots. Along the way, we’ve demonstrated the use of additional commands from the package [@Wickham:2009 <http://CRAN.R-project.org/package=ggplot2>] package for modifying and customizing plots from functions.
Computational details
---------------------
This document is a package vignette for the package for “Visually Exploring Random Forests” (<http://CRAN.R-project.org/package=ggRandomForests>). The package is designed for use with the package [@Ishwaran:RFSRC:2014 <http://CRAN.R-project.org/package=randomForestSRC>] for growing survival, regression and classification random forest models and uses the package [@Wickham:2009 <http://CRAN.R-project.org/package=ggplot2>] for plotting diagnostic and variable association results. is structured to extract data objects from objects and provides functions for printing and plotting these objects.
The vignette is a tutorial for using the package with the package for building and post-processing random survival forests. In this tutorial, we explore a random forest for survival model constructed for the primary biliary cirrhosis (PBC) of the liver data set [@fleming:1991], available in the package. We grow a random survival forest and demonstrate how can be used when determining how the survival response depends on predictive variables within the model. The tutorial demonstrates the design and usage of many of functions and features and also how to modify and customize the resulting `ggplot` graphic objects along the way.
The vignette is written in LaTeXusing the package [@Xie:2015; @Xie:2014; @Xie:2013 <http://CRAN.R-project.org/package=knitr>], which facilitates weaving [@rcore] code, results and figures into document text.
This vignette is available within the package on the Comprehensive R Archive Network (CRAN) [@rcore <http://cran.r-project.org>]. Once the package has been installed, the vignette can be viewed directly from within with the following command:
R> vignette(“randomForestSRC-Survival”, package = “ggRandomForests”)
A development version of the package is also available on GitHub (<https://github.com>). We invite comments, feature requests and bug reports for this package at <https://github.com/ehrlinger/ggRandomForests>.
Acknowledgement
---------------
This work was supported in part by the National Institutes of Health grant R01-HL103552-01A1.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: |
We start by studying the Goldberger-Treiman discrepancy (GTd) $\Delta =
(2.259\pm 0.591)\%$. Then we look at the $\pi N$ $\sigma$ term, with the dimensionless ratio $\sigma_N/2m_N=3.35\%$. Finally we return to predicting (via the quark model) the $\pi N$ coupling constant, with GTd $\Delta\to 0$ as $m_q\to m_N/3$.
---
=16truecm =23.5truecm -1.5cm -1.2cm =24.pt
**PION NUCLEON COUPLING CONSTANT, GOLDBERGER-TREIMAN**
**DISCREPANCY AND $\pi N$ $\sigma$ TERM**
0.5cm
Miroslav Nagy[^1]
*Institute of Physics, Slovak Academy of Sciences, 845 11 Bratislava, Slovakia*
Michael D. Scadron[^2]
*Physics Department, University of Arizona, Tucson, Arizona 85721,USA*
Gerald E. Hite[^3]
*Texas A&M University at Galveston, TX 77553, USA*
1.0cm
0.5truecm PACS: 12.39.Ki, 12.39.Mk, 13.25.-k, 14.40.-n 1.0truecm Given the recent new value of the $\pi NN$ coupling constant [@Bugg] $$g^2_{\pi NN}/4\pi = 13.80\pm 0.12 ~~~{\rm or}~~~ g_{\pi NN}=13.169\pm
0.057, \label{g^2}$$ along with the observed axial current coupling [@PDG] $$g_A = 1.267 \pm 0.004, \label{g_A}$$ combined with the measured pion decay constant [@PDG] $$f_{\pi} = (92.42\pm 0.26) {\rm MeV}, \label{f_pi}$$ the Goldberger-Treiman discrepancy (GTd) is then $$\Delta = 1 - \frac{m_Ng_A}{f_{\pi}g_{\pi NN}} = (2.259\pm 0.591)\% .
\label{Delta}$$ Here we have used the mean nucleon mass $m_N$= 938.9 MeV and have computed the overall mean square error.
To verify this GTd in Eq.(\[Delta\]), we employ the constituent quark loop with imaginary part [@Coon] $${\rm Im} f_{\pi}(q^2) = \frac{3g_{\pi qq}}{2}\frac{4{\hat
m}}{8\pi}\left(1 -\frac{4{\hat m}^2}{q^2} \right)^{1/2}\Theta(q^2 -
4{\hat m}^2) . \label{Imag}$$ This follows from unitarity with the inclusion of a factor of 3 from colour. Following ref. [@Coon] using the quark level Goldberger-Treiman relation $f_{\pi}g_{\pi qq} = {\hat m}$, the GTd to fourth order in $q'^2$ predicts for a once-subtracted dispersion relation assuming a quark-level GTR: $$\frac{f_{\pi}(q^2) - f_{\pi}(0)}{f_{\pi}(0)} = \frac{q^2}{\pi}
\int_{4{\hat m}^2}^{\infty}\frac{dq'^2\left(1-\frac{4{\hat m}^2}{q'^2}
\right)^{1/2}}{q'^2(q'^2-q^2)}\frac{3g^2_{\pi qq}}{4\pi} , \label{frac}$$ or for $q^2=m_{\pi}^2$, the integral in Eq.(\[frac\]) gives a discrepancy for $f_{\pi}$ $${\bar \Delta} = \frac{f_{\pi}(m^2_{\pi})}{f_{\pi}(0)} - 1 =
\frac{3g^2_{\pi qq}}{2\pi^2}\left[1-r{\rm tan}^{-1}\left(\frac{1}{r}
\right)\right] \label{Delta1}$$ for $r^2=\frac{4{\hat m}^2}{m^2_{\pi}}-1\geq 0$. Since $m^2_{\pi}/
4{\hat m}^2\ll 1$, a Taylor series expansion leads to $$1-r{\rm tan}^{-1}\left(\frac{1}{r}\right) = \frac{1}{3r^2}-
\frac{1}{5r^4}+...= \frac{m^2_{\pi}}{12{\hat m}^2}\left(1+
\frac{1}{10}\frac{m^2_{\pi}}{{\hat m}^2}+...\right)$$ and a discrepancy [^4] $${\bar \Delta} = \frac{f_{\pi}(m^2_{\pi})}{f_{\pi}(0)} - 1 =
\frac{m^2_{\pi}}{8\pi^2f_{\pi}^2}\left(1+\frac{1}{10}
\frac{m^2_{\pi}}{{\hat m}^2}\right) \approx 2.946\% .
\label{Delta'}$$ The first term on the rhs is independent of ${\hat m}$, while in the small second term we take ${\hat m}= m_N/3$. This then leads to a net 2.946% correction in Eq.(\[Delta’\]).
Since the physical GT relation becomes exact $(f_{\pi}g_{\pi NN}=
m_Ng_A)$ when $m_{\pi}\to 0$ for a conserved axial current, it should not be surprising that the measured GTd in Eq.(\[Delta\]) of $(2.259
\pm 0.591)\%$ is within 1.16 standard deviations from the dispersion-theoretical $\bar{\rm GTd}$ ${\bar \Delta} = 2.946 \%$ in Eq.(\[Delta’\]). Appreciate that $g_A$ is measured at $q^2 = 0$ while $f_{\pi}$ is measured at $q^2 = m^2_{\pi}$ but $f_{\pi}(0)$ is inferred at $q^2=0$ via Eq.(\[Delta’\]).
Just as the chiral-breaking $SU(2)$ GTd is 2–3%, the $SU(2)\times SU(2)$ $\pi N$ $\sigma$ term of 63 MeV corresponds to a dimensionless ratio of about 3%: $$\frac{\sigma_N}{2m_N}=\frac{63~{\rm MeV}}{2\times 938.9~ {\rm MeV}}
\approx 3.35\% . \label{sigma}$$ Alternatively the chiral-limiting (CL) nucleon mass is related to the $\pi N$ $\sigma$ term as [@GMcNS] $$m_N^2 = (m^{CL}_N)^2 +m_N\sigma_N,~~~ {\rm or~ with}~~
\sigma_N=63~{\rm MeV},
\label{mnmCL}$$ $$\frac{m_N}{m_N^{CL}} - 1 = 3.53\%,~~~ {\rm with}~~
m_N^{CL}=906.85~{\rm MeV}. \label{fracmn}$$ Note the many $3\%$ CL relations in Eqs. (\[Delta\]),(\[Delta’\]),(\[sigma\]),(\[fracmn\]) above. Now we justify the $\sigma$ term $\sigma_N=63$ MeV.
The explicit $SU(2)\times SU(2)$ chiral-breaking $\sigma$ term is the sum of the perturbative GMOR [@GMOR] or quenched APE [@APE] part $$\sigma_N^{GMOR} = (m_{\Xi} + m_{\Sigma} - 2m_N) \frac{m^2_{\pi}}{m^2_K
- m^2_{\pi}} = 26~ {\rm MeV}, \label{GMOR}$$ $$\sigma_N^{APE} = (24.5 \pm 2)~ {\rm MeV}, \label{APE}$$ plus the nonperturbative linear $\sigma$ model (L$\sigma$M) nonquenched part [@GML] due to $\sigma$ tadpoles for the chiral-broken $m^2_{\pi}$ and $\sigma_N$, with ratio predicting $$\sigma_N^{L\sigma M} = \left(\frac{m_{\pi}}{m_{\sigma}}\right)^2 m_N
\approx 40~ {\rm MeV} \label{LsM}$$ for $m_{\sigma}\approx 665$ MeV [@SKN], a model-independent coupled channel dispersion theory and parameter-free relation. Specifically, Eq.(\[LsM\]) stems from semi-strong L$\sigma$M tadpole graphs generating $\sigma_N$ and $m^2_{\pi}$. Their ratio cancels out the $\langle\sigma|H_{ss}|0\rangle$ factor. The L$\sigma$M couplings $2g_{\sigma\pi\pi}$=$m_{\sigma}^2/f_{\pi}$ and $f_{\pi}g_{\sigma
NN}=m_N$ then give $\sigma^{L\sigma M}_N$ = $(m_{\pi}/m_{\sigma})^2m_N$ as found in Eq.(\[LsM\]). Since the $\sigma(600)$ has been observed [@PDG], with a broad width, but the central model-independent value [@SKN] is known to be 665 MeV, the chiral L$\sigma$M mass ratio in Eq.(\[LsM\]) is expected to be quite accurate - while being free of model-dependent parameters. The authors of [@TaR] find the $\sigma$ meson between 400 MeV and 900 MeV, with the average mass 650 MeV near 665 MeV from [@SKN]. Then the sum of (\[GMOR\],\[APE\]) plus (\[LsM\]) is $$\sigma_N = \sigma_N^{GMOR,APE} + \sigma_N^{L\sigma M}\approx (25+40)
~{\rm MeV}= 65 ~{\rm MeV}. \label{sigmaN}$$ Rather than add the perturbative plus nonperturbative parts as in Eq.(\[sigmaN\]), one can instead work in the infinite momentum frame (IMF) requiring squared masses [@MDS] and only one term (tadpole terms ${}\to 0$ in the IMF) [@CSS] $$\sigma_N^{IMF} = \frac{m^2_{\Xi}+m^2_{\Sigma}-2m^2_N}{2m_N}\left(
\frac{m^2_{\pi}}{m^2_K-m^2_{\pi}}\right) = 63~ {\rm MeV}. \label{IMF}$$ Note that Eqs.(\[sigmaN\]) and (\[IMF\]) are both very near the observed value [@Koch] $(65\pm 5)$ MeV.
With hindsight, we can also deduce the $\pi N$ $\sigma$ term via PCAC (partially conserved axial current) at the Cheng-Dashen (CD) point [@ChDa] with background isospin-even $\pi$N amplitude $${\bar F}^+(\nu=0,t=2m^2_{\pi})= \sigma_N/f^2_{\pi} + O(m^4_{\pi}).
\label{F+nu}$$ At this CD point, a recent Karlsruhe data analysis by G. Höhler [@Koch] finds $${\bar F}^+(0,2m^2_{\pi})= \sigma_N/f^2_{\pi} + 0.002m^{-1}_{\pi}=
1.02m^{-1}_{\pi}, \label{F2mpi}$$ implying $\sigma_N=63$ MeV for $f_{\pi}=93$ MeV, $m_{\pi}=139.57$ MeV.
We can unify the earlier parts of this paper by first inferring from Eq.(\[Delta’\]) the chiral limit (CL) pion decay constant $$f^{CL}_{\pi} = f_{\pi}/1.029 46\approx 89.775~{\rm MeV} \label{fpiCL}$$ using Eq.(\[Delta’\]) and the observed [@PDG] $f_{\pi}=(92.42\pm
0.26)$ MeV. Then the quark-level GTr using the meson-quark coupling $g=2\pi/\sqrt 3$ [@Elias] predicts the nonstrange quark mass in the CL as $${\hat m}^{CL} = f_{\pi}^{CL}g = 325.67 ~{\rm MeV}, \label{mCL}$$ close to the expected ${\hat m}^{CL} = m_N/3\approx 313$ MeV. This in turn predicts the scalar $\sigma$ mass in the CL as [@GML; @NJL] $$m^{CL}_{\sigma}= 2{\hat m}^{CL} = 651.34~ {\rm MeV} \label{msCL}$$ and then the on-shell L$\sigma$M $\sigma$ mass obeys $$m^2_{\sigma} - m^2_{\pi} = (m^{CL}_{\sigma})^2\approx (651.34~{\rm
MeV})^2 ~~~{\rm or}~~~ m_{\sigma}\approx 665.76~{\rm MeV}, \label{ms2}$$ almost exactly the model-independent $\sigma$ mass found in ref. [@SKN], also predicting $\sigma_N^{L\sigma M}$ in Eq.(\[LsM\]).
In this letter we have linked the GT discrepancy Eqs.(\[Delta\]),(\[Delta’\]) and the $\pi N$ $\sigma$ term Eqs.(\[sigmaN\]),(\[IMF\]) with the L$\sigma$M values Eqs.(\[fpiCL\])-(\[ms2\]). The predicted L$\sigma$M value of $g_{\pi
NN}$ is $$g_{\pi NN} = N_c gg_A= 3(2\pi/\sqrt3)1.267\approx 13.79, \label{gpiNN}$$ near the observed value in Eq.(\[g\^2\]) with meson-quark coupling g. Substituting Eq.(\[gpiNN\]) into the GTd (Eq.(\[Delta\])) in turn predicts in the quark model $$\Delta = 1 - \frac{m_N}{3m_q}\to 0~~~{\rm as}~~~ m_q\to m_N/3.
\label{DmN}$$ However meson-baryon couplings for pseudoscalars (P), axial-vectors (A) and $SU(6)$-symmetric states are known [@MMN] to obey $$(d/f)_P\approx 2.0, ~~~ (d/f)_A\approx 1.74, ~~~ (d/f)_{SU(6)}
= 1.50, \label{SU6}$$ where the scales of $d,f$ characterize the symmetric, antisymmetric $SU(3)$ structure constants. Note that the ratio remains the same: $$\frac{(d/f)_A}{(d/f)_P}=\frac{1.74}{2.0}=0.87,~~~
\frac{(d/f)_{SU(6)}}{(d/f)_A}=\frac{1.50}{1.74}\approx 0.86.
\label{dfAP}$$ Thus to predict the quark-based $\pi NN$ coupling constant we weight Eq.(\[gpiNN\]) by the scale factor of Eq.(\[dfAP\]) in order to account for the $SU(6)$ quark content of $g_A$: $$g_{\pi NN}= 3\times 2\pi/{\sqrt 3}\times 1.267\times 0.87\approx 12.00
\label{gpiNN'}$$ and this predicted coupling constant is near 13.169 from ref. [@Bugg], or 13.145 from ref. [@AWP], or nearer still to 13.054 from ref. [@STS]. One could alter this 0.87 reduction of $g_A$ in Eq.(\[gpiNN’\]) by using the quark-based factor 3/5=0.6, where the $SU(6)$ factor for $g_A$ of 5/3 becomes inverted for quarks as suggested in [@TNP]. In any case the predicted $\pi NN$ coupling lies between 12.00 and 13.79 in Eqs.(\[gpiNN’\]),(\[gpiNN\]), midway near the recent data in Eq.(\[g\^2\]).
In passing, we note that the large model-independent [@SKN] scalar $\sigma$ mass of $m_{\sigma}\approx 665$ MeV is recovered via the L$\sigma$M combined with the CL quark-level GTR Eqs.(\[fpiCL\])-(\[ms2\]). Also the large almost model-independent interior dispersion relation version of the $\pi N~
\sigma$ term [@HJS; @KH] is between 65-80 MeV. While this $\sigma$ term follows from the two GMOR + L$\sigma$M terms in Eq.(\[sigmaN\]) or from the IMF term in Eq.(\[IMF\]), original chiral perturbation theory (ChPT) of the 1970s suggested [@HP] $\sigma_N\approx 25$ MeV near the GMOR value.
Modern ChPT now predicts [@GLS] a $\sigma_N$ of 45 MeV at $t=0$ extended up to the above presumably measured value of 60 MeV according to [@GLS; @HL] $$60~{\rm MeV} = \sigma^{\rm GMOR}_N(25~{\rm MeV}) + \sigma_N^{\rm
higher~order~ChPT}(10~{\rm MeV}) + \sigma_N^{\rm t-dep.}(15~{\rm MeV})
+ \sigma_N^{{\bar s}s}(10~{\rm MeV}) \label{60MeV}$$ and the latter ”three pieces happen to have the same sign as $\sigma_N^{\rm GMOR}$” [@HL].
In summary, as $m_{\pi}\to 0$, $\partial A_{\pi}\to 0$, the quark-level GT relation requires the observed $2-3\%$ GTd and $3\%$ $\sigma$ term ratio to predict $g_{\pi NN}$, with $\Delta\to 0$ as $m_q\to m_N/3$ or ${\bar \Delta}\to 0$ when $m^2_{\pi}\to 0$. We have computed the $\pi N$ $\sigma$ term in many. different ways to find approximately $\sigma_N
=63$ MeV.
Acknowledgements: One of us (MDS) appreciates D.V. Bugg’s help in first obtaining Eq.(\[Delta’\]) via a computer. Nonetheless the reader may be more convinced with Eq.(\[Delta’\]) using the Dwight integral solution combined with the Taylor series in footnote 4. This work is in part supported by the Slovak Agency for Science, Grant 2/3105/23.
[99]{} D.V. Bugg: [*Eur. Phys. J.*]{} C [**33**]{} (2004) 505 Particle Data Group (PDG), K. Hagiwara et al.: [*Phys. Rev.*]{} D [**66**]{} (2002) 1 S.A. Coon, M.D. Scadron: [*Phys. Rev.*]{} C [**23**]{} (1981) 1150 J.F. Gunion, P.C. McNamee, M.D. Scadron: [*Phys. Lett.*]{} B [**63**]{} (1976) 81; [*Nucl. Phys.*]{} B [**123**]{} (1977) 445 M. Gell-Mann, R. Oakes, B. Renner \[GMOR\]: [*Phys. Rev.*]{} [**175**]{} (1968) 2195 APE Collab., S. Cabasino et al.: [*Phys. Lett.*]{} B [**258**]{} (1991) 202 For L$\sigma$M see e.g. M. Gell-Mann, M. Levy: [*Nuovo Cim.*]{} [**16**]{} (1960) 706. For quark-level L$\sigma$M see R. Delbourgo, M.D. Scadron: [*Mod. Phys. Lett.*]{} A [**10**]{} (1995) 251, hep-ph/9910242. For a recent summary, see M.D. Scadron, G. Rupp, F. Kleefeld, E. van Beveran: [*Phys. Rev.*]{} D [**69**]{} (2004) 014010, hep-ph/0309109 Yu.S. Surovtsev, D. Krupa, M. Nagy: [*Eur. Phys. J.*]{} A [**15**]{} (2002) 409 N. Törnqvist, M. Roos: [*Phys. Rev. Lett.*]{} [**76**]{} (1996) 1575 For IMF see e.g. M.D. Scadron: [*Mod. Phys. Lett.*]{} A[**7**]{} (1992) 669 G. Clement, M.D. Scadron, J. Stern: [*Jour. Phys.*]{} G [**17**]{} (1991) 199, [*Zeit. Phys.*]{} C [**60**]{} (1993) 307; S.A. Coon, M.D. Scadron: [*Jour. Phys.*]{} G [**18**]{} (1992) 1923; S. Fubini, G. Furlan: [*Physics*]{} [**1**]{} (1965) 229 R. Koch: [*Zeit. Phys.*]{} C[**15**]{} (1982) 161; G. Höhler: $\pi N$ Scattering, (Springer 1983) T.P. Cheng, R.F. Dashen: [*Phys. Rev. Lett.*]{} [**26**]{} (1971) 594 For $g=2\pi/\sqrt 3$, see infrared QCD, V. Elias, M.D. Scadron: [*Phys. Rev. Lett.*]{} [**53**]{} (1984) 1129; L.R. Baboukhadia, V. Elias, M.D. Scadron: [*J. Phys.*]{} G [**23**]{} (1997) 1065. Also see L$\sigma$M ref. [@GML] and the $Z=0$ compositeness condition A. Salam: [*Nuovo Cimento*]{} [**25**]{} (1962) 234; S. Weinberg: [*Phys. Rev.*]{} [**130**]{} (1963) 776; M.D. Scadron: [*Phys. Rev.*]{} D [**57**]{} (1998) 5307 Y. Nambu, G. Jona-Lasinio: [*Phys. Rev.*]{} [**122**]{} (1961) 345 M.M. Nagels et al.: [*Nucl. Phys.*]{} B [**109**]{} (1979) 1; Z. Dziembowski, J. Franklin: [*J. Phys.*]{} G [**17**]{} (1991) 213; F.E. Close: [*Phys. Rev. Lett.*]{} [**64**]{} (1990) 361; R.L. Jaffe, A.V. Manohar: [*Nucl. Phys.*]{} B [**337**]{} (1990) 509 R.A. Arndt, R.L. Workman, M. Pavan: [*Phys. Rev.*]{} C [**49**]{} (1994) 2729 V. Stoks, R. Timmermans, J.J. deSwart: [*Phys. Rev.*]{} C [**47**]{} (1993) 512 T.N. Pham: [*Phys. Rev.*]{} D [**25**]{} (1982) 2955 G.E. Hite, R.J. Jacob, M.D. Scadron: [*Phys. Rev.*]{} D [**14**]{} (1976) 1306 W.B. Kaufmann, G.E. Hite: [*Phys. Rev.*]{} C [**60**]{} (1999) 055204. Also see [*TRIUMF Conf. MENU 1997, $\pi N$ Newsletter,*]{} W.B. Kaufmann, G.E. Hite: p.16, M.D. Scadron: p. 362 H. Pagels: [*Phys. Rep.*]{} C [**16**]{} (1975) 219 J. Gasser, H. Leutwyler, M. Sainio: [*Phys. Lett.*]{} B [**253**]{} (1991) 260. However the ChPT paper of J. Gasser, H. Leutwyler: [*Ann. Phys.*]{} (NY) [**158**]{} (1984) 142 attempts to rule out the L$\sigma$M in App. B. H. Leutwyler: [*XXVI Int. Conf. on High Energy Physics*]{}, Dallas 1992. The 15 MeV in Eq.(\[60MeV\]) was found not to occur in our c-number $\sigma_N$ analysis, since the t-dependence is minimal in Eqs.(\[F+nu\]),(\[F2mpi\]). The latter $\sigma_N^{{\bar
s}s}$ of 10 MeV in (\[60MeV\]) is 8 MeV higher than the $y\leq 6$ % value in refs. [@FR] with $\sigma_N^{{\bar s}s}$ = $(1-y)^{-1}\sigma_N^{\rm GMOR}$ - $\sigma_N^{\rm GMOR}\leq 2$ MeV. Lastly the $\sigma_N^{\rm ChPT}$ correction of 10 MeV in (\[60MeV\]) does not apply to our L$\sigma$M $\sigma_N$ addition to $\sigma_N^{\rm
GMOR}$. C. Foudas et al.: [*Phys. Rev. Lett.*]{} [**64**]{} (1990) 1207; S.A. Rabinowitz et al.: [*ibid.*]{} [**70**]{} (1993) 134
[^1]: E-mail address: [email protected]
[^2]: E-mail address: [email protected]
[^3]: E-mail address: [email protected]
[^4]: From Dwight Integral tables, Eq.(\[Delta1\]) above stems from Eq.122.1 on p.31, and the needed Taylor series of Eq.505.1, p.118: ${\rm tan}^{-1}x=x-\frac{x^3}{3}
+\frac{x^5}{5}+...,$ for $x<1$.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'We study the almost-sure reachability problem in a distributed system obtained as the asynchronous composition of $N$ copies (called processes) of the same automaton (called protocol), that can communicate via a shared register with finite domain. The automaton has two types of transitions: write-transitions update the value of the register, while read-transitions move to a new state depending on the content of the register. Non-determinism is resolved by a stochastic scheduler. Given a protocol, we focus on almost-sure reachability of a target state by one of the processes. The answer to this problem naturally depends on the number $N$ of processes. However, we prove that our setting has a cut-off property: the answer to the almost-sure reachability problem is constant when $N$ is large enough; we then develop an [aux auxempty$\mathsf{\mathit{}}$EXPSPACE]{}algorithm deciding whether this constant answer is positive or negative.'
author:
- Patricia Bouyer
- Nicolas Markey
- 'Mickael Randour[^1]'
- Arnaud Sangnier
- Daniel Stan
title: 'Reachability in Networks of Register Protocols under Stochastic Schedulers[^2][^3]'
---
Introduction
============
##### Verification of systems with many identical processes.
It is a classical pattern in distributed systems to have a large number of identical components running concurrently (a.k.a. networks of processes). In order to verify the correctness of such systems, a naive option consists in fixing an upper bound on the number of processes, and applying classical verification techniques on the resulting system. This has several drawbacks, and in particular it gives no information whatsoever about larger systems. Another option is to use parameterized-verification techniques, taking as a parameter the number of copies of the protocol in the system being considered. In such a setting, the natural question is to find and characterize the set of parameter values for which the system is correct. Not only the latter approach is more general, but it might also turn out to be easier and more efficient, since it involves orthogonal techniques.
##### Different means of communication lead to different models.
A seminal paper on parameterized verification of such distributed systems is the work of German and Sistla [@GS92]. In this work, the authors consider networks of processes all following the same finite-state automaton; the communication between processes is performed thanks to *rendez-vous* communication. Various related settings have been proposed and studied since then, which mainly differ by the way the processes communicate. Among those, let us mention broadcast communication [@EFM99; @DSZ10], token-passing [@CTTV04; @AJKR14], message passing [@BGS14], shared register with ring topologies [@ABG15], or shared memory [@EGM13]. In his nice survey on such parameterized models [@Esp14], Esparza shows that minor changes in the setting, such as the presence of a controller in the system, might drastically change the complexity of the verification problems. The relative expressiveness of some of those models has been studied recently in [@ARZ15], yielding several reductions of the verification problems for some of those classes of models.
##### Asynchronous shared-memory systems.
We consider a communication model where the processes asynchronously access a shared register, and where read and write operations on this register are performed non-atomically. A similar model has been proposed by Hague in [@Hag11], where the behavior of processes is defined by a pushdown automaton. The complexity of some reachability and liveness problems for shared-memory models have then been established in [@EGM13] and [@DEGM15], respectively. These works consider networks in which a specific process, called the leader, runs a different program, and address the problem whether, for some number of processes, the leader can satisfy a given reachability or liveness property. In the case where there is no leader, and where processes are finite-state, the parameterized control-state reachability problem (asking whether one of the processes can reach a given control state) can be solved in polynomial time, by adapting the approach of [@DSTZ12] for lossy broadcast protocols.
##### Fairness and cut-off properties.
=-1 In this work, we further insert fairness assumptions in the model of parameterized networks with asynchronous shared memory, and consider reachability problems in this setting. There are different ways to include fairness in parameterized models. One approach is to enforce fairness expressed as a temporal-logic properties on the executions (e.g., any action that is available infinitely often must be performed infinitely often); this is the option chosen for parameterized networks with rendez-vous [@GS92] and for systems with disjunctive guards (where processes can query the states of other processes) in [@AJK16]. We follow another choice, by equipping our networks with a stochastic scheduler that, at each step of the execution, assigns the same probability to the available actions of all the processes. From a high-level perspective, both forms of fairness are similar. However, expressing fairness via temporal logic allows for very regular patterns (e.g., round-robin execution of the processes), whereas the stochastic approach leads to consider all possible interleavings with probability $1$. Under this stochastic scheduler assumption, we focus on almost-sure reachability of a given control state by any of the processes of the system. More specifically, as in [@AJK16], we are interested in determining the existence of a *cut-off*, i.e., an integer $k$ such that networks with more than $k$ processes almost-surely reach the target state. Deciding the existence and computing such cut-offs is important for at least two aspects: first, it ensures that the system is correct for arbitrarily large networks; second, if we are able to derive a bound on the cut-off, then using classical verification techniques we can find the exact value of the cut-off and exactly characterize the sizes of the networks for which the behavior is correct.
##### Our contributions.
We prove that for finite-state asynchronous shared-memory protocols with a stochastic scheduler, and for almost-sure reachability of some control state by some process of the network, there always exists a positive or negative cut-off; positive cut-offs are those above which the target state is reached with probability $1$, while negative cut-offs are those above which the target state is reached with probability strictly less than $1$. Notice that both cut-offs are not complement of one another, so that our result is not trivial.
We then prove that the “sign” (positive or negative) of a cut-off can be decided in [aux auxempty$\mathsf{\mathit{}}$EXPSPACE]{}, and that this problem is -hard. Finally, we provide lower and upper bounds on the values of the cut-offs, exhibiting in particular protocols with exponential (negative) cut-off. Notice how these results contrast with classical results in related areas: in the absence of fairness, reachability can be decided in polynomial time, and in most settings, when cut-offs exist, they generally have polynomial size [@AJK16; @EN03; @EK00].
Presentation of the model and of the considered problem
=======================================================
Preliminaries.
--------------
Let $S$ be a finite set. A multiset over $S$ is a mapping $\mu\colon S \to
\bbN$. The cardinality of a multiset $\mu$ is $\Msize\mu=\sum_{s\in S}\mu(s)$. The support ${\ensuremath{\overline{\vphantom{}\mu}\xspace}}$ of $\mu$ is the subset $\nu\subseteq S$ s.t. for all $s\in S$, it holds $s\in \nu$ if, and only if, $\mu(s)>0$. For $k\in\bbN$, we write $\bbN^S_k$ for the set of multisets of cardinality $k$ over $S$, and $\bbN^S$ for the set of all multisets over $S$. For any $s\in S$ and $k\in\bbN$, we write $s^k$ for the multiset where $s^k(s)=k$ and $s^k(s')=0$ for all $s'\not=s$. We may write $s$ instead of $s^1$ when no ambiguity may arise. A multiset $\mu$ is included in a multiset $\mu'$, written $\mu\Msubseteq \mu'$, if $\mu(s)\leq
\mu'(s)$ for all $s\in S$. Given two multisets $\mu$ and $\mu'$, their union $\mu\Mcup\mu'$ is still a multiset s.t. $(\mu\Mcup\mu')(s)=\mu(s) + \mu'(s)$ for all $s\in S$. Assuming $\mu\Msubseteq\mu'$, the difference $\mu'\Msetminus\mu$ is still a multiset s.t. $(\mu'\Msetminus\mu)(s)=\mu'(s)-\mu(s)$.
A quasi-order $\tuple{A,\preceq}$ is a *well quasi-order* (wqo for short) if for every infinite sequence of elements $a_1,a_2,\ldots$ in $A$, there exist two indices $i<j$ such that $a_i \preceq a_j$. For instance, for $n>0$, $\tuple{\bbN^n,\leq}$ (with lexicographic order) is a wqo. Given a set $A$ with an ordering $\preceq$ and a subset $B \subseteq A$, the set $B$ is said to be *upward closed* in $A$ if for all $a_1 \in B$ and $a_2 \in A$, in case $a_1 \preceq a_2$, then ${a_2 \in B}$. The *upward-closure* of a set $B$ (for the ordering $\preceq$), denoted by ${\def\@tempa{\preceq}\ifx\@tempa\@empty\relax
\ensuremath{\mathord{\uparrow}}\else
\ensuremath{\mathord{\uparrow_{\preceq}}}\fi}(B)$ (or sometimes ${\def\@tempa{}\ifx\@tempa\@empty\relax
\ensuremath{\mathord{\uparrow}}\else
\ensuremath{\mathord{\uparrow_{}}}\fi}(B)$ when the ordering is clear from the context), is the set ${\{a \in A \mid \exists b \in B \mbox{ s.t. } b \preceq a\}}$. If $\tuple{A,\preceq}$ is a wqo and $B$ is an upward closed set in $A$, there exists a finite set of minimal elements ${\{b_1,\ldots,b_k\}}$ such that $B={\def\@tempa{}\ifx\@tempa\@empty\relax
\ensuremath{\mathord{\uparrow}}\else
\ensuremath{\mathord{\uparrow_{}}}\fi}{{\{b_1,\ldots,b_k\}}}$.
Register protocols and associated distributed system.
-----------------------------------------------------
We focus on systems that are defined as the (asynchronous) product of several copies of the same protocol. Each copy communicates with the others through a single register that can store values from a finite alphabet.
A *register protocol* is given by $\Prot = \tuple{Q, D, q_0, T}$, where $Q$ is a finite set of control locations, $D$ is a finite alphabet of data for the shared register, $q_0\in Q$ is an initial location, $T\subseteq Q\times \{R,W\}\times D\times Q$ is the set of transitions of the protocol. Here $R$ means *read* the content of the shared register, while $W$ means *write* in the register.
In order to avoid deadlocks, it is required that each location has at least one outgoing transition. We also require that whenever some $R$-transition $(q,R,d',q')$ appears in $T$, then for all $d\in D$, there exists at least one $q_d\in Q$ such that $(q,R,d,q_d)\in T$. The size of the protocol $\Prot$ is given by $|Q| + |T|$.
[running]{}\[ex1\] Figure \[fig-ex-protocol\] displays a small register protocol with four locations, over an alphabet of data $D=\{0,1,2\}$. In this figure (and in the sequel), omitted $R$-transitions (e.g., transitions $R(1)$ and $R(2)$ from $q_0$) are assumed to be self-loops. When the register contains $0$, this protocol may move from initial location $q_0$ to location $q_1$. From there it can write $1$ in the register, and then move to $q_2$. From $q_2$, as long as the register contains $1$, the process can either stay in $q_2$ (with the omitted self-loop $R(1)$), or write $2$ in the register and jump back to $q_1$. It is easily seen that if this process executes alone, it cannot reach state $q_f$.
(0,0) node\[rond,bleu\] (a) [$q_0$]{}; (2,0) node\[rond,vert\] (b) [$q_1$]{}; (4,0) node\[rond,jaune\] (c) [$q_2$]{}; (6,0) node\[rond,rouge\] (d) [$q_f$]{}; (a.180) – +(180:3mm); (a) edge\[-latex’\] node\[below\] [$R(0)$]{} (b); (b) edge\[-latex’,out=60,in=120,looseness=6\] node\[pos=.75,left\] [$W(1)$]{} (b); (b) edge\[-latex’,bend right\] node\[above\] [$R(1)$]{} (c); (c) edge\[-latex’,bend right\] node\[above\] [$W(2)$]{} (b); (c) edge\[-latex’\] node\[below\] [$R(2)$]{} (d); (d) edge\[-latex’,out=60,in=120,looseness=6\] node\[pos=.75,left\] [$W(2)$]{} (d);
We now present the semantics of distributed systems associated with our register protocols. We consider the *asynchronous* composition of several copies of the protocol (the number of copies is not fixed a priori and can be seen as a parameter). We are interested in the behavior of such a composition under a fair scheduler. Such distributed systems involve two sources of non-determinism: first, register protocols may be non-deterministic; second, in any configuration, all protocols have at least one available transition, and non-determinism arises from the asynchronous semantics. In the semantics associated with a register protocol, non-determinism will be solved by a randomized scheduler, whose role is to select at each step which process will perform a transition, and which transition it will perform among the available ones. Because we will consider qualitative objectives (almost-sure reachability), the exact probability distributions will not really matter, and we will pick the uniform one (arbitrary choice). Note that we assume non-atomic read/write operations on the register, as in [@Hag11; @EGM13; @DEGM15]. More precisely, when one process performs a transition, then all the processes that are in the same state are allowed to also perform the same transition just after, in fact write are always possible, and if a process performs a read of a specific value, since this read does not alter the value of the register, all processes in the same state can perform the same read (until one process performs a write). We will see later that dropping this hypothesis has a consequence on our results. We now give the formal definition of such a system.
The configurations of the distributed system built on register protocol $\Prot=\tuple{Q, D, q_0, T}$ belong to the set $\Confs=\bbN^Q \times D$. The first component of a configuration is a multiset characterizing the number of processes in each state of $Q$, whereas the second component provides the content of the register. For a configuration $\aconf=\tuple{\mu,d}$, we denote by ${st(\aconf)}$ the multiset $\mu$ in $\bbN^Q$ and by ${data(\aconf)}$ the data $d$ in $D$. We overload the operators defined over multisets; in particular, for a multiset $\delta$ over $Q$, we write $\aconf\Mcup\delta$ for the configuration $\tuple{\mu\Mcup\delta,d}$. Similarly, we write ${\ensuremath{\overline{\vphantom{}\aconf}\xspace}}$ for the support of ${st(\gamma)}$.
A configuration $\aconf'=\tuple{\mu',d'}$ is a *successor* of a configuration $\aconf=\tuple{\mu,d}$ if, and only if, there is a transition $(q,\textsf{op},d'',q')\in T$ such that $\mu(q)>0$, $\mu'=\mu\Msetminus q\Mcup
q'$ and either $\textsf{op}=R$ and $d=d'=d''$, or $\textsf{op}=W$ and $d'=d''$. In that case, we write $\aconf \rightarrow \aconf'$. Note that since $\mu(q)>0$ and $\mu'=\mu\Msetminus q\Mcup q'$, we have necessarily $\Msize\mu=\Msize{\mu'}$. In our system, we assume that there is no creation or deletion of processes during an execution, hence the size of configurations (i.e., $\Msize{{st(\aconf)}}$) remains constant along transitions. We write $\Gamma_k$ for the set of configurations of size $k$. For any configuration $\aconf\in\Gamma_k$, we denote by ${\mathrm{Post}}(\aconf)\subseteq
\Gamma_k$ the set of successors of $\aconf$, and point out that such a set is finite and non-empty.
Now, the *distributed system* $\Sys_\Prot$ associated with a register protocol $\Prot$ is a discrete-time Markov chain $\tuple{\Confs,{\mathit{Pr}}}$ where ${\mathit{Pr}}\colon \Gamma \times \Gamma \to [0,1]$ is the transition probability matrix defined as follows: for all $\aconf$ and $\aconf'\in \Confs$, we have ${\mathit{Pr}}(\aconf,\aconf')=\frac{1}{|{\mathrm{Post}}(\aconf)|}$ if $\aconf \rightarrow \aconf'$, and ${\mathit{Pr}}(\aconf,\aconf')=0$ otherwise. Note that ${\mathit{Pr}}$ is well defined: by the restriction imposed on the transition relation $T$ of the protocol, we have $0<|{\mathrm{Post}}(\aconf)| < \infty$ for all configuration $\aconf$, and hence we also get $\Sigma_{\aconf' \in \Confs}{\mathit{Pr}}(\aconf,\aconf')=1$. For a fixed integer $k$, we define the distributed system of size $k$ associated with $\Prot$ as the finite-state discrete-time Markov chain $\Sys_\Prot^k=\tuple{\Gamma_k,\Pr_k}$, where $\Pr_k$ is the restriction of $\Pr$ to $\Gamma_k\times\Gamma_k$.
We are interested in analyzing the behavior of the distributed system for a large number of participants. More precisely, we are interested in determining whether almost-sure reachability of a specific control state holds when the number of processes involved is large. We are therefore seeking a *cut-off* property, which we formalize in the following.
A finite path in the system $\Sys_\Prot$ is a finite sequence of configurations $\aconf_0
\rightarrow \aconf_1 \ldots \rightarrow \aconf_k$. In such a case, we say that the path starts in $\aconf_0$ and ends in $\aconf_k$. We furthermore write $\aconf \rightarrow^\ast \aconf'$ if, and only if, there exists a path that starts in $\aconf$ and ends in $\aconf'$. Given a location $q_f$, we denote by ${\llbracket \Diamond q_f \rrbracket}$ the set of paths of the form $\aconf_0
\rightarrow \aconf_1 \ldots \rightarrow \aconf_k$ for which there is $i \in
[0;k]$ such that ${st(\aconf_i)}(q_f)>0$. Given a configuration $\aconf$, we denote by ${\mathbb{P}}(\aconf,{\llbracket \Diamond q_f \rrbracket})$ the probability that some paths starting in $\aconf$ belong to ${\llbracket \Diamond q_f \rrbracket}$ in $\Sys_\Prot$. This probability is well-defined since the set of such paths is measurable (see e.g., [@PoMC2008-BK]). Given a register protocol $\Prot=\tuple{Q, D, q_0, T}$, an initial register value $d_0$, and a target location $q_f\in Q$, we say that $q_f$ is almost-surely reachable for $k$ processes if ${\mathbb{P}}(\tuple{q^k_0,d_0},{\llbracket \Diamond q_f \rrbracket})=1$.
[running]{}Consider again the protocol depicted in Fig. \[fig-ex-protocol\], with initial register content $0$. As we explained already, for $k=1$, the final state is not reachable at all, for any scheduler (here as $k=1$, the scheduler only has to solve non-determinism in the protocol).
When $k=2$, one easily sees that the final state is reachable: it suffices that both processes go to $q_2$ together, from where one process may write value $2$ in the register, which the other process can read and go to $q_f$. Notice that this does not ensure that $q_f$ is reachable almost-surely for this $k$ (and actually, it is not; see Example \[ex-1c\]).
We aim here at finding cut-offs for almost-sure reachability, i.e., we seek the existence of a threshold such that almost-sure reachability (or its negation) holds for all larger values.
Fix a protocol $\Prot=\tuple{Q, D, q_0, T}$, $d_0\in D$, and $q_f\in Q$. An integer $k\in\bbN$ is a *cut-off for almost-sure reachability* (shortly a *cut-off*) for $\Prot$, $d_0$ and $q_f$ if one of the following two properties holds:
- for all $h\geq k$, we have ${\mathbb{P}}(\tuple{q^h_0,d_0},{\llbracket \Diamond q_f \rrbracket})=1$. In this case $k$ is a *positive* cut-off;
- for all $h\geq k$, we have ${\mathbb{P}}(\tuple{q^h_0,d_0},{\llbracket \Diamond q_f \rrbracket})< 1$. Then $k$ is a *negative* cut-off.
An integer $k$ is a *tight* cut-off if it is a cut-off and $k-1$ is not.
=-1 Notice that from the definition, cut-offs need not exist for a given distributed system. Our main result precisely states that cut-offs always exist, and that we can decide their nature.
\[thm-main\] For any protocol $\Prot$, any initial register value $d_0$ and any target location $q_f$, there always exists a cut-off for almost-sure reachability, whose value is at most doubly-exponential in the size of $\calP$. Whether it is a positive or a negative cut-off can be decided in [aux auxempty$\mathsf{\mathit{}}$EXPSPACE]{}, and is -hard.
When dropping the condition on non-atomic read/write operations, and allowing transitions with atomic read/ write operations (i.e., one process is ensured to perform a read and a write operation without to be interrupted by another process), the existence of a cut-off (Theorem \[thm-main\]) is not ensured. This is demonstrated with the protocol of Fig. \[fig-ex-nocutoff\]: one easily checks (e.g., inductively on the number of processes, since processes that end up in $q_2$ play no role anymore) that state $q_f$ is reached with probability $1$ if, and only if, the number of processes is odd.
Properties of register protocols
================================
Example of a register protocol {#subsec:example}
------------------------------
(-2.15,.4) – (2.4,-1.9); (0,0) node\[rond,bleu\] (a) [$q_0$]{}; (0,-1.5) node\[rond,jaune\] (b) [$q_1$]{}; (2,-.75) node\[rond,rouge\] (c) [$q_2$]{}; (-1.75,0) node\[rond,vert\] (d) [$q_f$]{}; (a.135) – +(135:3mm);
\(a) edge\[-latex’, bend right=20\] node\[left=-1pt\] [$\genfrac{}{}{0pt}{1}{R(0)}{W(1)}$]{} (b); (b) edge\[-latex’, bend right=20\] node\[right=-1pt\] [$\genfrac{}{}{0pt}{1}{R(1)}{W(0)}$]{} (a); (a) edge\[-latex’, bend left=0\] node\[pos=.2,above right=-2pt\] [$\genfrac{}{}{0pt}{1}{R(1)}{W(2)}$]{} (c); (b) edge\[-latex’\] node\[pos=.2,below right=-2pt\] [$\genfrac{}{}{0pt}{1}{R(2)}{W(0)}$]{} (c); (a) edge\[-latex’\] node\[above\] [$R(0)$]{} (d);
\(c) edge\[-latex’,out=-60,in=-120,looseness=6\] (c); (d) edge\[-latex’,out=-60,in=-120,looseness=6\] (d);
(-.4,1.15) – (7.7,-1.15); (0,0) node\[rond,bleu\] (q0) [$s_0$]{}; (q0.-135) – +(-135:3mm); (q1) [$s_1$]{}; (q2) [$s_2$]{}; (qmid) [$\dots$]{}; (qn1) [$s_{n-1}$]{}; (qn) [$s_n$]{}; (q0) edge\[-latex’, out=-60,in=-120,looseness=7\] node\[below,pos=.5\] [$W(0)$]{} (q0) edge\[-latex’, bend right=30\] node\[below=-1pt\] [$R(0)$]{} (q1); (q1) edge\[-latex’, bend right=30\] node\[below=-1pt\] [$W(1)$]{} (q0) edge\[-latex’, bend right=30\] node\[below\] [$R(1)$]{} (q2); (q2) edge\[-latex’, bend right, in=-120\] node\[above,pos=.2\] [$W(2)$]{} (q0) edge\[dashed, out=-30,in=180\] node\[below,pos=.8\] [$R(2)$]{} +(.8,-.3); ($(qn1)+(-.8,-.3)$) edge\[-latex’, dashed,in=-150,out=0\] node\[below,pos=.2\] [$R(n-2)$]{} (qn1); (qn1) edge\[-latex’, bend right=30\] node\[below\] [$R(n-1)$]{} (qn); (qn1.160) edge\[dashed\] node\[pos=.9,above right=-1pt\] [$W(n-1)$]{} +(160:1.2cm); (q0) edge\[dashed,latex’-,out=80,in=-160\] +(50:1.5cm);
\[fig-filtern\]
=-1 We illustrate our model with a family of register protocols $\left(\calF_n\right)_{n>0}$, depicted in Fig. \[fig-filtern\]. For a fixed $n$, protocol $\calF_n$ has $n+1$ states and $n$ different data; intuitively, in order to move from $s_i$ to $s_{i+1}$, two processes are needed: one writes $i$ in the register and goes back to $s_0$, and the second process can proceed to $s_{i+1}$ by reading $i$. Since backward transitions to $s_0$ are always possible and since states can always exit $s_0$ by writing a $0$ and reading it afterwards, no deadlock can ever occur so the main question remains to determine if $s_n$ is reachable by one of the processes as we increase the number of initial processes. As shown in Lemma \[lem-filter\], the answer is positive: $\calF_n$ has a tight linear positive cut-off; it actually behaves like a “filter”, that can test if at least $n$ processes are running together. We exploit this property later in Section \[sec-hardness\].
\[lem-filter\] Fix $n\in\bbN$. The “filter” protocol $\calF_n$, depicted in Fig. \[filtern\], with initial register value $0$ and target location $s_n$, has a tight positive cut-off equal to $n$.
We consider the system $\Sys_{\calF_n}^m$, made of $m$ copies of $\calF_n$, with initial register value $0$. We first prove that any reachable configuration $\aconf$ satisfies: $$\forall j\leq m.\ \sum_{k=0}^{j}{st(\gamma)}(s_k) \geq j +
\indic_{{data(\gamma)}=j+1}$$ The proof is by induction: the invariant is satisfied by the initial configuration $\aconf_0=\tuple{s_0^m,0}$. Let us now consider the run $\aconf_0 \rightarrow^* \aconf \rightarrow \aconf'$, in which $\aconf$ satisfies the invariant, and with last transition $(q,{\textrm{op}}, d,q')\in T$.
- If $({\textrm{op}},d)=(R,0)$, then $q=s_0$ and $q'=s_1$. Along that transition, the right-hand-side term is unchanged; so is the left-hand-side term as soon as $j>0$, so that the inequality is preserved for those cases. The case $j=0$ is trivial.
- If $({\textrm{op}},d)=(R,i)$ with $i>0$, then $q=s_i$ and $q'=s_{i+1}$. we have ${st(\gamma')}={st(\gamma)}\Msetminus s_i \Mcup s_{i+1}$ and ${data(\gamma')} = {data(\gamma)}=i$. Again, along this read-transition, the right-hand side term is unchanged, while the left-hand-side term is unchanged for all $j\not=i$.
It remains to prove the inequality for $j=i$. We apply the induction hypothesis in $\aconf$ for $j=i-1$: since the transition $(q,R,i,q')$ is available, it must hold that ${st(\gamma)}(s_i)\geq 1$ and ${{data(\gamma)}=i=j+1}$. Hence $\sum_{k=0}^{i-1} {st(\gamma)}(s_k) \geq i-1+1=i$, and $\sum_{k=0}^{i}{st(\gamma)}(s_k) \geq {i+1}$. This implies $\sum_{k=0}^{i} {st(\gamma')}(s_k) \geq i$.
- If $({\textrm{op}},d)=(W,i)$, then $q=s_i$ and $q'=s_0$. Thus ${st(\gamma')} = {st(\gamma)}\Msetminus s_i\Mcup s_0$. For $j=i-1$, the left-hand-side term of the inequality is increased by $1$, while the right-hand-side one is either unchanged or also increased by $1$. The property is preserved in both cases. For $j\not=i-1$, the left-hand-side term cannot decrease, while the right-hand-side term cannot increase. Hence the invariant is preserved.
As a consequence, if $m<n$, we have (for $j=m$) $\sum_{k=0}^{m}{st(\gamma)}(s_k) = m$ for any reachable configuration $\aconf$, so that ${st(\gamma)}(s_n) = 0$.
Conversely, if $m\geq n$, from any configuration $\aconf$, it is possible to reach $\gamma_i =
\tuple{s_0^{i}\Mcup s_{i+1}^{m-i},i}$ for any $0\leq i < n$:
- for $i=0$: all processes can go to $s_0$, then write $0$ in the register, and all move to $s_1$: $\aconf\rightarrow^*\tuple{s_0^m,d}\xrightarrow{(W,0)}\tuple{s_0^m,0}
{\xrightarrow{(R,0)}}{}^m\tuple{s_1^m,0}$;
- for $1<i<n-1$, assuming $\tuple{s_0^{i}\Mcup s_{i+1}^{m-i},i}$ can be reached, one of the processes in $s_{i+1}$ can write $i+1$ (going back to $s_0$), and the remaining $m-i-1$ processes in $s_{i+1}$ can go to $s_{i+2}$: $$\tuple{s_0^is_{i+1}^{m-i}, i}
\xrightarrow{(W,i+1)}\tuple{s_0^{i+1}s_{i+1}^{m-i-1},i+1}
\xrightarrow{(R,i+1)}{}^{m-i-1}\tuple{s_0^{i+1}s_{i+2}^{m-i-1},i+1}$$
Thus from any $\gamma\in\Gamma$, configuration $\gamma_{n-1} =
\tuple{s_0^{n-1}s_n^{m-n+1},n-1}$ is reachable. Furthermore, $\gamma_{n-1}$ contains the final state $s_n$ since $m\geq n$.
Hence, we deduce that there is a unique bottom strongly-connected component in $\Sys_{\calF_n}^m$, and that $\gamma_{n-1}$ belongs to it: this configuration is reached with probability $1$ from $\tuple{s_0^m,0}$. It follows that ${\mathbb{P}}(\tuple{s_0^m,0},{\llbracket \Diamond s_f \rrbracket}) = 1$.
Basic results
-------------
In this section, we consider a register protocol $\Prot = \tuple{Q, D, q_0, T}$, its associated distributed system $\Sys_\Prot=\tuple{\Confs,{\mathit{Pr}}}$, an initial register value $d_0\in D$ and a target state $q_f \in Q$. We define a partial order $\preceq$ over the set $\Confs$ of configurations as follows: $\tuple{\mu,d} \preceq \tuple{\mu',d'}$ if, and only if, $d=d'$ and ${\ensuremath{\overline{\vphantom{'}\mu}\xspace}}={\ensuremath{\overline{\vphantom{}\mu'}\xspace}}$ and $\mu\Msubseteq \mu'$. Note that with respect to the classical order over multisets, we require here that the supports of $\mu$ and $\mu'$ be the same (we add in fact a finite information to hold for the comparison). We know from Dickson’s lemma that $\tuple{\bbN^Q,\Msubseteq}$ is a wqo and since $Q$, $D$ and the supports of multisets in $\bbN^Q$ are finite, we can deduce the following lemma.
\[lem:conf-wqo\] $\tuple{\Confs,\preceq}$ is a wqo.
We will give some properties of register protocols, but first we introduce some further notations. Given a set of configuration $\SubConfs \subseteq \Confs$, we define ${\mathrm{Pre}}^\ast(\SubConfs)$ and ${\mathrm{Post}}^\ast(\SubConfs)$ as follows:
2 \^()&=[{ ’ . \^’}]{} & \^()&=[{’ . \^’}]{}
We also define the set ${\llbracket q_f \rrbracket}$ of configurations we aim to reach as ${\{\aconf \in \Confs \mid
{st(\aconf)}(q_f)>0\}}$. It holds that $\aconf \in
{\mathrm{Pre}}^\ast({\llbracket q_f \rrbracket})$ if, and only if, there exists a path in ${\llbracket \Diamond q_f \rrbracket}$ starting in $\aconf$.
=-1 As already mentioned, when $\tuple{\mu,d} \rightarrow
\tuple{\mu',d'}$ in $\Sys_\Prot$, the multisets $\mu$ and $\mu'$ have the same cardinality. This implies that given $k >0$, the set ${\mathrm{Post}}^\ast({\{\tuple{q^k_0,d_0}\}})$ is finite (remember that $Q$ and $D$ are finite). As a consequence, for a fixed $k$, checking whether ${\mathbb{P}}(\tuple{q^k_0,d_0},{\llbracket \Diamond q_f \rrbracket})=1$ can be easily achieved by analyzing the finite-state discrete-time Markov chain $\Sys_\Prot^k$ [@PoMC2008-BK].
\[lem:prob-pre-post\] Let $k\geq 1$. Then ${\mathbb{P}}(\tuple{q^k_0,d_0},{\llbracket \Diamond q_f \rrbracket})=1$ if, and only if, ${\mathrm{Post}}^\ast({\{\tuple{q_0^k,d_0}\}}) \subseteq {\mathrm{Pre}}^\ast({\llbracket q_f \rrbracket})$.
=-1 The difficulty here precisely lies in finding such a $k$ and in proving that, once we have found one correct value for $k$, all larger values are correct as well (to get the cut-off property). Characteristics of register protocols provide us with some tools to solve this problem. We base our analysis on reasoning on the set of configurations reachable from initial configurations in ${\def\@tempa{}\ifx\@tempa\@empty\relax
\ensuremath{\mathord{\uparrow}}\else
\ensuremath{\mathord{\uparrow_{}}}\fi}{\{\tuple{q_0,d_0}\}}$ (the upward closure of ${\{\tuple{q_0,d_0}\}}$ w.r.t. $\preceq$), remember that since the order $\tuple{\Confs,\preceq}$ requires equality of support for elements to be comparable, we have that ${\def\@tempa{}\ifx\@tempa\@empty\relax
\ensuremath{\mathord{\uparrow}}\else
\ensuremath{\mathord{\uparrow_{}}}\fi}{\{\tuple{q_0,d_0}\}}= \bigcup_{k\geq 1}
{\{\tuple{q_0^k,d_0}\}}$. We begin by showing that this set of reachable configurations and the set of configurations from which ${\llbracket q_f \rrbracket}$ is reachable are both upward-closed. Thanks to Lemma \[lem:conf-wqo\], they can be represented as upward closures of finite sets. To show that ${\mathrm{Post}}^\ast({\def\@tempa{}\ifx\@tempa\@empty\relax
\ensuremath{\mathord{\uparrow}}\else
\ensuremath{\mathord{\uparrow_{}}}\fi}{\{\tuple{q_0,d_0}\}})$ is upward-closed, we prove that register protocols enjoy the following monotonicity property. A similar property is given in [@DEGM15] and derives from the non-atomicity of operations.
\[lem:copycat\] Let $\aconf_1$, $\aconf_2$, and $\aconf'_2$ be configurations in $\Confs$. If $\aconf_1
\rightarrow^\ast \aconf_2$ and $\aconf_2 \preceq \aconf'_2$, then there exists $\aconf'_1 \in \Confs$ such that $\aconf'_1
\rightarrow^\ast \aconf'_2$ and $\aconf_1 \preceq \aconf'_1$.
Assume $\gamma_1\rightarrow\gamma_2$ with transition $(q_1,{\textrm{op}},d,q_2)\in T$ and $\gamma_2\preceq \gamma'_2$. Let $k={st(\gamma'_2)}(q_2)-{st(\gamma_2)}(q_2)\geq 0$. Then ${st(\gamma'_2)}(q_2) = k+{st(\gamma_2)}(q_2) \geq k+1$ so we define $\gamma'_1=\tuple{ {st(\gamma'_2)}\Msetminus q_2^{k+1}\Mcup q_1^{k+1},
{data(\gamma_1)} }$. Then:
- if ${\textrm{op}}=W$, the path $\gamma'_1\rightarrow^\ast\gamma'_2$, obtained by performing $k+1$ times the transition $(q_1,W,d,q_2)$, is a valid path since write operations can always be performed, independently of the content of the register;
- if ${\textrm{op}}=R$, the path $\gamma'_1\rightarrow^\ast\gamma'_2$, defined by applying $k+1$ times the transition $(q_1,R,d,q_2)$, is also a valid path, since the data $d$ in the register is unchanged.
By construction, we have ${st(\gamma_1)}(q_2)={st(\gamma'_1)}(q_2)$, and $1\leq {st(\gamma_1)}(q_1)\leq{st(\gamma'_1)}(q_1)$, and ${st(\gamma_1)}(q)={st(\gamma'_1)}(q)$ for all $q\neq q_2$. Hence $\gamma_1\preceq \gamma'_1$. The result is then generalized to arbitrary path $\gamma_1\rightarrow^\ast\gamma_2$ by induction.
${\mathrm{Pre}}^\ast({\llbracket q_f \rrbracket})$ is also upward-closed, since if ${\llbracket q_f \rrbracket}$ can be reached from some configuration $\aconf$, it can also be reached by a larger configuration by keeping the extra copies idle. Thus:
\[lem:prepost-upwardclosed\] ${\mathrm{Post}}^\ast({\def\@tempa{}\ifx\@tempa\@empty\relax
\ensuremath{\mathord{\uparrow}}\else
\ensuremath{\mathord{\uparrow_{}}}\fi}{\{\tuple{q_0,d_0}\}})$ and ${\mathrm{Pre}}^\ast({\llbracket q_f \rrbracket})$ are upward-closed sets in $\tuple{\Confs,\preceq}$.
Existence of a cut-off
----------------------
From Lemma \[lem:prepost-upwardclosed\], and from the fact that $\tuple{\Confs,\preceq}$ is a wqo, there must exist two finite sequences of configurations $(\theta_i)_{1 \leq i \leq n}$ and $(\eta_i)_{1
\leq i \leq m}$ such that ${{\mathrm{Post}}^\ast({\def\@tempa{}\ifx\@tempa\@empty\relax
\ensuremath{\mathord{\uparrow}}\else
\ensuremath{\mathord{\uparrow_{}}}\fi}{\{\tuple{q_0,d_0}\}})=
{\def\@tempa{}\ifx\@tempa\@empty\relax
\ensuremath{\mathord{\uparrow}}\else
\ensuremath{\mathord{\uparrow_{}}}\fi}{\{\theta_1,\ldots,\theta_n\}}}$ and ${\mathrm{Pre}}^\ast({\llbracket q_f \rrbracket})={\def\@tempa{}\ifx\@tempa\@empty\relax
\ensuremath{\mathord{\uparrow}}\else
\ensuremath{\mathord{\uparrow_{}}}\fi}{\{\eta_1,\ldots,\eta_m\}}$. By analyzing these two sequences, we now prove that any register protocol has a cut-off (for any initial register value and any target location).
We let $\Delta, \Delta' \subseteq \Confs$ be two upward-closed sets (for $\preceq$). We say that *$\Delta$ is included in $\Delta'$ modulo single-state incrementation* whenever for every $\gamma \in \Delta$, for every $q \in
{\ensuremath{\overline{\vphantom{}\aconf}\xspace}}$, there is some $k \in \bbN$ such that $\aconf \Mcup q^k \in \Delta'$. Note that this condition can be checked using only comparisons between minimal elements of $\Delta$ and $\Delta'$. In particular, we have the following lemma.
\[lemma:inc\] ${\mathrm{Post}}^\ast({\def\@tempa{}\ifx\@tempa\@empty\relax
\ensuremath{\mathord{\uparrow}}\else
\ensuremath{\mathord{\uparrow_{}}}\fi}{\{\tuple{q_0,d_0}\}})$ is included in ${\mathrm{Pre}}^\ast({\llbracket q_f \rrbracket})$ modulo single-state incrementation if, and only if, for all $i \in [1;n]$, and for all $q\in{\ensuremath{\overline{\vphantom{}\theta_i}\xspace}}$, there exists $j \in [1;m]$ such that ${data(\theta_i)}={data(\eta_j)}$ and ${\ensuremath{\overline{\vphantom{}\theta_i}\xspace}} =
{\ensuremath{\overline{\vphantom{\theta}\eta_j}\xspace}}$ and ${st(\eta_j)}(q') \leq
{st(\theta_i)}(q')$ for all $q' \in Q \setminus{\{q\}}$.
Suppose that ${\mathrm{Post}}^\ast({\def\@tempa{}\ifx\@tempa\@empty\relax
\ensuremath{\mathord{\uparrow}}\else
\ensuremath{\mathord{\uparrow_{}}}\fi}{\{\tuple{q_0,d_0}\}})$ is included in ${\mathrm{Pre}}^\ast({\llbracket q_f \rrbracket})$ modulo single-state incrementation. Let $i \in [1;n]$ and $q \in
{\ensuremath{\overline{\vphantom{}\theta_i}\xspace}}$. By definition, there exists some $k \in \bbN$ such that $\theta_i \Mcup q^k \in {\mathrm{Pre}}^\ast({\llbracket q_f \rrbracket})$. Hence there is $j \in [1;m]$ such that $\eta_j \preceq \theta_i \Mcup q^k$. Hence we have ${data(\theta_i)}={data(\eta_j)}$, ${\ensuremath{\overline{\vphantom{}\theta_i}\xspace}} =
{\ensuremath{\overline{\vphantom{}\eta_j}\xspace}}$ and ${st(\eta_j)}(q') \leq
{st(\theta_i)}(q')$ for all $q' \in Q \setminus{\{q\}}$.
Now assume that for all $i \in [1;n]$, and for all $q\in{\ensuremath{\overline{\vphantom{}\theta_i}\xspace}}$, there exists $j \in [1;m]$ such that ${data(\theta_i)}={data(\eta_j)}$, ${\ensuremath{\overline{\vphantom{}\theta_i}\xspace}} =
{\ensuremath{\overline{\vphantom{\theta}\eta_j}\xspace}}$ and ${st(\eta_j)}(q') \leq
{st(\theta_i)}(q')$ for all $q' \in Q \setminus{\{q\}}$. Let $\aconf \in {\mathrm{Post}}^\ast({\def\@tempa{}\ifx\@tempa\@empty\relax
\ensuremath{\mathord{\uparrow}}\else
\ensuremath{\mathord{\uparrow_{}}}\fi}{\{\tuple{q_0,d_0}\}})$. Hence there exists $i \in [1;n]$ such that $\theta_i \preceq \aconf$ (note that hence ${\ensuremath{\overline{\vphantom{}\theta_i}\xspace}}={\ensuremath{\overline{\vphantom{}\aconf}\xspace}}$). Let $q \in
{\ensuremath{\overline{\vphantom{}\aconf}\xspace}}$. Then there exists $j \in [1;m]$ such that ${data(\theta_i)}={data(\eta_j)}$, ${\ensuremath{\overline{\vphantom{}\theta_i}\xspace}} =
{\ensuremath{\overline{\vphantom{\theta}\eta_j}\xspace}}$ and ${st(\eta_j)}(q') \leq
{st(\theta_i)}(q')$ for all $q' \in Q \setminus{\{q\}}$. Take $k=|{st(\eta_j)}(q)-{st(\theta_i)}(q)|$. We consider the configuration $\aconf'=\aconf
\Mcup q^k$. For all $q' \in Q \setminus{\{q\}}$, we have ${st(\eta_j)}(q') \leq
{st(\theta_i)}(q') \leq {st(\aconf')}(q')$. And we have ${st(\eta_j)}(q) \leq {st(\theta_i)}(q) + k \leq
{st(\aconf')}(q)$. This allows us to deduce that $\eta_j \preceq
\aconf \Mcup q^k$ and consequently $\aconf \Mcup q^k \in
{\mathrm{Pre}}^\ast({\llbracket q_f \rrbracket})$. Consequently ${\mathrm{Post}}^\ast({\def\@tempa{}\ifx\@tempa\@empty\relax
\ensuremath{\mathord{\uparrow}}\else
\ensuremath{\mathord{\uparrow_{}}}\fi}{\{\tuple{q_0,d_0}\}})$ is included in ${\mathrm{Pre}}^\ast({\llbracket q_f \rrbracket})$ modulo single-state incrementation.
Using the previous characterization of inclusion modulo single-state incrementation for ${\mathrm{Post}}^\ast({\def\@tempa{}\ifx\@tempa\@empty\relax
\ensuremath{\mathord{\uparrow}}\else
\ensuremath{\mathord{\uparrow_{}}}\fi}{\{\tuple{q_0,d_0}\}})$ and ${\mathrm{Pre}}^\ast({\llbracket q_f \rrbracket})$ together with the result of Lemma \[lem:prob-pre-post\], we are able to provide a first characterization of the existence of a negative cut-off.
\[lem:negative-cut-off\] If ${\mathrm{Post}}^\ast({\def\@tempa{}\ifx\@tempa\@empty\relax
\ensuremath{\mathord{\uparrow}}\else
\ensuremath{\mathord{\uparrow_{}}}\fi}{\{\tuple{q_0,d_0}\}})$ is not included in ${\mathrm{Pre}}^\ast({\llbracket q_f \rrbracket})$ modulo single-state incrementation, then $\max_{1 \leq i \leq n}(\Msize{{st(\theta_i)}} )$ is a negative cut-off.
Applying the previous lemma, there is $i \in [1;n]$ and $q\in{\ensuremath{\overline{\vphantom{}\theta_i}\xspace}}$ such that for every $j\in [1;m]$, either ${data(\theta_i)}\neq {data(\eta_j)}$, or ${\ensuremath{\overline{\vphantom{}\theta_i}\xspace}} \neq
{\ensuremath{\overline{\vphantom{}\eta_j}\xspace}}$, or there is $q_j\neq q$ such that ${st(\eta_j)}(q_j) > {st(\theta_i)}(q_j)$.
Let $k_i = \Msize{{st(\theta_{i})}}$, and fix $k \ge k_i$. We define $\gamma_{i,k}=\theta_i \Mcup q^{k-k_i}$. Clearly $\theta_i
\preceq \gamma_{i,k} \in {\mathrm{Post}}^\ast(\{q_0^k,d_0\})$. On the opposite, for every $j \in [1;m]$, $\eta_j \not\preceq
\gamma_{i,k}$; hence we conclude that $\gamma_{i,k}\not\in
{\mathrm{Pre}}^\ast({\llbracket q_f \rrbracket})$.
Applying Lemma \[lem:prob-pre-post\], we get that ${\mathbb{P}}(\tuple{q^k_0,d_0},{\llbracket \Diamond q_f \rrbracket})<1$ for every $k \ge k_i$.
We now prove that if the condition of Lemma \[lem:negative-cut-off\] fails to hold, then there is a positive cut-off.In order to make our claim precise, for every $i \in [1;n]$ and for any $q \in {\ensuremath{\overline{\vphantom{}\theta_i}\xspace}}$, we let $d_{i,q}=\max\{(|{st(\eta_j)}(q) - {st(\theta_i)}(q)|) \mid 1 \leq
j \leq m\ \text{and}\ {\ensuremath{\overline{\vphantom{}\theta_i}\xspace}} = {\ensuremath{\overline{\vphantom{}\eta_j}\xspace}}\}$.
\[lem:positive-cut-off\] If ${\mathrm{Post}}^\ast({\def\@tempa{}\ifx\@tempa\@empty\relax
\ensuremath{\mathord{\uparrow}}\else
\ensuremath{\mathord{\uparrow_{}}}\fi}{\{\tuple{q_0,d_0}\}})$ is included in ${\mathrm{Pre}}^\ast({\llbracket q_f \rrbracket})$ modulo single-state incrementation, then $\max_{1
\leq i \leq n}(\Msize{{st(\theta_i)}} + \sum_{q\in
{\ensuremath{\overline{\vphantom{}\theta_i}\xspace}}} d_{i,q} )$ is a positive cut-off.
Let $k_0 = \max_{1 \leq i \leq n}(\Msize{{st(\theta_i)}} +
\sum_{q\in {\ensuremath{\overline{\vphantom{}\theta_i}\xspace}}} d_{i,q} )$, and $k \ge
k_0$. Consider a configuration $\gamma\in{\mathrm{Post}}^\ast(\tuple{q_0^k,d_0})$. There exists $i\in[1;n]$ with $\theta_i \preceq \gamma$.
Choose $q \in {\ensuremath{\overline{\vphantom{}\theta_i}\xspace}} = {\ensuremath{\overline{\vphantom{}\gamma}\xspace}}$ such that ${st(\gamma)}(q) \ge {st(\theta_i)}(q)+d_{i,q}$ (this $q$ should exist since $\Msize{{st(\gamma)}} \ge k_0$). We apply Lemma \[lemma:inc\], and exhibit $j \in [1;m]$ such that ${data(\theta_i)} = {data(\eta_j)}$, ${\ensuremath{\overline{\vphantom{}\theta_i}\xspace}} =
{\ensuremath{\overline{\vphantom{}\eta_j}\xspace}}$ and for every $q' \ne q$, ${st(\eta_j)}(q') \le
{st(\theta_i)}(q')$. Now, ${st(\eta_j)}(q) \le {st(\theta_i)}(q) + d_{i,q} \le
{st(\gamma)}(q)$. We conclude that $\eta_j \preceq \gamma$, and therefore that ${\mathrm{Post}}^\ast({\{\tuple{q_0^k,d_0}\}}) \subseteq
{\mathrm{Pre}}^\ast({\llbracket q_f \rrbracket})$. By Lemma \[lem:prob-pre-post\], we conclude that $k_0$ is a positive cut-off.
The last two lemmas entail our first result:
\[thm-cutoff\] Any register protocol admits a cut-off (for any given initial register value and target state).
Detecting negative cut-offs {#sec-algo}
===========================
We develop an algorithm for deciding whether a distributed system associated with a register protocol has a negative cut-off. Thanks to Theorem \[thm-cutoff\], this can also be used to detect the existence of a positive cut-off. Our algorithm relies on the construction and study of a *symbolic graph*, as we define below: for any given protocol $\Prot$, the symbolic graph has bounded size, but can be used to reason about *arbitrarily large* distributed systems built from $\Prot$. It will store sufficient information to decide the existence of a negative cut-off.
$k$-bounded symbolic graph
--------------------------
In this section, we consider a register protocol $\Prot=\tuple{Q,D,q_0,T}$, its associated distributed system $\Sys_\Prot=\tuple{\Confs,{\mathit{Pr}}}$, an initial register value $d_0 \in
D$, and a target location $q_f \in Q$ of $\Prot$. With $\Prot$, we associate a finite-state graph, called *symbolic graph of index $k$*, which for $k$ large enough contains enough information to decide the existence of a negative cut-off.
Let $k$ be an integer. The *symbolic graph of index $k$* associated with $\Prot$ and $d_0$ is the transition system $\SG=\tuple{V,v_0,E}$ where
- $V = \bbN^Q_k \times 2^Q\times D$ contains triples made of a multiset of states of $Q$ of size $k$, a subset of $Q$, and the content of the register; the multiset (called *concrete part*) is used to exactly keep track of a fixed set of $k$ processes, while the subset of $Q$ (the *abstract part*) encodes the support of the arbitrarily many remaining processes;
- $v_0=\tuple{q_0^k, \{q_0\}, \{d_0\}}$;
- transitions are of two types, depending whether they involve a process in the concrete part or a process in the abstract part. Formally, there is a transition $\tuple{\mu,S,d} \to \tuple{\mu',S',d'}$ whenever there is a transition $(q,O,d'',q')\in T$ such that $d=d'=d''$ if $O=R$ and $d'=d''$ if $O=W$, and one of the following two conditions holds:
- either $S'=S$ and $q\Msubseteq\mu$ (that is, $\mu(q)>0$) and $\mu'=\mu\Msetminus q\Mcup q'$;
- or $\mu=\mu'$ and $q\in S$ and $S'\in\{S\setminus\{q\}\cup
\{q'\},S\cup \{q'\}\}$.
The symbolic graph of index $k$ can be used as an abstraction of distributed systems made of at least $k+1$ copies of $\Prot$: it keeps full information of the states of $k$ processes, and only gives the support of the states of the other processes. In particular, the symbolic graph of index $0$ provides only the states appearing in each configuration of the system.
=\[draw,rounded corners=2mm,minimum height=6mm,inner sep=3pt\] (0,0) node\[sg,fill=fbleuc\] (q0-0) [$\{q_0\},0$]{};
(2,-.5) node\[sg,fill=fvert\] (q1-1) [$\{q_1\},1$]{}; (2,-1.5) node\[sg,fill=fvert\] (q1-0) [$\{q_1\},0$]{}; (8,-.5) node\[sg,fill=fvert\] (q1-2) [$\{q_1\},2$]{};
(5,-.5) node\[sg,fill=fjaune\] (q2-1) [$\{q_2\},1$]{};
(2,1.5) node\[sg,path picture=[iin [-2,-1.5,...,2]{} [(i+.25,-1) – +(0,2); (i,-1) – +(0,2);]{}]{}\] (q01-0) [$\{q_0,q_1\},0$]{}; (2,.5) node\[sg,path picture=[iin [-2,-1.5,...,2]{} [(i+.25,-1) – +(0,2); (i,-1) – +(0,2);]{}]{}\] (q01-1) [$\{q_0,q_1\},1$]{}; (8,1.5) node\[sg,path picture=[iin [-2,-1.5,...,2]{} [(i+.25,-1) – +(0,2); (i,-1) – +(0,2);]{}]{}\] (q01-2) [$\{q_0,q_1\},2$]{};
(5,1.5) node\[sg,path picture=[iin [-2,-1.5,...,2]{} [(i+.25,-1) – +(0,2); (i,-1) – +(0,2);]{}]{}\] (q02-1) [$\{q_0,q_2\},1$]{};
(5,.5) node\[sg,path picture=[iin [-2,-1.25,...,2]{} [(i+.25,-1) – +(0,2); (i+.5,-1) – +(0,2); (i,-1) – +(0,2);]{}]{}\] (q012-1) [$\{q_0,q_1,q_2\},1$]{}; (8,.5) node\[sg,path picture=[iin [-2,-1.25,...,2]{} [(i+.25,-1) – +(0,2); (i+.5,-1) – +(0,2); (i,-1) – +(0,2);]{}]{}\] (q012-2) [$\{q_0,q_1,q_2\},2$]{};
(5,-1.5) node\[sg,path picture=[iin [-2,-1.5,...,2]{} [(i+.25,-1) – +(0,2); (i,-1) – +(0,2);]{}]{}\] (q12-1) [$\{q_1,q_2\},1$]{}; (8,-1.5) node\[sg,path picture=[iin [-2,-1.5,...,2]{} [(i+.25,-1) – +(0,2); (i,-1) – +(0,2);]{}]{}\] (q12-2) [$\{q_1,q_2\},2$]{};
(11,0) node\[sg,fill=frouge,minimum height=3.3cm,text width=1.75cm,align=center\] (qf) [all sets\
containing $q_f$]{}; (q0-0) |- (q1-0); (q0-0) |- (q01-0); (q01-0) edge\[-latex’,out=210,in=150\] (q1-0); (q01-0) – (q01-1); (q1-0) – (q1-1); (q01-1) – (q012-1); (q01-1) – (q02-1); (q1-1) – (q12-1); (q1-1) – (q2-1); (q02-1) – (q012-2); (q02-1) – (q01-2); (q01-2) edge\[-latex’,out=160,in=40\] (q01-1.30); (q012-1) – (q01-2); (q012-1) – (q02-1); (q012-1) – (q012-2); (q2-1) – (q12-2); (q12-1) – (q2-1); (q12-1) – (q2-1); (q2-1) – (q1-2); (q12-1) – (q1-2); (q1-2) edge\[-latex’,out=165,in=15\] (q1-1); (q12-1) – (q12-2); (q12-2) – (q1-2); (q012-2) – (q01-2); (q12-2) – ($(qf.180)+(0,-1.5)$); (q012-2) – ($(qf.180)+(0,.5)$);
[running]{}\[ex-1c\] Consider the protocol depicted in Fig. \[fig-ex-protocol\]. Its symbolic graph of index $0$ is depicted in Fig. \[fig-ex-symbg\]. Notice that the final state (representing all configurations containing $q_f$) is reachable from any state of this symbolic graph. However, our original protocol $\Prot$ of Fig. \[fig-ex-protocol\] does not have a positive cut-off (assuming initial register value $0$): indeed, with positive probability, a single process will go to $q_1$ and immediately write $1$ in the register, thus preventing any other process to leave $q_0$; then one may check that the process in $q_1$ alone cannot reach $q_f$, so that the probability of reaching $q_f$ from $q_0^k$ is strictly less than $1$, for any $k>0$. This livelock is not taken into account in the symbolic graph of index $0$, because from any configuration with support $\{q_0,q_1\}$ and register data equal to $1$, the symbolic graph has a transition to the configuration with support $\{q_0,q_1,q_2\}$, which only exists in the concrete system when there are at least two processes in $q_1$. As we prove in the following, analyzing the symbolic graph for a sufficiently large index guarantees to detect such a situation.
For any index $k$, the symbolic graph achieves the following correspondence:
\[lemma-gsymb\] Given two states $\tuple{\mu,S,d}$ and $\tuple{\mu',S',d'}$, there is a transition from $\tuple{\mu,S,d}$ to $\tuple{\mu',S',d'}$ in the symbolic graph of index $k$ if, and only if, there exist multisets $\delta$ and $\delta'$ with respective supports $S$ and $S'$, and such that $\tuple{\mu\Mcup\delta,d}\to
\tuple{\mu'\Mcup\delta',d'}$ in $\Sys_\Prot$.
We begin with the reverse implication: if there is a transition from $\tuple{\mu\Mcup\delta,d}$ to $\tuple{\mu'\Mcup\delta',d'}$ (assuming it is a write transition, the other case being similar) in the distributed system, then this transition originates from a transition $(q,W,d',q')$ in $\Prot$, and either this transition affects a process from the set of $k$ processes that are monitored exactly by the symbolic graph, or it affects a process in the abstract part, in which only the support is monitored. In the former case, $\delta=\delta'$, hence also their supports are equal, and the transition $(q,W,d',q')$ is applied to a location in $\mu$, which entails $q\Msubseteq \mu$ and $\mu'=\mu\Msetminus
q\Mcup q'$ and $d'=d''$. In the latter case, we get $\mu=\mu'$, and the transition $(q,W,d',q')$ is applied to a state in the support, so that $q\in
S$ and $S'$ is either $S\cup\{q'\}$ (in case $\delta(q)>1$), or it is $S\setminus\{q\}\cup\{q'\}$ (in case $\delta(q)=1$).
Conversely, if there is a transition $\tuple{\mu,S,d} \to
\tuple{\mu',S',d'}$ (assuming it originates from a $W$-transition $(q,W,d',q')$ in $\Prot$, the other case being similar), we again have to consider two separate cases.
- The first case is when $S'=S$, $q\Msubseteq\mu$ and $\mu'=\mu\Msetminus q\Mcup q'$, corresponding to the case where the transition is performed by one of the $k$ processes tracked exactly by the symbolic graph. In that case, for any $\delta$ with support $S$, there is a transition from $\tuple{\mu\Mcup\delta,d}$ to $\tuple{\mu'\Mcup\delta,d'}$ in the concrete distributed system.
- In the second case, $\mu'=\mu$, $q\in S$, and $S'$ is either $S\setminus\{q\}\cup\{q'\}$ or $S\cup \{q'\}$. Consider any multiset $\delta$ with support $S$, and such that $\delta(q)>1$ in case $S'=S\cup\{q'\}$, and $\delta(q)=1$ if $S'=S\setminus \{q\}\cup\{q'\}$. Let $\delta'=\delta\Msetminus q\Mcup q'$; then the support of $\delta'$ is $S'$, and there is a transition from $\tuple{\mu\Mcup\delta,d}$ to $\tuple{\mu'\Mcup \delta', d'}$, as required.
This concludes our proof.
Deciding the existence of a negative cut-off
--------------------------------------------
We now explain how the symbolic graph can be used to decide the existence of a negative cut-off. Since ${\mathrm{Pre}}^\ast({\llbracket q_f \rrbracket})$ is upward-closed in $\tuple{\Confs,\preceq}$, there is a finite set of configurations $\{\eta_i=\tuple{\mu_i,d_i} \mid
1\leq i\leq m\}$ such that ${\mathrm{Pre}}^\ast({\llbracket q_f \rrbracket})={\def\@tempa{}\ifx\@tempa\@empty\relax
\ensuremath{\mathord{\uparrow}}\else
\ensuremath{\mathord{\uparrow_{}}}\fi}{\{\eta_i\mid
1\leq i\leq m\}}$. We let $K=\max \{ {st(\eta_i)}(q) \mid q\in Q,\
1\leq i\leq m\}$, and show that for our purpose, it is enough to consider the symbolic graph of index $K\cdot|Q|$; we provide a bound on $K$ in the next section.
\[lem:algo\] There is a negative cut-off for $\Prot$, $d_0$ and $q_f$ if, and only if, there is a node in the symbolic graph of index $K\cdot |Q|$ that is reachable from $\tuple{q_0^{K\cdot|Q|},\{q_0\}, d_0}$ but from which no configuration involving $q_f$ is reachable.
We begin with the converse implication, assuming that there is a state $\tuple{\mu,S,d}$ in the symbolic graph of index $K\cdot |Q|$ that is reachable from $(q_0^{K\cdot |Q|},\{q_0\},d_0)$ and from which no configuration in ${\llbracket q_f \rrbracket}$ is reachable. Applying Lemma \[lemma-gsymb\], there exist multisets $\delta_0=q_0^N$ and $\delta$, with respective supports $\{q_0\}$ and $S$, such that $\tuple{\mu\Mcup\delta,d}$ is reachable from ${\tuple{q_0^{K\cdot
|Q|}\Mcup\delta_0,d_0}}$. If location $q_f$ was reachable from $\tuple{\mu\Mcup\delta,d}$ in the distributed system, then there would exist a path from $\tuple{\mu,S,d}$ to a state involving $q_f$ in the symbolic graph, which contradicts our hypothesis. By Lemma \[lem:copycat\], it follows that such a configuration $\tuple{\mu\Mcup\delta',d}$—which cannot reach $q_f$—can be reached from ${\tuple{q_0^{K\cdot
|Q|}\Mcup q_0^{N'},d_0}}$ for any $N' \geq N$: hence it cannot be the case that $q_f$ is reachable almost-surely for any $N' \geq N$. Therefore there cannot be a positive cut-off, which implies that there is a negative one (from Theorem \[thm-cutoff\]).
Conversely, if there is a negative cut-off, then for some $N>K\cdot |Q|$, the distributed system $\Sys_\Prot^N$ with $N$ processes has probability less than $1$ of reaching ${\llbracket q_f \rrbracket}$ from $q_0^N$. This system being finite, there must exist a reachable configuration $\tuple{\mu,d}$ from which $q_f$ is not reachable [@PoMC2008-BK]. Hence $\tuple{\mu,d}\notin{\mathrm{Pre}}^\ast({\llbracket q_f \rrbracket})$, and for all $i\leq m$, there is a location $q^i$ such that $\mu(q^i)<\mu_i(q^i)\leq K$. Then there must exist a reachable state $\tuple{\kappa,S,d}$ of the symbolic graph of index $K\cdot |Q|$ for which $\kappa(q^i)=\mu(q^i)$ and $q^i\notin S$, for all $1\leq i\leq m$: it indeed suffices to follow the path from $\tuple{q_0^N,d_0}$ to $\tuple{\mu,d}$ while keeping track of the processes that end up in some $q^i$ in the concrete part; this is possible because the concrete part has size at least $K\cdot |Q|$.
It remains to be proved that no state involving $q_f$ is reachable from $\tuple{\kappa,S,d}$ in the symbolic graph. If it were the case, then by Lemma \[lemma-gsymb\], there would exist $\delta$ with support $S$ such that ${\llbracket q_f \rrbracket}$ is reachable from $\tuple{\kappa\Mcup\delta,d}$ in the distributed system. Then $\tuple{\kappa\Mcup\delta,d}\in {\mathrm{Pre}}^\ast({\llbracket q_f \rrbracket})$, so that for some $1\leq i\leq m$, $(\kappa\Mcup\delta)(q^i)\geq \mu_i(q^i)$, which is not possible as $\kappa(q^i)<\mu_i(q^i)$ and $q^i$ is not in the support $S$ of $\delta$. This contradiction concludes the proof.
Besides the existence of a negative cut-off, this proof also provides us with an upper bound on the tight cut-off, as we shall see in Section \[sec-bounds\].
Complexity of the algorithm
---------------------------
We now consider the complexity of the algorithm that can be deduced from Lemma \[lem:algo\]. Using results by Rackoff on the coverability problem in Vector Addition Systems [@Rac78], we can bound $K$–and consequently the size of the needed symbolic graph–by a *double-exponential* in the size of the protocol. Therefore, it suffices to solve a reachability problem in [@Sip97] on this doubly-exponential graph: this boils down to $\NEXPSPACE$ with regard to the protocol’s size, hence ${\def\@aux{} \ifx\@aux\@empty\else\ensuremath{\mathsf{\mathit{}}}\hbox{-}\fi\@EXPSPACE}$ by Savitch’s theorem [@Sip97].
\[thm:expspace-membership\] Deciding the existence of a negative cut-off is in [aux auxempty$\mathsf{\mathit{}}$EXPSPACE]{}.
Recall that ${\mathrm{Pre}}^\ast({\llbracket q_f \rrbracket})$ is exactly the set of configurations that can cover $q_f$, i.e., configurations $\aconf$ from which there exists a path $\aconf \rightarrow^\ast \aconf'$ with ${st(\aconf')}(q_f) > 0$. Recall also that it can be written as an upward-closure of minimal elements: ${\mathrm{Pre}}^\ast({\llbracket q_f \rrbracket}) = {\def\@tempa{}\ifx\@tempa\@empty\relax
\ensuremath{\mathord{\uparrow}}\else
\ensuremath{\mathord{\uparrow_{}}}\fi}{\{\eta_1,\ldots,\eta_m\}}$. Now, consider the value $K$ in Lemma \[lem:algo\]: it is defined as $K=\max \{
{st(\eta_i)}(q) \mid q\in Q,\ 1\leq i\leq m\}$, i.e., the maximum number of states appearing in any multiset of any minimal configuration $\eta_i$. The value of $K$ can be bounded using classical results on the coverability problem in Vector Addition Systems (VAS) [@Rac78].
Intuitively, a $b$-dimensional VAS is a system composed of an initial $b$-dimensional vector $\mathbf{v_{0}}$ of naturals (the *axiom*), and a finite set of $b$-dimensional integer vectors (the *rules*). An *execution* is built as follows: it starts from the axiom and, at each step, the next vector is derived from the current one by adding a rule, provided that this derivation is *admissible*, i.e., that the resulting vector only contains non-negative integers. An execution ends if no derivation is admissible. The *coverability problem* asks if a given target vector $\mathbf{v} = (v_1, \ldots{}, v_b)$ can be covered, i.e., does there exists a (possibly extendable) execution $\mathbf{v_{0}}
\rightsquigarrow \mathbf{v_{1}} \rightsquigarrow \ldots{} \rightsquigarrow
\mathbf{v_{n}} = \mathbf{v'}$ such that, for all $1\leq i \leq b$, it holds that $v_i \leq v'_i$.
Our distributed system $\Sys_\Prot$ can be seen as a $\Msize Q$-dimensional VAS where each transition is modeled by a rule vector modifying the multiset of the current configuration. Formally, one has to take into account that available rules depend on the data stored in the shared register. This can be achieved by either considering the expressively equivalent model of VAS with states (VASS, see e.g., [@RY86]) or by adding $\mathcal{O}(\Msize D)$ dimensions to enforce this restriction. Over such a VAS(S), we are interested in the coverability of the vector corresponding to the multiset $q_f$ (i.e., containing only one copy of $q_f$ and no other state). In particular, we want to bound the size of vectors needed to cover $q_f$, as it will lead to a bound on minimal elements $\eta_i$ of ${\mathrm{Pre}}^\ast({\llbracket q_f \rrbracket})$, hence a bound on the value $K$.
Results by Rackoff (hereby as reformulated by Demri *et al.* [@DJLL13 Lemma 3]) state that if a covering execution exists from an initial vector $\mathbf{v_0}$, then there is one whose length may be doubly-exponential in the size of the input: singly-exponential in the size of the rule set and the target vector, and doubly-exponential in the dimension of the VAS. Hence, for our distributed system $\Sys_\Prot$, seen as a VAS, this implies that if $q_f$ can be covered from a configuration $\aconf$, there is a covering execution whose length is bounded by some $L$ in $2^{\calO\left(\Msize
Q \cdot \Msize D\right)^{\mathcal{O}\left(\Msize
Q + \Msize D\right)}}$. This bound on the *length* of the execution obviously also implies a bound on the *number of processes* actively involved in the execution (because at each transition, only one process is active). Hence, we can deduce that if a configuration $\aconf = \tuple{\mu,d}$ can cover $q_f$ (i.e., there exists a path $\aconf \rightarrow^\ast \aconf'$ with ${st(\aconf')}(q_f) > 0$), then it is also the case of configuration $\aconf'' =
\tuple{\mu'',d}$, which we build as follows: $\forall\, q \in Q,\;
\mu''(q) = \min\{\mu(q), L\}$. That is, it also holds that there exists a path $\aconf'' \rightarrow^\ast \aconf'''$ with ${st(\aconf''')}(q_f) > 0$.
By definition of $K$ as $K=\max \{ {st(\eta_i)}(q) \mid q\in Q,\ 1\leq
i\leq m\}$ and configurations $\eta_i$ as minimal elements for the upward-closure ${\mathrm{Pre}}^\ast({\llbracket q_f \rrbracket}) =
{\def\@tempa{}\ifx\@tempa\@empty\relax
\ensuremath{\mathord{\uparrow}}\else
\ensuremath{\mathord{\uparrow_{}}}\fi}{\{\eta_1,\ldots,\eta_m\}}$, we have that $K \leq L$ in any case. Hence, for our algorithm to be correct, it suffices to consider the symbolic graph of index $L\cdot \Msize Q$, as presented in Lemma \[lem:algo\], and to solve a reachability problem over this graph. Let us study the size of this graph. Its state space is $V =
\mathbb{N}_{L\cdot \Msize Q}^{Q} \times 2^Q\times D$. The multisets of $\mathbb{N}_{L\cdot \Msize Q}^{Q}$ are essentially mappings $Q \to \left[0; L\cdot \Msize Q \right]$. Hence, we have that: $
\Msize V \leq (L\cdot \Msize Q + 1)^{\Msize Q} \cdot 2^{\Msize Q} \cdot
\Msize D
$, which is doubly-exponential in both the state space of the protocol and the size of the data alphabet (because $L$ is). Since reachability over directed graphs lies in $\NLOGSPACE$ [@Sip97] with regard to the size of the graph, we obtain $\NEXPSPACE$-membership with regard to the size of the protocol. Finally, by Savitch’s theorem [@Sip97], we know that $\NEXPSPACE={\def\@aux{} \ifx\@aux\@empty\else\ensuremath{\mathsf{\mathit{}}}\hbox{-}\fi\@EXPSPACE}$, which concludes our proof.
-hardness for deciding cut-offs {#sec-hardness}
-------------------------------
Our proof is based on the encoding of a linear-bounded Turing machine [@Sip97]: we build a register protocol for which there is a negative cut-off if, and only if, the machine reaches its final state $\qhalt$ with the tape head reading the last cell of the tape.
\[thm:negative-cutoff\] Deciding the existence of a negative cut-off is -hard.
Write $n$ for the size of the tape of the Turing machine. We assume (without loss of generality) that the machine is deterministic, and that it accepts only if it ends in its halting state $q_{\halt}$ while reading the last cell of the tape. Our reduction works as follows (see Fig. \[fig-lbTM\]): some processes of our network will first be assigned an index $i$ in $[1;n]$ indicating the cell of the tape they shall encode during the simulation. The other processes are stuck in the initial location, and will play no role. The state $q$ and position $j$ of the head of the Turing machine are stored in the register. During the simulation phase, when a process is scheduled to play, it checks in the register whether the tape head is on the cell it encodes, and in that case it performs the transition of the Turing machine. If the tape head is not on the cell it encodes, the process moves to the target location (which we consider as the target for the almost-sure reachability problem). Finally, upon seeing $(q_{\halt},n)$ in the register, all processes move to a $(n+1)$-filter protocol $\calF_{n+1}$ (similar to that of Fig. \[fig-filtern\]) whose last location $s_{n+1}$ is the aforementioned target location.
(0,0) node\[rond,vert,inner sep=0pt\] (qinit) node ; (qinit.135) – +(135:3mm); (-2,0) node\[rond,rouge,inner sep=0pt\] (qinit’) node [$\pinit'$]{}; (-2,-1.3) node\[rond,bleu,inner sep=0pt\] (c1) node [$1,c_1$]{}; (-1,-1.3) node\[rond,bleu,inner sep=0pt\] (c2) node [$2,c_2$]{}; (0,-1.3) node\[rond,bleu,inner sep=0pt\] (c3) node [$3,c_3$]{}; (1,-1.3) node [...]{}; (2,-1.3) node\[rond,bleu,inner sep=0pt\] (cn) node [$n,c_n$]{}; (2,0) node\[rond,gris,inner sep=0pt\] (qsink) node [$\psink$]{};
(qinit) – (qinit’) node\[midway,above\] [$R(\#)$]{}; (qinit) – (qsink) node\[midway,above\] [$R(D\setminus\{\#\})$]{}; (qinit) – (c2) node\[left,midway\] [$R(\#)$]{}; (qinit) – (c3) node\[right,midway\] [$R(\#)$]{}; (qinit) – (cn) node\[right,midway\] [$R(\#)$]{}; (qinit’) – (c1) node\[left,midway\] [$W(q_0,1)$]{}; (qsink) edge\[out=-25,in=25,looseness=5\] (qsink);
(-2,0) node\[rond,bleu,inner sep=1pt\] (c’1) node [$1,\sigma$]{}; (-2,-1) node\[rounded corners=3.5mm,bleu,inner sep=1pt,minimum size=7mm\] (c’2) [$1,\sigma,q\phantom'$]{}; (-2,-2) node\[rond,bleu,inner sep=1pt\] (c’3) node [$1,\sigma'$]{}; (2,0) node\[rond,bleu,inner sep=1pt\] (c”1) node [$n,\sigma'$]{}; (2,-1) node\[rounded corners=3.5mm,bleu,inner sep=1pt,minimum size=7mm\] (c”2) [$n,\sigma',q''$]{}; (2,-2) node\[rounded corners=3.5mm,bleu,inner sep=1pt,minimum size=7mm\] (c”3) [$n,\sigma''$]{};
(c’1) – (c’2) node\[midway,right\] [$R(q,1)$]{}; (c’2) – (c’3) node\[midway,right\] [$W(q',2)$]{}; (c”1) – (c”2) node\[midway,left\] [$R(q'',n)$]{}; (c”2) – (c”3) node\[midway,left\] [$W(q,n-1)$]{};
(c’1.-135) – +(-135:3mm); (c’1.-112.5) – +(-112.5:3mm); (c’1.-45) – +(-45:3mm); (c”1.-135) – +(-135:3mm); (c”1.-67.5) – +(-67.5:3mm); (c”1.-45) – +(-45:3mm); (c’3.-135) – +(-135:3mm); (c’3.-112.5) – +(-112.5:3mm); (c’3.-67.5) – +(-67.5:3mm); (c’3.-45) – +(-45:3mm); (c’3.-90) – +(-90:3mm); (c”3.-135) – +(-135:3mm); (c”3.-67.5) – +(-67.5:3mm); (c”3.-112.5) – +(-112.5:3mm); (c”3.-45) – +(-45:3mm); (c”3.-90) – +(-90:3mm); iin [1,2,3,n]{} [(ci.-135) – +(-135:3mm); (ci.-112.5) – +(-112.5:3mm); (ci.-67.5) – +(-67.5:3mm); (ci.-45) – +(-45:3mm); (ci.-90) – +(-90:3mm);]{}
(-3.5,0) node\[rond,jaune,inner sep=0pt\] (s0) node [$s_0$]{}; (-2,0) node\[rond,jaune,inner sep=0pt\] (s1) node [$s_1$]{}; (-1,0) node\[rond,jaune,inner sep=0pt\] (s2) node [$s_2$]{}; (2,0) node\[rond,jaune,inner sep=0pt,double\] (sn1) node [$s_{n}$]{}; (3.5,0) node\[rond,jaune,inner sep=0pt,double\] (sn) node [$s_{n+1}$]{};
(s0) edge\[out=-160,in=-200,looseness=4\] node\[midway,left\] [$W(f_0)$]{} (s0); (s0) edge\[bend right\] node\[midway,below\] [$R(f_0)$]{} (s1); (s1) edge\[bend right\] node\[midway,below\] [$W(f_1)$]{} (s0); (s1) edge\[bend right\] node\[midway,below\] [$R(f_1)$]{} (s2); (s2) edge\[out=150,in=40\] node\[pos=.1,above\] [$W(f_2)$]{} (s0); (s2) edge\[bend right\] node\[midway,below\] [$R(f_2)$]{} +(.8,-.2); ($(sn1)+(-.8,-.2)$) edge\[bend right,-latex’\] node\[midway,below\] [$R(f_{n-1})$]{} (sn1); (sn1) edge\[bend right,-latex’\] node\[midway,below\] [$R(f_n)$]{} (sn); (sn) edge\[out=-20,in=20,looseness=4\] (sn);
(c’1) -| (s0) node\[pos=.9,left\] [$\genfrac{}{}{0pt}{1}{R(\qhalt,n)}{R(f_i), i\in[0;n]}$]{}; (c’3) -| (s0); (c”1) -| (sn) node\[pos=.95,right\] [$\genfrac{}{}{0pt}{1}{R(\cdot,j), j\not=n}{R(\#)}$]{}; (c”3) -| (sn); (c’1) – +(.7,0) node\[right,font=\] [(to $s_{n+1}$)]{}; (c’3) – +(.7,0) node\[right,font=\] [(to $s_{n+1}$)]{}; (c”1) – +(-.7,0) node\[left,font=\] [(to $s_{0}$)]{}; (c”3) – +(-.7,0) node\[left,font=\] [(to $s_{0}$)]{};
(5,0.4) – (5,-.8) node\[midway,right=3mm,font=\] [initialization phase]{}; (5,0); (5,-1.3) – (5,-5.2) node\[midway,right=3mm,text width=4cm,font=\]
simulation phase
(for transitions
$(q,\sigma) \to (q',\sigma',+1)$ and
$(q'',\sigma') \to (q,\sigma'',-1)$)
; (5,-6) – (5,-6.8) node\[midway,right=3mm,text width=4.2cm,font=\] [counting phase]{};
If the Turing machine halts, then the corresponding run can be mimicked with exactly one process per cell, thus giving rise to a finite run of the distributed system where $n$ processes end up in the $(n+1)$-filter (and the other processes are stuck in the initial location); from there $s_{n+1}$ cannot be reached. If the Turing machine does not halt, then assume that there is an infinite run of the distributed system never reaching the target location. This run cannot get stuck in the simulation phase forever, because it would end up in a strongly connected component from which the target location is reachable. Thus this run eventually reaches the $(n+1)$-filter, which requires that at least $n+1$ processes participate in the simulation (because with $n$ processes it would simulate the exact run of the machine, and would not reach $q_{\halt}$, while with fewer processes the tape head could not go over cells that are not handled by a process). Thus at least $n+1$ processes would end up in the $(n+1)$-filter, and with probability $1$ the target location should be reached.
We now formalize this construction, by describing the states and transitions of the protocol within these three phases. We fix a linear-bounded Turing machine $\mathcal{M} =
(Q,q_0,\qhalt,\Sigma,\delta)$, where $Q$ is the set of states, $q_0,
\qhalt \in Q$ are the initial and halting states, $\Sigma$ is the alphabet, and $\delta \subseteq Q \times \Sigma \times Q \times \Sigma
\times \{-1,+1\}$ is the set of transitions. We define the data alphabet $D=\{\#\}\uplus Q\times\Sigma \uplus \{f_i\mid
0\leq i\leq n\}$, and the set of locations $P=\{\pinit,\pinit',\psink\}
\uplus \bigl([1;n]\times \Sigma\times (Q\cup\{\epsilon\})\bigr) \uplus \{s_i\mid 0\leq i\leq
n+1\}$. The set of locations corresponds to three phases (see Fig. \[fig-lbTM\]):
- The initialization phase contains $\pinit$, $\pinit'$ and $\psink$. From the initial state $\pinit$, upon reading $\#$ (the initial content of the register), the protocol has transitions to each state $(i,\sigma_i)$ for all $2\leq i\leq n$, where $\sigma_i$ is the $i$-th letter of the initial content of the tape. If reading anything different from $\#$, the protocol moves to the sink state $\psink$. Finally, there are transitions $(\pinit,r(\#),\pinit')$ and $(\pinit',w(q_0,1),(1,c_1))$, where $q_0$ is the initial state of the Turing machine: this pair of transitions is used to initialize the computation, by setting the content of the first cell and modifying the register, so that the initialization phase is over (there are no transitions writing $\#$ in the register).
- The second phase, called *simulation phase*, uses register alphabet $Q \times [1,n]$, in order to encode the state and position of the head of the Turing machine. The state space for the simulation phase is $[1;n]\times \Sigma\times (Q\cup\{\epsilon\})$: state $(i,\sigma,\epsilon)$ (written $(i,\sigma)$ in the sequel) encodes the fact that the content of the $i$-th cell is $\sigma$; the states of the form $(i,\sigma,q)$ are intermediary states used during the simulation of one transition: when in state $(i,\sigma)$ and reading $(q,i)$ in the register, the protocol moves to $(i,\sigma,q)$, from which it moves to $(i,\sigma')$ and writes $(q',j)$ in the register, provided that the machine has a transition $(q,\sigma)\to(q',\sigma',j-i)$. If the active process does not encode the position that the tape head is reading (i.e., the process is in state $(i,\sigma)$ and reads $(q,j)$ with $j\not=i$) then it moves to the final state $s_{n+1}$ of the third phase.
- The role of the *counting phase* is to count the number of processes participating in the simulation. When seeing the halting state in the register, each protocol moves to a module whose role is to check whether at least $n+1$ protocols are still “running”. This uses data $\{f_i \mid
0\leq i \leq n\}$ and states $\{s_i \mid i\in[0,n+1]\}$, with transitions from any state of the simulation phase to $s_0$ if the register contains $(\qhalt,n)$ or any of $\{f_i \mid 0\leq i\leq n\}$.
We now prove that our construction is correct:
The register protocol $\calP_\calM$, with initial register content $\#$ and target location $s_{n+1}$, has a negative cut-off if, and only if, the Turing machine $\mathcal{M}$ reaches $\qhalt$ in the last cell of the tape.
First assume that there is a negative cut-off: there exists $N_0$ such that for any $N\geq N_0$, starting from the initial configuration $\tuple{\pinit^N,\#}$ of the system $\Sys_{\Prot_\calM}^{N}$ made of $N$ copies of $\Prot_\calM$, the probability that at least one process reaches $s_{n+1}$ is strictly less than $1$. Since $\Sys_{\Prot_\calM}^{N}$ is a finite Markov chain, this implies that there is a cone of executions never visiting $s_{n+1}$, i.e., a finite execution $\rho$ whose continuations never visit $s_{n+1}$. Since the register initially contains $\#$, this finite execution (or a finite continuation of it) must contain at least one configuration where some process has entered the simulation part.
Now, in the simulation phase, we notice that, right after taking a transition $((i,\sigma,q),\penalty1000\relax w(q',{i\pm 1}),\penalty1000\relax (i,\sigma'))$, the transition $((i,\sigma'),r(\cdot,j),s_{n+1})$ is always enabled. It follows that at the end of the finite run $\rho$, no simulation transition should be enabled; hence all processes that had entered the simulation part should have left it. Hence some process must have visited $s_0$ along $\rho$ (because we assume that $\rho$ does not involve $s_{n+1}$). Moreover, by Lemma \[lem-filter\], for $s_{n+1}$ not to be reachable along any continuation of $\rho$, no more than $n$ processes must be able to reach $s_0$ along any continuation of $\rho$, hence at most $n$ processes may have entered the simulation phase. On the other hand, for $s_0$ to be visited, some process has to first write $(\qhalt,n)$ in the register; since the register initially contains $(q_0,1)$, and no process can write $(\cdot{},i+1)$ without first reading $(\cdot{},i)$, then for each $i\in[1,n]$ there must be at least one process visiting some state $(i,\sigma_i)$, for some $\sigma_i$; It follows that at least $n$ processes must have entered the simulation phase.
In the end, along $\rho$, exactly one process visits $(i,c_i)$, for each $i\in [1,n]$, and encode the content of the $i$-th cell. As a consequence, along $\rho$, each cell of the tape of the Turing machine is encoded by exactly one process, and the execution mimics the exact computation of the Turing machine. Since the configuration $(\qhalt,n)$ is eventually reached, the Turing machine halts with the tape head on the last cell of the tape.
Conversely, assume the Turing machine halts, and consider the execution of $N\geq n$ processes where exactly one process goes in each of the $(i,c_i)$ and mimics the run of the Turing machine (the other processes going to $\psink$). We get a finite execution ending up in a configuration where all processes are either in $\pinit$ or in $\psink$, except for $n$ processes that are in the counting phase. No continuation of this prefix ever reaches $s_{n+1}$, so that the probability that some process reaches $s_{n+1}$ is strictly less than $1$.
Bounds on cut-offs {#sec-bounds}
==================
Existence of exponential tight negative cut-offs
------------------------------------------------
We exhibit a family of register protocols that admits negative cut-off exponential in the size of the protocol. The construction reuses ideas from the -hardness proof. Our register protocol has two parts: one part simulates a counter over $n$ bits, and requires a *token* (a special value in the register) to perform each step of the simulation. The second part is used to generate the tokens (i.e., writing $1$ in the register). Figure \[fig-expo\] depicts our construction. We claim that this protocol, with $\#$ as initial register value and $q_f$ as target location, admits a negative tight cut-off larger than $2^n$: in other terms, there exists $N>2^n$ such that the final state will be reached with probability strictly less than $1$ in the distributed system made of at least $N$ processes (starting with $\#$ in the register), while the distributed system with $2^n$ processes will reach the final state almost-surely. In order to justify this claim, we explain now the intuition behind this protocol.
(0,0.25) node\[rond,vert,inner sep=0pt\] (a) [$\init$]{}; (a.180) – +(180:3mm); (9,0.25) node\[rond,violet,inner sep=0pt\] (b) [$\token$]{}; (9,-2) node\[rond,violet,inner sep=0pt\] (c) [$\sent$]{}; (9,-4.25) node\[rond,violet,inner sep=0pt\] (d) [$\sink$]{}; (a) – (b); (b) – (c) node\[midway,right\] [$\scriptstyle W(1)$]{}; (c) – (d) node\[midway,right\] [$\scriptstyle R(\halt)$]{};
(5,.5) -| +(.6,-3.5) – +(-.75,-3.5) |- +(0,0) node\[pos=.2,coordinate\] (yn) ; i//in [0/2/1,2/3/2,5/n]{} [ n (i,.5) -| +(.75,-4.75) – +(-.75,-4.75) node\[midway,coordinate\] (z) |- +(0,0); (i,0) node\[rond,bleu,inner sep=0pt\] (a) [$a_{\n}$]{}; (i,-1.25) node\[rond,bleu,inner sep=0pt\] (b) [$b_{\n}$]{}; (i,-2.5) node\[rond,bleu,inner sep=0pt\] (c) [$c_{\n}$]{}; (i,-3.75) node\[rond,bleu,inner sep=0pt\] (d) [$d_{\n}$]{}; (a) – (b) node\[pos=.8,left\] [$R(\n)$]{}; (b) – (c) node\[pos=.8,left\] [$W(0)$]{}; (c) – (d) node\[pos=.8,left\] [$R(\n)$]{}; n (d) – ++(.5,.5) – ($(a\n)+(.5,-.5)$) node\[midway,fill=fgris,minimum height=3.5mm\] node\[midway,right=-2.5mm\] [$W(\j)$]{} – (a); ]{} (a) – ($(an)+(-.5,1)$) – (an); (a) – ($(a2)+(-.5,1)$) – (a2); (a) – ($(4,0)+(-.5,1)$) – (3.75,.5); (a) – ($(a1)+(-.5,1)$) node\[pos=.8,above\] [$\scriptstyle R(\#)$]{} – (a1);
(0,0) node\[rond,jaune,inner sep=0pt\] (s0) node [$s_0$]{}; (2,0) node\[rond,jaune,inner sep=0pt\] (s1) node [$s_1$]{}; (4,0) node\[rond,jaune,inner sep=0pt\] (s2) node [$s_2$]{}; (7,0) node\[rond,jaune,inner sep=0pt,double\] (sn1) node [$s_{n}$]{}; (9,0) node\[rond,rouge,inner sep=0pt,double\] (sn) node [$q_f$]{};
(s0) edge\[out=-160,in=-200,looseness=4\] node\[midway,left\] [$W(f_0)$]{} (s0); (s0) edge\[bend right\] node\[midway,below\] [$R(f_0)$]{} (s1); (s1) edge\[bend right\] node\[midway,below\] [$W(f_1)$]{} (s0); (s1) edge\[bend right\] node\[midway,below\] [$R(f_1)$]{} (s2); (s2) edge\[out=150,in=40\] node\[pos=.1,above\] [$W(f_2)$]{} (s0); (s2) edge\[bend right\] node\[midway,below\] [$R(f_2)$]{} +(.8,-.2); ($(sn1)+(-.8,-.2)$) edge\[bend right,-latex’\] node\[midway,below\] [$R(f_{n-1})$]{} (sn1); (sn1) edge\[bend right,-latex’\] node\[midway,below\] [$R(f_n)$]{} (sn);
\(c) – +(-.5,-.5) – ($(sn)+(-.5,.5)$) node\[pos=.85,right\] [$\scriptstyle R(m), m\not=\halt$]{} – (sn);
(a1) – +(.25,.4) – ($(a1)+(.9,.4)$) node [$\genfrac{}{}{0pt}{1}{R(i)}{i\not=1}$]{} – ($(an)+(1,.4)$) – +(0,-6) node\[pos=.1,right=-2pt\] [$\scriptstyle R(\#)$]{} – (sn); (a2) – +(.25,.4) – ($(a2)+(.9,.4)$) node [$\genfrac{}{}{0pt}{1}{R(i)}{i\not=2}$]{} – ($(an)+(.5,.4)$); (an) – +(.25,.4) – ($(an)+(.65,.4)$) node [$\genfrac{}{}{0pt}{1}{R(i)}{i\not=n}$]{} – ($(an)+(.9,.4)$); (c1) – +(.25,-.4) – ($(c1)+(.9,-.4)$) node [$\genfrac{}{}{0pt}{1}{R(i)}{i\not=1}$]{} – ($(cn)+(1,-.4)$) – +(0,-.5); (c2) – +(.25,-.4) – ($(c2)+(.9,-.4)$) node [$\genfrac{}{}{0pt}{1}{R(i)}{i\not=2}$]{} – ($(cn)+(.5,-.4)$); (cn) – +(.25,-.4) – ($(cn)+(.65,-.4)$) node [$\genfrac{}{}{0pt}{1}{R(i)}{i\not=n}$]{} – ($(cn)+(.5,-.4)$); (z1) |- +(-1,-.2) -| (s0.110) node\[pos=0.3,above\] [$\genfrac{}{}{0pt}{1}{R(\halt)}{R(f_i), i\in[0,n]}$]{}; (z2) |- +(-3,-.2); (yn) -| +(-.2,-1) |- ($(z2)+(-1,-.2)$); (dn) |- ($(z2)+(0,-.4)$) node\[below,pos=.8\] [$\scriptstyle W(\halt)$]{} -| (s0.70);
\[fig-expo\]
We first focus on the first part of the protocol, containing nodes named $a_i$, $b_i$, $c_i$, $d_i$ and $s_i$. This part can be divided into three phases: the initialization phase lasts as long as the register contains $\#$; the counting phase starts when the register contains $\halt$ for the first time; the simulation phase is the intermediate phase.
During the initialization phase, processes move to locations $a_i$ and $\token$, until some process in writes $1$ in the register (or until some process reaches $q_f$, using a transition from $a_i$ to $q_f$ while reading $\#$). Write $\aconf_0$ for the configuration reached when entering the simulation phase (i.e., when $1$ is written in the register for the first time). We assume that ${st(\aconf_0)}(a_i)>0$ for some $i$, as otherwise all the processes are in $\token$, and they all will eventually reach $q_f$. Now, we notice that if ${st(\aconf_0)}(a_i)=0$ for some $i$, then location $d_n$ cannot be reached, so that no process can reach the counting phase. In that case, some process (and actually all of them) will eventually reach $q_f$. We now consider the case where ${st(\aconf_0)}(a_i)\geq 1$ for all $i$. One can prove (inductively) that $d_i$ is reachable when ${st(\aconf_0)}(\token)\geq 2^i$. Hence $d_n$, and thus also $s_0$, can be reached when ${st(\aconf_0)}(\token)\geq 2^n$. Assuming $q_f$ is not reached, the counting phase must never contain more than $n$ processes, hence we actually have that ${st(\aconf_0)}(a_i)= 1$. With this new condition, $s_0$ is reached if, and only if, ${st(\aconf_0)}(\token)\geq 2^n$. When the latter condition is not true, $q_f$ will be reached almost-surely, which proves the second part of our claim: the final location is reached almost-surely in systems with strictly less than $n+2^n$ copies of the protocol.
We now consider the case of systems with at least $n+2^n$ processes. We exhibit a finite execution of those systems from which no continuation can reach $q_f$, thus proving that $q_f$ is reached with probability strictly less than $1$ in those systems. The execution is as follows: during initialization, for each $i$, one process enters $a_i$; all other processes move to $\token$, and one of them write $1$ in the register. The $n$ processes in the simulation phase then simulate the consecutive incrementations of the counter, consuming one token at each step, until reaching $d_n$. At that time, all the processes in move to , and the process in $d_n$ writes $\halt$ in the register and enters $s_0$. The processes in the simulation phase can then enter $s_0$, and those in $\sent$ can move to $\sink$. We now have $n$ processes in $s_0$, and the other ones in . According to Lemma \[lem-filter\], location $q_f$ cannot be reached from this configuration, which concludes our proof.
There exists a family of register protocols which, equipped with an initial register value and a target location, admit negative tight cut-offs whose size are exponential in the size of the protocol.
The question whether there exists protocols with exponential *positive* cut-offs remains open. The family of *filter* protocols described at Section \[subsec:example\] is an example of protocols with a linear positive cut-off.
Upper bounds on tight cut-offs
------------------------------
The results (and proofs) of Section \[sec-algo\] can be used to derive upper bounds on tight cut-offs. We make this explicit in the following theorem.
\[thm:bounds\] For a protocol $\Prot=\tuple{Q,D,q_0,T}$ equipped with an initial register value $d_0\in D$ and a target location $q_f\in Q$, the tight cut-off is at most doubly-exponential in $\Msize \Prot$.
First assume that the cut-off is negative. From Lemma \[lem:algo\], there is a state $\tuple{\mu,S,d}$ in the symbolic graph of index $K\cdot |Q|$ that is reachable from $(q_0^{(K\cdot |Q|)},\{q_0\},d_0)$ and from which no configutation containing $q_f$ is reachable. Applying Lemma \[lemma-gsymb\], there exist multisets $\delta_0=q_0^N$ and $\delta$, with respective supports $\{q_0\}$ and $S$, such that $\tuple{\mu\Mcup\delta,d}$ is reachable from ${\tuple{q_0^{K\cdot |Q|}\Mcup\delta_0,d_0}}$. Hence $N+K\cdot |Q|$ is a negative cut-off.
Let us evaluate the size of $N$: this number is extracted from the symbolic path $\tuple{q_0^{(K\cdot |Q|)},\{q_0\},d_0}\rightarrow^*
\tuple{\mu,S,d}$, which has length at most $|V|-1$ (where $V$ is the set of states of the symbolic graph of index $K\cdot |Q|$). Applying Lemma \[lem:algo\] $|V|-1$ times, increasing the size of the concrete representation of $S$ by one each time, we get $N\leq |V|$. Thus, both $K\cdot |Q|$ and $N$ are doubly-exponential in $\Msize\Prot$, thanks to the proof of Theorem \[thm:expspace-membership\].
The proof of Lemma \[lem:algo\] also entails that if the distributed system with some $N>K\cdot |Q|$ processes does not almost-surely reach the target state, then there is a negative cut-off. Hence for there to be a positive cut-off, the target has to be almost-surely reachable for all $N>K\cdot |Q|$, which makes $K\cdot |Q|$ a (doubly-exponential) positive cut-off.
Conclusions and future works
============================
We have shown that in networks of identical finite-state automata communicating (non-atomically) through a single register and equipped with a fair stochastic scheduler, there always exists a cut-off on the number of processes which either witnesses almost-sure reachability of a specific control-state (positive cut-off) or its negation (negative cut-off). This cut-off determinacy essentially relies on the monotonicity induced by our model, which allows to use well-quasi order techniques. By analyzing a well-chosen symbolic graph, one can decide in [aux auxempty$\mathsf{\mathit{}}$EXPSPACE]{}whether that cut-off is positive, or negative, and we proved this decision problem to be -hard. This approach allows us to deduce some doubly-exponential bounds on the value of the cut-offs. Finally, we gave an example of a network in which there is a negative cut-off, which is exponential in the size of the underlying protocol. Note however that no such lower-bound is known yet for positive cut-offs.
We have several further directions of research. First, it would be nice to fill the gap between the lower bound and the [aux auxempty$\mathsf{\mathit{}}$EXPSPACE]{}upper bound for deciding the nature of the cut-off. We would like also to investigate further atomic read[/]{}write operations, which generate non-monotonic transition systems, but for which we would like to decide whether there is a cut-off or not. Finally, we believe that our techniques could be extended to more general classes of properties, for instance, universal reachability (all processes should enter a distinguished state), or liveness properties.
[10]{}
C. Aiswarya, Benedikt Bollig, and Paul Gastin. An automata-theoretic approach to the verification of distributed algorithms. In Luca Aceto and David de Frutos[-]{}Escrig, editors, [ *[P]{}roceedings of the 26th [I]{}nternational [C]{}onference on [C]{}oncurrency [T]{}heory ([CONCUR]{}’15)*]{}, volume 42 of [*Leibniz International Proceedings in Informatics*]{}, pages 340–353. Leibniz-Zentrum f[ü]{}r Informatik, September 2015. [](http://dx.doi.org/10.4230/LIPIcs.CONCUR.2015.340).
Benjamin Aminof, Swen Jacobs, Ayrat Khalimov, and Sasha Rubin. Parametrized model checking of token-passing systems. In Kenneth L. McMillan and Xavier Rival, editors, [*[P]{}roceedings of the 15th [I]{}nternational [W]{}orkshop on [V]{}erification, [M]{}odel [C]{}hecking, and [A]{}bstract [I]{}nterpretation ([VMCAI]{}’14)*]{}, volume 8318 of [*Lecture Notes in Computer Science*]{}, pages 262–281. Springer-Verlag, January 2014. [](http://dx.doi.org/10.1007/978-3-642-54013-4_15).
Benjamin Aminof, Sasha Rubin, and Florian Zuleger. On the expressive power of communication primitives in parameterised systems. In Martin Davis, Ansgar Fehnker, Annabelle K. McIver, and Andrei Voronkov, editors, [*[P]{}roceedings of the 20th [I]{}nternational [C]{}onference [L]{}ogic [P]{}rogramming and [A]{}utomated [R]{}easoning ([LPAR]{}’15)*]{}, volume 9450 of [*Lecture Notes in Computer Science*]{}, pages 313–328. Springer-Verlag, November 2015.
Simon Au[ß]{}erlechner, Swen Jacobs, and Ayrat Khalimov. Tight cutoffs for guarded protocols with fairness. In Barbara Jobstmann and K. Rustan M. Leino, editors, [ *[P]{}roceedings of the 17th [I]{}nternational [W]{}orkshop on [V]{}erification, [M]{}odel [C]{}hecking, and [A]{}bstract [I]{}nterpretation ([VMCAI]{}’16)*]{}, volume 9583 of [*Lecture Notes in Computer Science*]{}, pages 476–494. Springer-Verlag, January 2016. [](http://dx.doi.org/10.1007/978-3-662-49122-5_23).
ristel Baier and Joost-Pieter Katoen. . MIT Press, May 2008.
Benedikt Bollig, Paul Gastin, and Len Schubert. Parameterized verification of communicating automata under context bounds. In Jo[ë]{}l Ouaknine, Igor Potapov, and James Worrell, editors, [ *[P]{}roceedings of the 8th [W]{}orkshop on [R]{}eachability [P]{}roblems in [C]{}omputational [M]{}odels ([RP]{}’14)*]{}, volume 8762 of [*Lecture Notes in Computer Science*]{}, pages 45–57. Springer-Verlag, September 2014. [](http://dx.doi.org/10.1007/978-3-319-11439-2_4).
Patricia Bouyer, Nicolas Markey, Mickael Randour, Arnaud Sangnier, and Daniel Stan. Reachability in networks of register protocols under stochastic schedulers. In [*[P]{}roceedings of the 43rd [I]{}nternational [C]{}olloquium on [A]{}utomata, [L]{}anguages and [P]{}rogramming ([ICALP]{}’16)*]{}, Leibniz International Proceedings in Informatics. Leibniz-Zentrum f[ü]{}r Informatik, 2016.
Edmund M. Clarke, Muralidhar Talupur, Tayssir Touili, and Helmut Veith. Verification by network decomposition. In [Ph]{}ilippa Gardner and Nobuko Yoshida, editors, [ *[P]{}roceedings of the 15th [I]{}nternational [C]{}onference on [C]{}oncurrency [T]{}heory ([CONCUR]{}’04)*]{}, volume 3170 of [*Lecture Notes in Computer Science*]{}, pages 276–291. Springer-Verlag, August-September 2004. [](http://dx.doi.org/10.1007/978-3-540-28644-8_18).
Giorgio Delzanno, Arnaud Sangnier, Riccardo Traverso, and Gianluigi Zavattaro. On the complexity of parameterized reachability in reconfigurable broadcast networks. In Deepak D’Souza, Telikepalli Kavitha, and Jaikumar Radhakrishnan, editors, [*[P]{}roceedings of the 32nd [C]{}onference on [F]{}oundations of [S]{}oftware [T]{}echnology and [T]{}heoretical [C]{}omputer [S]{}cience ([FSTTCS]{}’12)*]{}, volume 18 of [*Leibniz International Proceedings in Informatics*]{}, pages 289–300. Leibniz-Zentrum f[ü]{}r Informatik, December 2012. [](http://dx.doi.org/LIPIcs.FSTTCS.2012.289).
Giorgio Delzanno, Arnaud Sangnier, and Gianluigi Zavattaro. Parameterized verification of ad hoc networks. In Paul Gastin and Fran[ç]{}ois Laroussinie, editors, [ *[P]{}roceedings of the 21st [I]{}nternational [C]{}onference on [C]{}oncurrency [T]{}heory ([CONCUR]{}’10)*]{}, volume 6269 of [*Lecture Notes in Computer Science*]{}, pages 313–327. Springer-Verlag, September 2010. [](http://dx.doi.org/10.1007/978-3-642-15375-4_22).
St[é]{}phane Demri, Marcin Jurdzi[ń]{}ski, Oded Lachish, and Ranko Lazi[ć]{}. The covering and boundedness problems for branching vector addition systems. , 79(1):23–38, February 2013. [](http://dx.doi.org/10.1016/j.jcss.2012.04.002).
Antoine Durand[-]{}Gasselin, Javier Esparza, Pierre Ganty, and Rupak Majumdar. Model checking parameterized asynchronous shared-memory systems. In Daniel Kroening and Corina S. Pasareanu, editors, [ *[P]{}roceedings of the 27th [I]{}nternational [C]{}onference on [C]{}omputer [A]{}ided [V]{}erification ([CAV]{}’15)*]{}, volume 9206 of [*Lecture Notes in Computer Science*]{}, pages 67–84. Springer-Verlag, July 2015. [](http://dx.doi.org/10.1007/978-3-319-21690-4_5).
E. Allen Emerson and Vineet Kahlon. Reducing model checking of the many to the few. In David McAllester, editor, [*[P]{}roceedings of the 17th [I]{}nternational [C]{}onference on [A]{}utomated [D]{}eduction ([CADE]{}’00)*]{}, volume 1831 of [*Lecture Notes in Artificial Intelligence*]{}, pages 236–254. Springer-Verlag, June 2000. [](http://dx.doi.org/10.1007/10721959_19).
E. Allen Emerson and Kedar Namjoshi. On reasoning about rings. , 14(4):527–550, August 2003. [](http://dx.doi.org/10.1142/S0129054103001881).
Javier Esparza. Keeping a crowd safe: On the complexity of parameterized verification (invited talk). In Ernst W. Mayr and Natacha Portier, editors, [*[P]{}roceedings of the 31st [S]{}ymposium on [T]{}heoretical [A]{}spects of [C]{}omputer [S]{}cience ([STACS]{}’14)*]{}, volume 25 of [*Leibniz International Proceedings in Informatics*]{}, pages 1–10. Leibniz-Zentrum f[ü]{}r Informatik, March 2014. [](http://dx.doi.org/10.4230/LIPIcs.STACS.2014.1).
Javier Esparza, Alain Finkel, and Richard Mayr. On the verification of broadcast protocols. In [*[P]{}roceedings of the 14th [A]{}nnual [S]{}ymposium on [L]{}ogic in [C]{}omputer [S]{}cience ([LICS]{}’99)*]{}, pages 352–359. IEEE Comp. Soc. Press, July 1999. [](http://dx.doi.org/10.1109/LICS.1999.782630).
Javier Esparza, Pierre Ganty, and Rupak Majumdar. Parameterized verification of asynchronous shared-memory systems. In Natasha Sharygina and Helmut Veith, editors, [*[P]{}roceedings of the 25th [I]{}nternational [C]{}onference on [C]{}omputer [A]{}ided [V]{}erification ([CAV]{}’13)*]{}, volume 8044 of [*Lecture Notes in Computer Science*]{}, pages 124–140. Springer-Verlag, July 2013. [](http://dx.doi.org/10.1007/978-3-642-39799-8_8).
Steven M. German and A. Prasad Sistla. Reasoning about systems with many processes. , 39(3):675–735, July 1992.
Matthew Hague. Parameterised pushdown systems with non-atomic writes. In Supratik Chakraborty and Amit Kumar, editors, [*[P]{}roceedings of the 31st [C]{}onference on [F]{}oundations of [S]{}oftware [T]{}echnology and [T]{}heoretical [C]{}omputer [S]{}cience ([FSTTCS]{}’11)*]{}, volume 13 of [*Leibniz International Proceedings in Informatics*]{}, pages 457–468. Leibniz-Zentrum f[ü]{}r Informatik, December 2011. [](http://dx.doi.org/10.4230/LIPIcs.FSTTCS.2011.457).
Charles Rackoff. The covering and boundedness problems for vector addition systems. , 6:223–231, 1978. [](http://dx.doi.org/10.1016/0304-3975(78)90036-1).
Louis E. Rosier and Hsu-Chun Yen. A multiparameter analysis of the boundedness problem for vector addition systems. , 32(1):105–135, February 1986. [](http://dx.doi.org/10.1016/0022-0000(86)90006-1).
Michael Sipser. . PWS Publishing Company, 1997.
[^1]: F.R.S.-FNRS Postdoctoral Researcher.
[^2]: This paper extends the conference version presented in [@BMR+16b].
[^3]: This work has been partly supported by ERC Starting grant EQualIS (FP7-308087), by European FET project Cassting (FP7-601148), and by the ANR research program PACS (ANR-14-CE28-0002).
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'We give a short proof of the main result of [@A]: every Schmidt subspace of a Hankel operator is the image of a model space by an isometric multiplier. This class of subspaces is closely related to nearly $S^*$-invariant subspaces, and our proof uses Hitt’s theorem on the structure of such subspaces. We also give a formula for the action of a Hankel operator on its Schmidt subspace.'
address:
- 'Department of Mathematics, King’s College London, Strand, London, WC2R 2LS, U.K.'
- 'Université Paris-Sud XI, Laboratoire de Mathématiques d’Orsay, CNRS, UMR 8628, and Institut Universitaire de France'
author:
- Alexander Pushnitski
- Patrick Gérard
date: 12 July 2019
title: 'The structure of Schmidt subspaces of Hankel operators: a short proof'
---
Introduction and main result {#sec.a}
============================
Hankel operators
----------------
Let ${{\mathcal H}}^2\subset L^2({{\mathbb T}})$ be the standard Hardy space of the unit disk, and let $P$ be the orthogonal projection onto ${{\mathcal H}}^2$ in $L^2({{\mathbb T}})$. For a *symbol* $u\in\operatorname{BMOA}({{\mathbb T}})$, we define the Hankel operator $H_u$ acting on ${{\mathcal H}}^2$ by $$H_uf=P(u\overline{f}), \quad f\in {{\mathcal H}}^2.
\label{b0}$$ Thus, $H_u$ is an *anti-linear* operator. Denoting by $(\cdot,\cdot)$ the standard inner product in ${{\mathcal H}}^2$, we have $$(H_uz^n,z^m)=(P(u\overline{z^n}),z^m)=(u\overline{z^n},z^m)=(u,z^{m+n})={\widehat}u(n+m),$$ where $n,m\geq0$ and ${\widehat}u(\cdot)$ are the Fourier coefficients of $u$. Thus, $H_u$ is the anti-linear realisation of the Hankel matrix $\{{\widehat}u(n+m)\}_{n,m\geq0}$ in the Hardy class ${{\mathcal H}}^2$. In Section \[sec.a5\] we recall the relation of $H_u$ to a linear realisation of the Hankel matrix in ${{\mathcal H}}^2$.
Our aim is to describe the *Schmidt subspaces* $$E_{H_u}(s):=\operatorname{Ker}(H_u^2-s^2I), \quad s>0,$$ as a class of subspaces in ${{\mathcal H}}^2$. Since $H_u$ commutes with $H_u^2$, we see that $E_{H_u}(s)$ is an invariant subspace for $H_u$ (this is one of the advantages of working with the anti-linear realisation $H_u$). We give the formula for the action of $H_u$ on this subspace.
Model spaces and isometric multipliers
--------------------------------------
For an inner function $\theta$ on the unit disk we use the standard notation $${K}_\theta={{\mathcal H}}^2\cap (\theta {{\mathcal H}}^2)^\perp$$ for the corresponding model space. A convenient equivalent description of ${K}_\theta$ is $$h\in {K}_\theta \quad \Leftrightarrow \quad h\in {{\mathcal H}}^2 \text{ and } \overline{z}\theta \overline{h}\in {{\mathcal H}}^2.
\label{a2}$$ Observe that for $h\in{K}_\theta$, the combination $\overline{z}\theta \overline{h}$ is again in ${K}_\theta$.
As usual, we denote by $Sf(z)=zf(z)$ the shift operator in ${{\mathcal H}}^2$ and $S^*$ is the adjoint of $S$ in ${{\mathcal H}}^2$. Recall that the significance of model spaces stems from Beurling’s theorem which implies that a proper subspace of ${{\mathcal H}}^2$ is invariant under $S^*$ if and only if it is a model space.
If $p$ is an analytic function in the unit disk, we will say that $p$ is an *isometric multiplier* on ${K}_\theta$, if for every $f\in {K}_\theta$ we have $pf\in {{\mathcal H}}^2$ and ${\lVertpf\rVert}={\lVertf\rVert}$. In this case we denote $$p{K}_\theta:=\{pf: f\in {K}_\theta\}\, .$$ We note that for a subspace $p{K}_\theta\subset{{\mathcal H}}^2$, the choice of the parameters $p$ and $\theta$ in this representation is not unique. One can multiply $p$ and $\theta$ by arbitrary unimodular constants and one can also perform *Frostman shifts* on $p{K}_\theta$, see Section \[sec.frostman\] for the details.
Main result and discussion
--------------------------
\[thm1\][@A] Let $H_u$ be a bounded Hankel operator in ${{\mathcal H}}^2$. Every non-trivial Schmidt subspace $E_{H_u}(s)$, $s>0$, is of the form $p{K}_\theta$, where $\theta$ is an inner function and $p$ is an isometric multiplier on ${K}_\theta$. Moreover, there exists a unimodular constant $e^{i\varphi}$ such that the action of $H_u$ on this subspace is given by $$H_u(ph)=se^{i\varphi} p \overline{z} \theta \overline{h}, \quad h\in {K}_\theta\, .
\label{a00}$$
**Remarks:**
1. By , the combination $\overline{z} \theta \overline{h}$ in is in ${{\mathcal H}}^2$; in fact, it is in ${K}_\theta$.
2. The constant $e^{i\varphi}$ depends on the choice of the parameters $p$, $\theta$ in the representation $E_{H_u}(s)=p{K}_\theta$. In particular, by choosing a unimodular constant in the definition of $p$ and $\theta$, one can achieve $e^{i\varphi}=1$ in .
3. For an inner function $\theta$, consider the Hankel operator $H_{S^*\theta}$. It is not difficult to see that $$\operatorname{Ran}H_{S^*\theta}=E_{H_{S^*\theta}}(1)={K}_\theta$$ and the action of $H_{S^*\theta}$ on ${K}_\theta$ is given by the anti-linear involution $$H_{S^*\theta} h= \overline{z}\theta\overline{h}, \quad h\in {K}_\theta\,.$$ Comparing with Theorem \[thm1\], we see that such Hankel operators can be regarded as the “simplest" ones from the point of view of our analysis: they have only one non-trivial Schmidt subspace and one can choose $p={\mathbbm{1}}$. Here and in what follows, ${\mathbbm{1}}$ is the function identically equal to $1$ in ${{\mathcal H}}^2$.
4. Denoting by $T_p$ the Toeplitz operator with the symbol $p$ (which in this case, by the analyticity of $p$, coincides with the operator of multiplication by $p$), we can rewrite formula as $$H_u T_p=s e^{i\varphi}T_p H_{S^*\theta} \quad \text{ on ${K}_\theta$.}$$
5. In [@A], formula was discussed only in the case $\theta(0)=0$. This case is important because condition $\theta(0)=0$ is equivalent to ${\mathbbm{1}}\in{K}_\theta$ and thus to $p\in p{K}_\theta$. In fact, in this case for every $h\in {K}_\theta$ we have $$(ph,\overline{p(0)}p)=p(0)(ph,p)
=p(0)(h,{\mathbbm{1}})=p(0)h(0)=(ph,{\mathbbm{1}}),
\label{a3}$$ and therefore $\overline{p(0)}p$ is the orthogonal projection of ${\mathbbm{1}}$ onto the subspace $p{K}_\theta$.
6. There seems to be a close analogy between Theorem \[thm1\] and the structure of Toeplitz eigenspaces. Let $v\in L^\infty({{\mathbb T}})$ and let $T_v$ be the Toeplitz operator with the symbol $v$. Then (see [@Hayashi]) all eigenspaces of $T_v$ have the form $p{K}_\theta$. In fact, Toeplitz operators and operators of the form $H_u^2$ satisfy similar commutation relations, see Remark \[rmk\] below.
Hitt’s theorem
--------------
A closed subspace $M\subset {{\mathcal H}}^2$ is called *nearly $S^*$-invariant*, if $M\not\perp{\mathbbm{1}}$ and $$f\in M, \quad f\perp {\mathbbm{1}}\quad \Rightarrow \quad S^*f\in M.
\label{a1}$$ The proof of Theorem \[thm1\] given in the present paper relies on the following fundamental result by D. Hitt [@Hitt] (see also [@Sarason]).
\[thm2\][@Hitt] Let $M\subset {{\mathcal H}}^2$ be a non-trivial nearly $S^*$-invariant subspace. Then $M=pN$, where $N\subset {{\mathcal H}}^2$ is an $S^*$-invariant subspace and $p$ is an isometric multiplier on $N$.
By Beurling’s theorem, $N$ is either a model space or $N={{\mathcal H}}^2$; in the second case $p$ must be an inner function.
A partial converse of Hitt’s theorem is obvious: if $p$ is an isometric multiplier on an $S^*$-invariant subspace $N$, and $p(0)\not=0$ (i.e. $p\not\perp{\mathbbm{1}}$), then $pN$ is nearly $S^*$-invariant. However, if $p(0)=0$, then $pN$ is not nearly $S^*$-invariant.
Hitt’s theorem seems to be closely related to Theorem \[thm1\]. However, in [@A] the authors were unable to use Hitt’s theorem directly (even though its key ideas were used in the proof). The reason for this is that the Schmidt subspaces $E_{H_u}(s)$ are *not necessarily nearly $S^*$-invariant!* Indeed, the weight $p$ in the representation $E_{H_u}(s)=p{K}_\theta$ may vanish at zero (see e.g. Example in [@A Section 6]).
This obstacle is overcome in the present paper through the use of conformal mapping. More precisely, our plan of the proof is as follows. At the first step we consider the case $E_{H_u}(s)\not\perp{\mathbbm{1}}$. We prove that in this case $E_{H_u}(s)$ is nearly $S^*$-invariant and use Hitt’s theorem to obtain the representation $E_{H_u}(s)=p{K}_\theta$. Some additional algebra yields the formula for the action of $H_u$.
At the second step we consider the case $E_{H_u}(s)\perp{\mathbbm{1}}$. We choose a point $\alpha$ in the unit disk such that $\alpha$ is not a common zero of all elements of $E_{H_u}(s)$. We then use a Möbius map $\mu$ sending $\alpha$ to $0$ and consider the associated unitary operator $U_\mu$ on ${{\mathcal H}}^2$. It is easy to check that $U_\mu E_{H_u}(s)$ is a Schmidt subspace of another bounded Hankel operator $H_w$ and that the point $0$ is not a common zero of all elements of $E_{H_w}(s)$, i.e. $E_{H_w}(s)\not\perp{\mathbbm{1}}$. This reduces the problem to the one considered at the first step of the proof.
The proofs of this paper are self-contained, apart from the reliance on Hitt’s theorem. It is informed by the intuition coming from [@A], and in fact we reproduce some simple elements of the argument of [@A].
Linear Hankel operators {#sec.a5}
-----------------------
Here we rewrite Theorem \[thm1\] in terms of linear (rather than anti-linear) Hankel operators on the Hardy space. We follow [@A Appendix] almost verbatim. Let $J$ be the linear involution in $L^2({{\mathbb T}})$, $$Jf(z)=f(\overline{z}), \quad z\in {{\mathbb T}},$$ and let ${\mathbf C}$ be the anti-linear involution in ${{\mathcal H}}^2$, $${\mathbf C}f(z)=\overline{f(\overline{z})}, \quad z\in {{\mathbb T}}.$$ For a symbol $u\in \operatorname{BMOA}({{\mathbb T}})$, let us define the *linear* Hankel operator $G_u$ in ${{\mathcal H}}^2$ by $$G_u f=P(u\cdot Jf), \quad f\in {{\mathcal H}}^2.$$ We have $$G_u=H_u{\mathbf C}, \quad G_u^*={\mathbf C}H_u,$$ and so from Theorem \[thm1\] we obtain
Let $s$ be a singular value of $G_u$. Then there exists an inner function $\theta$ and an isometric multiplier $p$ on ${K}_\theta$ such that $$\begin{aligned}
\operatorname{Ker}(G_u^*G_u-s^2I)&={\mathbf C}(p{K}_\theta),
\\
\operatorname{Ker}(G_uG_u^*-s^2I)&=p{K}_\theta. \end{aligned}$$ The action $$G_u:\operatorname{Ker}(G_u^*G_u-s^2I)\to \operatorname{Ker}(G_uG_u^*-s^2I)$$ is given by $$G_u{\mathbf C}(pf)=sp\theta\overline{f}, \quad
f\in {K}_\theta.$$
Acknowledgements
----------------
The authors are grateful to V. Kapustin for useful discussions.
The case $E_{H_u}(s)\not\perp{\mathbbm{1}}$ {#sec.b}
===========================================
Frostman shifts {#sec.frostman}
---------------
Let $\theta$ be an inner function; then (see e.g. [@Crofoot Theorem 10]) for any ${{\lvert\alpha\rvert}}<1$ one has $${K}_\theta=g_\alpha {K}_{\theta_\alpha},$$ where $$\theta_\alpha(z)=\frac{\alpha-\theta(z)}{1-\overline{\alpha}\theta(z)},
\quad
g_\alpha(z)=\frac{1-\overline{\alpha}\theta(z)}{\sqrt{1-{{\lvert\alpha\rvert}}^2}},
\label{b00}$$ and $g_\alpha$ is an isometric multiplier on ${K}_{\theta_\alpha}$.
It follows that if $p$ is an isometric multiplier on ${K}_\theta$, then $$p{K}_\theta=pg_\alpha {K}_{\theta_\alpha},
\label{b1}$$ where $pg_\alpha$ is an isometric multiplier on ${K}_{\theta_\alpha}$. Conversely, if $$p{K}_\theta={\widetilde}p{K}_{{\widetilde}\theta},$$ where $p$ is an isometric multiplier on ${K}_\theta$ and ${\widetilde}p$ is an isometric multiplier on ${K}_{{\widetilde}\theta}$, then, again by [@Crofoot Theorem 10], $${\widetilde}p=c_1 p g_\alpha, \quad {\widetilde}\theta=c_2 \theta_\alpha,$$ where $\alpha\in{{\mathbb D}}$ and $c_1$, $c_2$ are unimodular complex numbers.
Some algebra of model spaces
----------------------------
\[lma.b1\] Let $\theta$ be an inner function in the unit disk. Then $$S^*({K}_\theta\cap {\mathbbm{1}}^\perp)={K}_\theta\cap (S^*\theta)^\perp.$$
It suffices to prove the identity $${K}_\theta\cap {\mathbbm{1}}^\perp=S({K}_\theta\cap (S^*\theta)^\perp).$$ Let $h\in {K}_\theta\cap {\mathbbm{1}}^\perp$. Write $h=Sg$, with $g=S^*h\in {K}_\theta$. Then $$(g,S^*\theta)=(Sg,\theta)=(h,\theta)=0,$$ and so $g\in {K}_\theta\cap (S^*\theta)^\perp$. Conversely, let $g\in {K}_\theta\cap (S^*\theta)^\perp$. Write any $f\in {{\mathcal H}}^2$ as $f=c{\mathbbm{1}}+Sw$, $w\in{{\mathcal H}}^2$; then $$(Sg,\theta f)
=
\overline{c}(Sg,\theta)+(Sg,\theta Sw)
=
\overline{c}(g,S^*\theta)+(g,\theta w)=0,$$ and so $Sg\in {K}_\theta$. Clearly, $Sg\perp{\mathbbm{1}}$, and so $Sg\in {K}_\theta\cap{\mathbbm{1}}^\perp$.
Some identities for $H_u$ {#sec.b2}
-------------------------
Hankel operators $H_u$ satisfy the key identity $$S^*H_u=H_uS;
\label{b2}$$ in fact, this identity characterises the class of all Hankel operators. Recalling that $$SS^*=I-(\cdot,{\mathbbm{1}}){\mathbbm{1}},$$ from and from $u=H_u{\mathbbm{1}}$ one obtains $$S^*H_u^2S=H_u^2-(\cdot,u)u.
\label{b3}$$ Multiplying by $S^*$ on the right and rearranging, we arrive at $$S^*H_u^2-H_u^2S^*=(\cdot,{\mathbbm{1}})S^*H_uu-(\cdot,Su)u.
\label{b4}$$ This relation is key to checking the definition of nearly $S^*$-invariance. Finally, it is straightforward to check that $H_u$ satisfies $$(H_uf,g)=(H_ug,f), \quad f,g\in {{\mathcal H}}^2.
\label{b5}$$
\[rmk\] Observe that Toeplitz operators $T_v$ on ${{\mathcal H}}^2$ satisfy the commutation relation $$S^*T_vS=T_v\, ;$$ formula can be viewed as a rank one perturbation of this relation.
Proof of the representation $E_{H_u}(s)=p{K}_\theta$ in the case $E_{H_u}(s)\not\perp{\mathbbm{1}}$ {#sec.b3}
---------------------------------------------------------------------------------------------------
Here we assume that $E_{H_u}(s)\not\perp{\mathbbm{1}}$ and prove the first part of Theorem \[thm1\].
1\) For $f\in E_{H_u}(s)\cap{\mathbbm{1}}^\perp$ and $g\in E_{H_u}(s)$ we have $$(S^*H_u^2f,g)-(H_u^2S^*f,g)
=
s^2(S^*f,g)-(S^*f,H_u^2g)
=
s^2(S^*f,g)-s^2(S^*f,g)=0$$ and therefore, by , $$(f,Su)(u,g)=0.$$ By assumption, there exists an element $h\in E_{H_u}(s)$ with $(h,{\mathbbm{1}})\not=0$. Take $g=H_uh$; then, using , $$(u,g)=(H_u{\mathbbm{1}},g)=(H_ug,{\mathbbm{1}})=(H_u^2h,{\mathbbm{1}})=s^2(h,{\mathbbm{1}})\not=0,$$ and so $(f,Su)=(S^*f,u)=0$. Now applying to $f$, we find $$(H_u^2-s^2 I)S^*f=0,$$ i.e. $S^*f\in E_{H_u}(s)$. Putting this together, we see that we have checked the inclusion $$S^*(E_{H_u}(s)\cap {\mathbbm{1}}^\perp)\subset E_{H_u}(s)\cap u^\perp.
\label{d3a}$$
2\) By Hitt’s theorem, $E_{H_u}(s)=pN$, where $p$ is an isometric multiplier on $N$ and $N$ is either a model space or $N={{\mathcal H}}^2$; we need to eliminate the second possibility. Suppose $N={{\mathcal H}}^2$. Since by assumption $E_{H_u}(s)\not\perp{\mathbbm{1}}$, we see that $p(0)\not=0$. Then $$p{{\mathcal H}}^2\cap {\mathbbm{1}}^\perp=zp{{\mathcal H}}^2.$$ It follows that $$S^*(p{{\mathcal H}}^2\cap {\mathbbm{1}}^\perp)=p{{\mathcal H}}^2.$$ Comparing with , we conclude that $u\perp E_{H_u}(s)$. Then for any $h\in E_{H_u}(s)$, $$s^2(h,{\mathbbm{1}})=(H_u^2 h,{\mathbbm{1}})=(H_u{\mathbbm{1}},H_uh)=(u,H_uh)=0,$$ and so $E_{H_u}(s)\perp{\mathbbm{1}}$, contrary to our assumption.
Proof of in the case $E_{H_u}(s)\not\perp{\mathbbm{1}}$
-------------------------------------------------------
Here we assume $E_{H_u}(s)\not\perp{\mathbbm{1}}$ and prove formula for the action of $H_u$ on $E_{H_u}(s)=p{K}_\theta$.
1\) Let us first assume that $\theta(0)=0$; then ${\mathbbm{1}}\in {K}_\theta$, $p\in p{K}_\theta$ and by the element $\overline{p(0)}p$ is the orthogonal projection of ${\mathbbm{1}}$ onto $p{K}_\theta$. Next, let $u_s$ be the orthogonal projection of $u$ onto $E_{H_u}(s)$. Since $H_u$ commutes with $H_u^2$, and therefore with the operator of the orthogonal projection onto $E_{H_u}(s)$, we see that $$u_s=H_u(\overline{p(0)}p)=p(0)H_up.
\label{d9}$$ Further, by , we have $$S^*(p{K}_\theta \cap p^\perp)\subset p{K}_\theta\cap u_s^\perp.
\label{d3aa}$$ Using Lemma \[lma.b1\], $$S^*(p {K}_{\theta}\cap p^\perp)
=
S^*(p({K}_{\theta}\cap{\mathbbm{1}}^\perp))
=
pS^*({K}_{\theta}\cap{\mathbbm{1}}^\perp)
=
p({K}_\theta\cap(S^*\theta)^\perp)
=
p{K}_\theta\cap (pS^*\theta)^\perp.$$ Comparing this with , we obtain $$u_s=cpS^*\theta
\label{d7}$$ with some constant $c$. Putting this together with , we get $$H_up=\frac{c}{p(0)}pS^*\theta.
\label{d8}$$ In order to evaluate $c$, let us compute the norms on both sides of : $$\begin{aligned}
{\lVertu_s\rVert}^2&={\lVertp(0)H_u p\rVert}^2={{\lvertp(0)\rvert}}^2
(H_u^2p,p)=s^2{{\lvertp(0)\rvert}}^2({\mathbbm{1}},{\mathbbm{1}})=s^2{{\lvertp(0)\rvert}}^2,
\\
{\lVertcpS^*\theta\rVert}&={{\lvertc\rvert}}{\lVertpS^*\theta\rVert}={{\lvertc\rvert}}{\lVertS^*\theta\rVert}={{\lvertc\rvert}}{\lVert\overline{z}\theta\rVert}={{\lvertc\rvert}}.\end{aligned}$$ It follows that ${{\lvertc\rvert}}=s{{\lvertp(0)\rvert}}$. Substituting this into , we obtain $$H_u p=se^{i\varphi}p\overline{z}\theta$$ with some unimodular complex number $e^{i\varphi}$. This is exactly the required formula for $h={\mathbbm{1}}$. From here we easily get formula for a general $h\in {K}_\theta$: $$H_u(p h)
=P(\overline{h}u\overline{p})
=P(\overline{h}P(u\overline{p}))
=P(\overline{h}H_u p)
=se^{i\varphi}P(\overline{h}p\overline{z}\theta)
=s e^{i\varphi}\overline{h}p\overline{z}\theta.$$
2\) Now let $E_{H_u}(s)=p{K}_{\theta}$, with $\theta(0)\not=0$. Choose $\alpha=\theta(0)$ and write $$p{K}_\theta=pg_\alpha {K}_{\theta_\alpha}$$ according to . Since $\theta_\alpha(0)=0$, by the previous step of the proof we have $$H_u(pg_\alpha h)=se^{i\varphi} \overline{z}pg_\alpha \theta_\alpha\overline{h},
\quad h\in {K}_{\theta_\alpha}.$$ Directly from the definitions one has $$g_\alpha \theta_\alpha=-\theta \overline{g_\alpha},$$ and so, denoting $g_\alpha h=v\in {K}_\theta$, we obtain $$H_u(pv)=-se^{i\varphi}\overline{z}\theta\overline{v},\quad v\in {K}_\theta,$$ as required.
The case $E_{H_u}(s)\perp{\mathbbm{1}}$ {#sec.c}
=======================================
Conformal maps {#sec.c1}
--------------
For $\alpha\in{{\mathbb D}}$, let $\mu:{{\mathbb D}}\to{{\mathbb D}}$ be the conformal map $$\mu(z)=\frac{\alpha-z}{1-\overline{\alpha}z},$$ and consider the corresponding unitary operator on the Hardy class, $$U_\mu f(z)=\frac{\sqrt{1-{{\lvert\alpha\rvert}}^2}}{1-\overline{\alpha}z}f(\mu(z)), \quad z\in{{\mathbb D}}.$$ Observe that $\mu$ is an involution, $\mu\circ\mu=\text{id}$ and $U_\mu^2=I$.
\[lma.c1\] Let $U_\mu$ be as defined above, and let $u\in\operatorname{BMOA}({{\mathbb T}})$. Then $$U_\mu H_uU_\mu=H_{w}, \quad\text{ where }\quad w=-S^*((Su)\circ\mu).$$
Computing the Jacobian of the change of variable $e^{it}\mapsto \mu(e^{it})$ on the unit circle, we get for $h_1,h_2\in {{\mathcal H}}^2$ $$(H_uh_1,h_2)=(u,h_1h_2)=(u\circ \mu,(h_1\circ \mu) (h_2\circ \mu) \tfrac{1-{{\lvert\alpha\rvert}}^2}{{{\lvert1-\overline{\alpha}z\rvert}}^2}).$$ Writing for ${{\lvertz\rvert}}=1$ $$\frac{1-{{\lvert\alpha\rvert}}^2}{{{\lvert1-\overline{\alpha}z\rvert}}^2}
=
-\left(\frac{\sqrt{1-{{\lvert\alpha\rvert}}^2}}{1-\overline{\alpha}z}\right)^2
\overline{(\overline{z}\mu(z))},$$ we get $$\begin{gathered}
(H_uh_1,h_2)=-\bigl(\overline{z}\mu(z)u\circ \mu, (U_\mu h_1)(U_\mu h_2)\bigr)
=
\bigl(w, (U_\mu h_1) (U_\mu h_2)\bigr)
\\
=
(H_{w}U_\mu h_1, U_\mu h_2)
=(U_\mu H_{w}U_\mu h_1,h_2). \qedhere\end{gathered}$$
\[lma.c2\] Let $\theta$ be an inner function and let $p$ be an isometric multiplier on $K_\theta$. Then $U_\mu(p{K}_\theta)=(p\circ\mu){K}_{\theta\circ\mu}$, and $p\circ\mu$ is an isometric multiplier on ${K}_{\theta\circ\mu}$.
Clearly, $U_\mu(p {K}_\theta)=(p\circ\mu)U_\mu({K}_\theta)$. Also, $U_\mu(\theta {{\mathcal H}}^2)=(\theta\circ\mu) {{\mathcal H}}^2$ and so $U_\mu {K}_\theta={K}_{\theta\circ\mu}$.
Proof of Theorem \[thm1\] in the case $E_{H_u}(s)\perp{\mathbbm{1}}$
--------------------------------------------------------------------
Let us choose $\alpha\in{{\mathbb D}}$ which is not a common zero of all elements of $E_{H_u}(s)$. Consider the conformal map $\mu$ and the unitary operator $U_\mu$ corresponding to this point $\alpha$. By the choice of $\alpha$, the point $z=0$ is not a common zero for $U_\mu E_{H_u}(s)$, i.e. $U_\mu E_{H_u}(s)\not\perp{\mathbbm{1}}$. Moreover, by Lemma \[lma.c1\], the latter subspace is a Schmidt subspace of a Hankel operator $H_w$, $$U_\mu E_{H_u}(s)
=\operatorname{Ker}(U_\mu H_u^2 U_\mu-s^2I)
=E_{H_w}(s)\, .$$ Thus, by the already proven case of Theorem \[thm1\], applied to $H_w$, we obtain that $$E_{H_w}(s)=p{K}_\theta,$$ where $p$ is an isometric multiplier on ${K}_\theta$, and that $H_w$ acts on $E_{H_w}(s)$ according to the formula : $$H_w(ph)=se^{i\varphi}p\overline{z}\theta \overline{h}, \quad h\in {K}_\theta.
\label{d6}$$ By Lemma \[lma.c2\] we obtain $$E_{H_u}(s)=U_\mu E_{H_w}(s)=U_\mu(p{K}_\theta)=(p\circ\mu){K}_{\theta\circ\mu}\, ,$$ which proves the first part of the theorem. It remains to check formula for the action of $H_u$.
Denote $U_\mu h=v\in {K}_{\theta\circ\mu}$. Let us apply $U_\mu$ on both sides of . For the left hand side, we have $$U_\mu H_w(ph)
=
U_\mu H_w(pU_\mu v)
=
U_\mu H_wU_\mu ((p\circ\mu)v)
=
H_u((p\circ\mu)v).$$ For the right hand side, we have $$U_\mu(se^{i\varphi} p\overline{z}\theta\overline{h})
=
se^{i\varphi}(p\circ\mu)(\theta\circ\mu)U_\mu(\overline{z}\overline{h}).$$ By the definition of $U_\mu$, $$\begin{aligned}
U_\mu(\overline{z}\overline{h})
&=
\frac{\sqrt{1-{{\lvert\alpha\rvert}}^2}}{1-\overline{\alpha}z}
\overline{\mu(z)}
\overline{h(\mu(z))}
=
\frac{\sqrt{1-{{\lvert\alpha\rvert}}^2}}{1-\overline{\alpha}z}
\frac{\overline{\alpha}-\overline{z}}{1-\alpha\overline{z}}
\overline{h(\mu(z))}
\\
&=
-\overline{z}
\frac{\sqrt{1-{{\lvert\alpha\rvert}}^2}}{1-\overline{\alpha}z}
\frac{1-\overline{\alpha}z}{1-\alpha\overline{z}}
\overline{h(\mu(z))}
=
-\overline{z}\overline{U_\mu h}\, .\end{aligned}$$ Putting this together, we obtain $$H_u((p\circ\mu)v)=-se^{i\varphi}(p\circ\mu)\overline{z}(\theta\circ\mu)\overline{v},$$ for all $v\in {K}_{\theta\circ\mu}$. This is the required formula .
[5]{}
*Multipliers between invariant subspaces of the backwards shift,* Pacific J. Math. **166**, no. 2 (1994), 225–256.
*Weighted model spaces and Schmidt subspaces of Hankel operators,* to appear in Journal of London Math. Soc.; preprint arXiv:1803.04295.
*The kernel of a Toeplitz operator,* Integral Equations Operator Theory **9** (1986), no. 4, 588–591.
*Invariant subspaces of ${{\mathcal H}}^2$ of an annulus,* Pacific Journal of Mathematics, **134**, no.1 (1988), 101–120.
*Nearly invariant subspaces of the backward shift,* Operator Theory: Advances and Applications, **35** (1988), 481–493.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'An orthogonal representation of a graph is an assignment of nonzero real vectors to its vertices such that distinct non-adjacent vertices are assigned to orthogonal vectors. We prove general lower bounds on the dimension of orthogonal representations of graphs using the Borsuk-Ulam theorem from algebraic topology. Our bounds strengthen the Kneser conjecture, proved by Lovász in 1978, and some of its extensions due to [Bárány]{}, Schrijver, [Donikov]{}, and [Kriz]{}. As applications, we determine the integrality gap of fractional upper bounds on the Shannon capacity of graphs and the quantum one-round communication complexity of certain promise equality problems.'
author:
- 'Ishay Haviv[^1]'
bibliography:
- 'kneser.bib'
title: '[**Topological Bounds on the Dimension of Orthogonal Representations of Graphs**]{}'
---
Introduction
============
A $t$-dimensional [*orthogonal representation*]{} of a graph $G=(V,E)$ over a field ${\mathbb{F}}$ is an assignment of a vector $u_i \in {\mathbb{F}}^t$ to each vertex $i \in V$ such that $\langle u_i,u_i \rangle \neq 0$ for every $i \in V$, and $\langle u_i,u_j \rangle = 0$ for every distinct non-adjacent vertices $i$ and $j$ in $G$. The [*orthogonality dimension*]{} of a graph $G$ over a field ${\mathbb{F}}$, denoted by $\xi_{\mathbb{F}}(G)$, is defined as the smallest integer $t$ for which there exists a $t$-dimensional orthogonal representation of $G$ over ${\mathbb{F}}$. It is easy to verify that the orthogonality dimension of a graph is sandwiched between its independence number and its clique cover number, that is, for every graph $G$ and a field ${\mathbb{F}}$, $\alpha(G) \leq \xi_{\mathbb{F}}(G) \leq \chi(\overline{G})$.
The notion of orthogonal representations over the real field was introduced by Lovász [@Lovasz79] in the study of the Shannon capacity of graphs and was later involved in a geometric characterization of connectivity properties of graphs by Lovász, Saks, and Schrijver [@LovaszSS89]. The orthogonality dimension over the complex field was used by de Wolf [@deWolfThesis] in a characterization of the quantum one-round communication complexity of promise equality problems and by Cameron et al. [@CameronMNSW07] in the study of the quantum chromatic number of graphs (see also [@ScarpaS12; @BrietBLPS15; @BrietZ17]). An extension of orthogonal representations, called orthogonal bi-representations, was introduced by Haemers [@Haemers81] (see also [@Peeters96]). Their smallest possible dimension, known as the minrank parameter of graphs, has found further applications in information theory, e.g., [@BBJK06; @mazumdar2014duality; @maleki2014index], and in theoretical computer science, e.g., [@Valiant92; @PudlakRS97; @HavivL13] (see Section \[sec:minrank\]).
The present paper provides lower bounds on the orthogonality dimension of graphs over the real and complex fields using topological methods. The use of topological methods in combinatorics was initiated in the study of the chromatic number of the Kneser graph defined as follows. For integers $d \geq 2s$, the Kneser graph $K(d,s)$ is the graph whose vertices are all the $s$-subsets of $[d]=\{1,\ldots,d\}$ where two sets are adjacent if they are disjoint. In 1955, Kneser [@Kneser55] observed that $K(d,s)$ admits a proper coloring with $d-2s+2$ colors, simply by coloring every set $A$ by the smallest integer in $A \cup \{d-2s+2\}$, and conjectured that fewer colors do not suffice. In 1978, Lovász [@LovaszKneser] confirmed the conjecture by a breakthrough application of a tool from algebraic topology, the Borsuk-Ulam theorem [@Borsuk33]. Since then, topological methods have led to additional important results in combinatorics, discrete geometry, and theoretical computer science. For an in-depth background to the topic we refer the interested reader to Matou[š]{}ek’s excellent book [@MatousekBook].
Following Lovász’s proof of the Kneser conjecture, several alternative proofs were given in the literature. The simplest known proof of the conjecture is the one of Greene [@Greene02], inspired by a proof found by [Bárány]{} [@Barany78] soon after Lovász’s. Other proofs were given by [Donikov]{} [@Dolnikov82], Sarkaria [@Sarkaria90], and Matou[š]{}ek [@Matousek04], where Matou[š]{}ek’s proof is the only one derived from a combinatorial argument (but the topological inspiration is still around). Schrijver considered in [@SchrijverKneser78] the graph $S(d,s)$ defined as the subgraph of $K(d,s)$ induced by the collection of all $s$-subsets of $[d]$ that include no consecutive integers modulo $d$ (that is, the $s$-subsets $A \subseteq [d]$ such that if $i \in A$ then $i+1 \notin A$, and if $d \in A$ then $1 \notin A$). It was shown in [@SchrijverKneser78] that $S(d,s)$ is a vertex-critical subgraph of $K(d,s)$, that is, its chromatic number is equal to that of $K(d,s)$ and a removal of any vertex of $S(d,s)$ decreases its chromatic number.
The various known proofs of the Kneser conjecture extend far beyond the chromatic number of the Kneser graph. Extensions of these proofs to lower bounds on the chromatic number of general graphs were given by [Donikov]{} [@Dolnikov82], by [Kriz]{} [@Kriz92], and by Matou[š]{}ek and Ziegler [@MatousekZ04] who generalized the proof techniques of Lovász, Sarkaria, and [Bárány]{}. Such extensions are usually stated for a generalized Kneser graph $K(\calF)$, defined as the graph whose vertex set is a set system $\calF$ where two sets are adjacent if they are disjoint. The generalized bounds are tight for the collection $\calF$ of all $s$-subsets of $[d]$ which corresponds to the graph $K(d,s)$, and some of them imply a tight lower bound on the chromatic number of the graph $S(d,s)$ as well. It is not difficult to see that every graph is isomorphic to $K(\calF)$ for some set system $\calF$ (see, e.g., [@MatousekZ04]), hence the bounds hold for all graphs (but for certain graphs they are quite weak). It was shown in [@MatousekZ04] that the extensions of the proofs of Lovász, Sarkaria, [Bárány]{}, [Donikov]{}, and [Kriz]{} can be (almost) linearly ordered by strength, where Lovász’s original proof technique is the strongest.
Topological Bounds on the Orthogonality Dimension
-------------------------------------------------
We prove two general lower bounds on the orthogonality dimension of graphs over the real and complex fields and on the minrank parameter over the real field (see Definition \[def:minrank\]). For convenience, we state the results for the complements of the generalized Kneser graphs $K(\calF)$. As mentioned before, every graph can be represented in this form. The two bounds are proved using the Borsuk-Ulam theorem from algebraic topology.
The statement of our first bound is purely combinatorial. It strengthens the lower bounds on the chromatic number obtained independently by [Donikov]{} [@Dolnikov82] and by [Kriz]{} [@Kriz92]. Our proof is inspired by the proof of the Kneser conjecture by Greene [@Greene02]. To state the bound, we need the following definition (see, e.g., [@MatousekBook Section 3.4]).
\[def:cd2\] Let $\calF$ be a set system with ground set $[d]$ such that $\emptyset \notin \calF$. The [*$2$-colorability-defect*]{} of $\calF$, denoted by ${\mathop{\mathrm{cd}}}_2(\calF)$, is the minimum size of a set $X \subseteq [d]$ such that the hypergraph on the vertex set $[d] \setminus X$ with the hyperedge set $\{ A \in \calF \mid A \cap X = \emptyset \}$ is $2$-colorable. Equivalently, ${\mathop{\mathrm{cd}}}_2(\calF)$ is the minimum number of white elements in a coloring of $[d]$ by red, blue and white, such that no set of $\calF$ is completely red or completely blue (but it may be completely white).
\[thm:cd\] For every set system $\calF$ such that $\emptyset \notin \calF$,
1. \[thm:cd\_itm:1\] $\xi_\R(\overline{K(\calF)}) \geq {\mathop{\mathrm{cd}}}_2(\calF)$,
2. \[thm:cd\_itm:2\] $\xi_\C(\overline{K(\calF)}) \geq \frac{{{\mathop{\mathrm{cd}}}}_2(\calF)}{2}$, and
3. \[thm:cd\_itm:3\] ${{\mathop{\mathrm{minrk}}}}_\R(\overline{K(\calF)}) \geq \sqrt{\frac{{{\mathop{\mathrm{cd}}}}_2(\calF)}{2}}$.
As an easy consequence of Theorem \[thm:cd\], we obtain that the orthogonality dimension of the complement of the Kneser graph $K(d,s)$ over the reals is $d-2s+2$ (see Corollary \[cor:xiKneser\]).
Our second bound has a geometric nature. Its proof employs the approach of [Bárány]{} [@Barany78] to the Kneser conjecture and strengthens a general lower bound on the chromatic number that follows from [@Barany78] and is given explicitly in [@MatousekZ04]. In what follows, $\S^t$ stands for the $t$-dimensional unit sphere $\{ x \in \R^{t+1} \mid \|x\|=1 \}$, and an open hemisphere of $\S^t$ is a set of the form $\{ z \in \S^t \mid \langle x,z \rangle >0 \}$ for some $x \in \S^t$.
\[thm:geo\] Let $\calF$ be a set system with ground set $[d]$ such that $\emptyset \notin \calF$. Suppose that for an integer $t \geq 2$ there exist points $y_1,\ldots,y_d \in \S^{t-2}$ such that every open hemisphere of $\S^{t-2}$ contains the points of $\{y_i \mid i \in A\}$ for some $A \in \calF$. Then,
1. \[thm:geom\_itm:1\] $\xi_\R(\overline{K(\calF)}) \geq t$,
2. \[thm:geom\_itm:2\] $\xi_\C(\overline{K(\calF)}) \geq \frac{t}{2}$, and
3. \[thm:geom\_itm:3\] ${{\mathop{\mathrm{minrk}}}}_\R(\overline{K(\calF)}) \geq \sqrt{\frac{t}{2}}$.
The theorem is used to prove that the orthogonality dimension of the complement of the Schrijver graph $S(d,s)$ over the reals is $d-2s+2$ (see Corollary \[cor:xiSchrijver\]). The proof technique of Theorem \[thm:geo\] is also used to prove a lower bound on the orthogonality dimension over the reals of the complement of the Borsuk graph defined by Erd[ő]{}s and Hajnal [@ErdosH67] (see Section \[sec:Borsuk\]).
Applications
------------
We describe below applications of our results to information theory and to quantum communication complexity.
### Shannon Capacity {#sec:Shannon}
The strong product $G_1 \cdot G_2$ of two graphs $G_1=(V_1,E_1)$ and $G_2=(V_2,E_2)$ is defined as the graph whose vertex set is $V_1 \times V_2$ where two distinct vertices $(u_1,u_2)$ and $(v_1,v_2)$ are adjacent if for every $i \in \{1,2\}$ the vertices $u_i$ and $v_i$ are either equal or adjacent in $G_i$. The $k$-th power of a graph $G$, denoted by $G^k$, is defined as the product of $k$ copies of $G$. The Shannon capacity of a graph $G$, introduced in 1956 by Shannon [@Shannon56], is the limit $c(G) = \lim_{k \rightarrow \infty}{(\alpha(G^k))^{1/k}}$ whose existence follows from super-multiplicativity and Fekete’s lemma. This graph parameter is motivated by a question in information theory on the capacity of noisy channels. Indeed, if $G$ is the graph whose vertices are the symbols that a channel can transmit, and two symbols are adjacent if they may be confused in the transmission, then $c(G)$ can be intuitively interpreted as the effective alphabet size of the channel. The Shannon capacity parameter of graphs is very far from being well understood. It is not known if the problem of deciding whether the Shannon capacity of an input graph exceeds a given value is decidable, and the exact Shannon capacity is unknown even for small and fixed graphs, e.g., the cycle on $7$ vertices.
Several upper bounds on the Shannon capacity of graphs were presented in the literature over the years. It is easy to see that $c(G) \leq \chi(\overline{G})$, and Shannon showed already in [@Shannon56] the stronger bound $c(G) \leq \chi_f(\overline{G})$, where $\chi_f$ stands for the fractional chromatic number. A useful way to obtain an upper bound on the Shannon capacity is to come up with a real-valued non-negative sub-multiplicative function on graphs that forms an upper bound on the independence number, that is, a function $f$ satisfying $f(G_1 \cdot G_2) \leq f(G_1) \cdot f(G_2)$ for every two graphs $G_1,G_2$ and $\alpha(G) \leq f(G)$ for every graph $G$. Indeed, for such an $f$ we have $$c(G) = \lim_{k \rightarrow \infty}{(\alpha(G^k))^{1/k}} \leq \lim_{k \rightarrow \infty}{(f(G^k))^{1/k}} \leq \lim_{k \rightarrow \infty}{(f(G)^k)^{1/k}} = f(G).$$ For example, it is not difficult to verify that the orthogonality dimension $\xi_{\mathbb{F}}$ over any field ${\mathbb{F}}$ is a sub-multiplicative upper bound on the independence number, hence $c(G) \leq \xi_{\mathbb{F}}(G)$ for every graph $G$. Other upper bounds on the Shannon capacity obtained in this way are the $\vartheta$-function due to Lovász [@Lovasz79], the ${{\mathop{\mathrm{minrk}}}}_{\mathbb{F}}$ parameter due to Haemers [@Haemers81], and the minimum dimension of polynomial representations due to Alon [@AlonUnion98].
In a recent work, Hu, Tamo, and Shayevitz [@HuTS17] defined for every function $f$ as above a fractional linear programming variant $f^*$. For a graph $G$ on the vertex set $V$, $f^*(G)$ is the value of the following linear program. $$\begin{aligned}
\label{Primal}
\begin{split}
\textup{maximize } &\sum_{x \in V} w(x) \\
\textup{subject to } & \sum_{x\in S}w(x) \le f(G[S]) \textup{~~~ for each set $S \subseteq V$},\\
& w(x)\ge 0 \textup{~~~ for each vertex $x \in V$},
\end{split}\end{aligned}$$ where $G[S]$ stands for the subgraph of $G$ induced by $S$. It was proved in [@HuTS17] that if $f$ is a sub-multiplicative upper bound on the independence number then so is $f^*$, hence $f^*$ also forms an upper bound on the Shannon capacity. Moreover, the upper bound $f^*$ is at least as strong as $f$, that is, $c(G) \leq f^*(G) \leq f(G)$ for every graph $G$ (see Section \[sec:Shannon4\]). It was shown in [@HuTS17] that the Lovász $\vartheta$-function satisfies $\vartheta(G)=\vartheta^*(G)$ for every graph $G$, whereas for other upper bounds $f$ on the Shannon capacity one can have $f^*(G) < f(G)$. For example, for every odd integer $n \geq 5$ the cycle $C_n$ satisfies ${{\mathop{\mathrm{minrk}}}}^*_\R(C_n) = \frac{n}{2} < \frac{n+1}{2} = {{\mathop{\mathrm{minrk}}}}_\R(C_n)$.
In this work, we aim to study the integrality gap of the fractional quantities $f^*$ as a function of the number of vertices. Namely, we would like to estimate the largest possible ratio $f(G)/f^*(G)$ over all $n$-vertex graphs $G$. We start with a general upper bound. A function $f$ on graphs is said to be sub-additive if for every graph $G$ on the vertex set $V$ and every sets $S_1$ and $S_2$ such that $V = S_1 \cup S_2$, $f(G) \leq f(G[S_1]) + f(G[S_2])$. The following theorem shows that for sub-additive functions $f$ the ratio between $f$ and $f^*$ is at most logarithmic in the number of vertices.
\[thm:fractional\_upper\] For every sub-additive function $f$ and every $n$-vertex graph $G$, $$f(G) \leq O(\log n) \cdot f^*(G).$$
It is easy to verify that all the aforementioned upper bounds on the Shannon capacity are sub-additive, hence their fractional variants cannot improve the bound on the Shannon capacity by more than a logarithmic multiplicative term.
As an application of our results on the Kneser graph, we obtain a matching lower bound on the integrality gap of the fractional orthogonality dimension over the real and complex fields.
\[thm:IntroIntegrality\] For every fixed ${{\varepsilon}}>0$, there exists an explicit family of $n$-vertex graphs $G$ such that $$\xi_\C^*(G) \leq \xi_\R^*(G) \leq 2+{{\varepsilon}}\mbox{~~whereas~~} \xi_\R (G)= \Theta(\log n) \mbox{~~and~~} \xi_\C (G)= \Theta(\log n).$$
We also show an unbounded integrality gap for the fractional minrank parameter over various fields (see Theorem \[thm:AllBounds\]).
### Quantum Communication Complexity
In the standard model of communication complexity, two parties Alice and Bob get inputs $x,y$ from two sets $\calX,\calY$ respectively, and they have to compute by a communication protocol the value of $g(x,y)$ for a two-variable function $g$. In a promise communication problem, the inputs are guaranteed to be drawn from a subset of $\calX \times \calY$ known to the parties in advance. In a one-round protocol, the communication flows only from Alice to Bob. The classical, respectively quantum, communication complexity of a problem is the minimal number of bits, respectively qubits, that the parties have to exchange on worst-case inputs in a communication protocol for the problem.
The orthogonality dimension of graphs over the complex field plays a central role in the study of the quantum communication complexity of promise equality problems.[^2] In such problems, Alice and Bob get either equal or adjacent vertices of a graph $G$ and their goal is to decide whether their inputs are equal. De Wolf [@deWolfThesis Section 8.5] showed that the classical one-round communication complexity of the promise equality problem associated with a graph $G$ is $\lceil \log_2 \chi(G) \rceil$, and that its quantum one-round communication complexity is $\lceil \log_2 \xi_\C(\overline{G}) \rceil$. Bri[ë]{}t et al. [@BrietBLPS15] proved that any classical protocol for such a problem can always be reduced to a classical one-round protocol with no extra communication, while in the quantum setting the one-round and two-round communication complexities of a promise equality problem can have an exponential gap. This separation was obtained using the Lovász $\vartheta$-function and the relation $\xi_\C(G) \geq \vartheta(G)$ (see [@BrietBLPS15 Lemma 2.5]).
For a set system $\calF$ with $\emptyset \notin \calF$, consider the promise equality problem in which Alice and Bob get either equal or disjoint sets from $\calF$, and their goal is to decide whether their inputs are equal. Observe that the graph associated with this problem is the generalized Kneser graph $K(\calF)$, hence its quantum one-round communication complexity is precisely $\lceil \log_2 \xi_\C(\overline{K(\calF)}) \rceil$. Our bounds on the orthogonality dimension of such graphs over $\C$ have applications to the quantum one-round communication complexity of promise equality problems, as demonstrated below.
For integers $d \geq 2s$, consider the communication complexity problem in which Alice and Bob get two $s$-subsets of $[d]$, their sets are guaranteed to be either equal or disjoint, and their goal is to decide whether the sets are equal. The graph associated with this problem is the Kneser graph $K(d,s)$, so its classical communication complexity is $\lceil \log_2 \chi(K(d,s)) \rceil = \lceil \log_2 (d-2s+2) \rceil$. As an application of Theorem \[thm:cd\], we get that $\lceil \log_2 \xi_\C(\overline{K(d,s)}) \rceil \geq \lceil \log_2 (d-2s+2)-1 \rceil$ (see Corollary \[cor:xiKneser\]), yielding the precise quantum one-round communication complexity of the problem up to an additive $1$. We note that the lower bound on $\xi_\C(\overline{K(d,s)})$ obtained from the Lovász $\vartheta$-function would not suffice here, since $\vartheta(\overline{K(d,s)}) = \frac{d}{s}$ (see [@Lovasz79]).
The orthogonality dimension over the complex field was also used by Bri[ë]{}t et al. [@BrietBLPS15] to characterize the quantum one-round communication complexity of a family of problems called list problems, originally studied by Witsenhausen [@Witsenhausen76]. In the list problem that corresponds to the Kneser graph $K(d,s)$, Alice gets an $s$-subset $A$ of $[d]$, Bob gets a list of pairwise disjoint $s$-subsets of $[d]$ that includes the set $A$, and his goal is to discover $A$. It follows from [@BrietBLPS15] that the quantum one-round communication complexity of this problem is equal to the quantity $\lceil \log_2 \xi_\C(\overline{K(d,s)}) \rceil$ determined above.
Outline
-------
The rest of the paper is organized as follows. In Section \[sec:preliminaries\] we provide some background on the Borsuk-Ulam theorem and on the minrank parameter of graphs. In Section \[sec:bounds\] we prove Theorems \[thm:cd\] and \[thm:geo\] and obtain our results on the Kneser, Schrijver, and Borsuk graphs. Finally, in Section \[sec:Shannon4\] we study the integrality gap of fractional upper bounds on the Shannon capacity and prove Theorems \[thm:fractional\_upper\] and \[thm:IntroIntegrality\].
Preliminaries {#sec:preliminaries}
=============
The Borsuk-Ulam Theorem {#sec:BU}
-----------------------
For an integer $t \geq 0$, let $\S^{t} = \{ x \in \R^{t+1} \mid \|x\|=1 \}$ denote the $t$-dimensional unit Euclidean sphere. For a point $x \in \S^t$, let $H(x) = \{ z \in \S^t \mid \langle x,z \rangle >0 \}$ denote the open hemisphere of $\S^t$ centered at $x$. We state below the Borsuk-Ulam theorem, proved by Borsuk in 1933 [@Borsuk33].
\[thm:BU\] For every continuous function $f: \S^t \rightarrow \R^t$ there exists $x \in \S^t$ such that $f(x)=f(-x)$.
Equivalently, if a continuous function $f: \S^t \rightarrow \R^{t'}$ satisfies $f(x) \neq f(-x)$ for all $x \in \S^t$ then $t' > t$. For several other equivalent versions of Theorem \[thm:BU\], see [@MatousekBook Section 2.1].
We also need a variant of the Borsuk-Ulam theorem for complex-valued functions. We start with some notations. For a complex number $z \in \C$ we denote by ${\mathsf{Re}}(z)$ and ${\mathsf{Im}}(z)$ the real and imaginary parts of $z$ respectively, hence $z = {\mathsf{Re}}(z)+{\mathsf{Im}}(z) \cdot i$. For an integer $t \geq 1$, let $\phi_t:\C^t \rightarrow \R^{2t}$ be the natural embedding of $\C^t$ in $\R^{2t}$ defined by $$\begin{aligned}
\label{eq:phi}
\phi_t(x) = ({\mathsf{Re}}(x_1),{\mathsf{Im}}(x_1),\ldots,{\mathsf{Re}}(x_t),{\mathsf{Im}}(x_t)) \in \R^{2t}.\end{aligned}$$ Clearly, $\phi_t$ is a bijection from $\C^t$ to $\R^{2t}$. Notice that we have $\phi_t(-x)=-\phi_t(x)$ for every $x \in \C^t$. Our variant of the Borsuk-Ulam theorem for complex-valued functions is given below.
\[thm:BU\_complex\] For every continuous function $f: \S^{2t} \rightarrow \C^t$ there exists $x \in \S^{2t}$ such that $f(x)=f(-x)$.
For a continuous function $f: \S^{2t} \rightarrow \C^t$ consider the function $\widetilde{f}: \S^{2t} \rightarrow \R^{2t}$ defined by the composition $\widetilde{f} = \phi_t \circ f$. Applying Theorem \[thm:BU\] to $\widetilde{f}$, we get that there exists $x \in \S^{2t}$ such that $\widetilde{f}(x)=\widetilde{f}(-x)$. By the invertibility of $\phi_t$, this implies that $f(x) = f(-x)$, as required.
Minrank {#sec:minrank}
-------
The minrank parameter of graphs, introduced by Haemers in [@Haemers81], is defined as follows.
\[def:minrank\] Let $G$ be a graph on the vertex set $V = [n]$ and let ${\mathbb{F}}$ be a field. We say that an $n \times n$ matrix $M$ over ${\mathbb{F}}$ [*represents*]{} $G$ if $M_{i,i} \neq 0$ for every $i \in V$, and $M_{i,j}=0$ for every distinct non-adjacent vertices $i,j \in V$. The [*minrank*]{} of $G$ over ${\mathbb{F}}$ is defined as $${{\mathop{\mathrm{minrk}}}}_{\mathbb{F}}(G) = \min\{{{\mathop{\mathrm{rank}}}}_{{\mathbb{F}}}(M)\mid M \mbox{ represents }G\mbox{ over }{\mathbb{F}}\}.$$
The minrank parameter can be equivalently defined in terms of orthogonal bi-representations. A $t$-dimensional [*orthogonal bi-representation*]{} of a graph $G=(V,E)$ over a field ${\mathbb{F}}$ is an assignment of a pair $(u_i,v_i) \in {\mathbb{F}}^t \times {\mathbb{F}}^t$ to each vertex $i \in V$ such that $\langle u_i,v_i \rangle \neq 0$ for every $i \in V$, and $\langle u_i,v_j \rangle = \langle u_j,v_i \rangle = 0$ for every distinct non-adjacent vertices $i$ and $j$ in $G$. It can be verified that ${{\mathop{\mathrm{minrk}}}}_{\mathbb{F}}(G)$ is the smallest integer $t$ for which there exists a $t$-dimensional orthogonal bi-representation of $G$ over ${\mathbb{F}}$ (see, e.g., [@Peeters96; @ChlamtacH14]). Since orthogonal bi-representations generalize orthogonal representations, we clearly have ${{\mathop{\mathrm{minrk}}}}_{\mathbb{F}}(G) \leq \xi_{\mathbb{F}}(G)$ for all graphs $G$ and fields ${\mathbb{F}}$.
The minrank parameter is always bounded from above by the clique cover number. The following lemma shows a lower bound on the minrank parameter over finite fields in terms of the clique cover number. Its proof is implicit in [@LangbergS08] and we give here a quick proof for completeness.
\[claim:minrankF\] For every graph $G$ and a finite field ${\mathbb{F}}$, ${{\mathop{\mathrm{minrk}}}}_{\mathbb{F}}(G) \geq \log_{|{\mathbb{F}}|}\chi(\overline{G})$.
Denote $t = {{\mathop{\mathrm{minrk}}}}_{\mathbb{F}}(G)$. Then there exists an assignment of a pair $(u_i,v_i) \in {\mathbb{F}}^t \times {\mathbb{F}}^t$ to each vertex $i \in V$ that forms an orthogonal bi-representation of $G$ over ${\mathbb{F}}$. Consider the coloring that assigns to every vertex $i \in V$ the color $u_i \in {\mathbb{F}}^t$. We claim that this is a proper coloring of $\overline{G}$. Indeed, for two vertices $i$ and $j$ adjacent in $\overline{G}$ we have $\langle u_i,v_i \rangle \neq 0$ and $\langle u_j,v_i \rangle = 0$, hence $u_i \neq u_j$. Since the number of used colors is at most $|{\mathbb{F}}|^t$, it follows that $\chi(\overline{G}) \leq |{\mathbb{F}}|^t$, as required.
Topological Bounds on the Orthogonality Dimension {#sec:bounds}
=================================================
In this section we prove Theorems \[thm:cd\] and \[thm:geo\] and obtain our results on the Kneser, Schrijver, and Borsuk graphs. The proofs employ the Borsuk-Ulam theorem given in Section \[sec:BU\].
Proof of Theorem \[thm:cd\]
---------------------------
We start with the lower bound on the orthogonality dimension over the real field.
Let $\calF$ be a set system with ground set $[d]$ such that $\emptyset \notin \calF$, and put $t = \xi_\R(\overline{K(\calF)})$. Then there exists an assignment of a nonzero vector $u_A \in \R^t$ to every set $A \in \calF$, such that $\langle u_A,u_B \rangle = 0$ for every disjoint sets $A,B \in \calF$. It can be assumed without loss of generality that the first nonzero coordinate in every vector $u_A$ is positive (otherwise replace $u_A$ by $-u_A$).
Let $y_1,\ldots,y_d \in \S^t$ be $d$ points in a general position, that is, no $t+1$ of them lie on a $(t-1)$-dimensional sphere. For a set $A \subseteq [d]$ denote $y_A = \{ y_i \mid i \in A \}$. Define a function $f:\S^t \rightarrow \R^t$ by $$f(x) = \sum_{A \in \calF}{u_A \cdot \prod_{j \in A}{\max(\langle x, y_j\rangle,0)}}.$$ Observe that for every $x \in \S^{t}$, $f(x)$ is a linear combination with positive coefficients of the vectors $u_A$ such that $y_A \subseteq H(x)$, where $H(x)$ is the open hemisphere of $\S^t$ centered at $x$. The function $f$ is clearly continuous, hence by Theorem \[thm:BU\] there exists $x \in \S^t$ such that $f(x) = f(-x)$. However, $f(x)$ is a linear combination of the vectors $u_A$ with $y_A \subseteq H(x)$ whereas $f(-x)$ is a linear combination of the vectors $u_A$ with $y_A \subseteq H(-x)$. Since $H(x) \cap H(-x) = \emptyset$, it follows that the sets $A$ involved in the linear combination of $f(x)$ are all disjoint from those involved in the linear combination of $f(-x)$. The fact that $\langle u_A,u_B \rangle = 0$ for every disjoint sets $A,B \in \calF$ yields that the vectors $f(x)$ and $f(-x)$ are orthogonal, and by $f(x) = f(-x)$ they must be equal to the zero vector.
We claim now that there is no $A \in \calF$ with $y_A \subseteq H(x)$. To see this, assume in contradiction that $A_1, \ldots, A_m$ are the sets with this property ($m \geq 1$), and let $j \in [t]$ be the least coordinate in which at least one of the vectors $u_{A_1},\ldots,u_{A_m}$ is nonzero. Since $f(x)$ is a linear combination of these vectors with positive coefficients, using the assumption that the first nonzero coordinate of every $u_{A}$ is positive, it follows that the $j$th coordinate of $f(x)$ is positive in contradiction to $f(x)$ being the zero vector. By the same reasoning, there is no $A \in \calF$ with $y_A \subseteq H(-x)$.
Finally, let $X$ denote the set of indices $i \in [d]$ for which $y_i$ does not belong to $H(x)$ nor to $H(-x)$. By the assumption of general position, $|X| \leq t$. We color the elements of $[d] \setminus X$ as follows: If $y_i \in H(x)$ then $i$ is colored red, and if $y_i \in H(-x)$ then $i$ is colored blue. Since no set $A \in \calF$ satisfies $y_A \subseteq H(x)$ or $y_A \subseteq H(-x)$, we get a proper $2$-coloring of the hypergraph $([d] \setminus X, \{ A \in \calF \mid A \cap X = \emptyset \})$. This implies that ${\mathop{\mathrm{cd}}}_2(\calF) \leq t$, as required.
We next prove our lower bound on the orthogonality dimension over the complex field. Its proof is similar to the proof over the reals but requires the Borsuk-Ulam theorem for complex-valued functions (Theorem \[thm:BU\_complex\]). Recall that $\phi_t$ stands for the natural embedding of $\C^t$ in $\R^{2t}$ given in .
Let $\calF$ be a set system with ground set $[d]$ such that $\emptyset \notin \calF$, and put $t = \xi_\C(\overline{K(\calF)})$. Then there exists an assignment of a nonzero vector $u_A \in \C^t$ to every set $A \in \calF$, such that $\langle u_A,u_B \rangle = 0$ for every disjoint sets $A,B \in \calF$. It can be assumed without loss of generality that the first nonzero coordinate in every vector $\phi_t(u_A)$ is positive (otherwise replace $u_A$ by $-u_A$).
Let $y_1,\ldots,y_d \in \S^{2t}$ be $d$ points in a general position. As before, for a set $A \subseteq [d]$ denote $y_A = \{ y_i \mid i \in A \}$. Define a function $f:\S^{2t} \rightarrow \C^t$ by $$f(x) = \sum_{A \in \calF}{u_A \cdot \prod_{j \in A}{\max(\langle x, y_j\rangle,0)}}.$$ Observe that for every $x \in \S^{2t}$, $f(x)$ is a linear combination with real positive coefficients of the vectors $u_A$ satisfying $y_A \subseteq H(x)$. The function $f$ is clearly continuous, hence by Theorem \[thm:BU\_complex\] there exists $x \in \S^{2t}$ such that $f(x) = f(-x)$. However, $f(x)$ is a linear combination of the vectors $u_A$ with $y_A \subseteq H(x)$ whereas $f(-x)$ is a linear combination of the vectors $u_A$ with $y_A \subseteq H(-x)$. Since $H(x) \cap H(-x) = \emptyset$, it follows that the sets $A$ involved in the linear combination of $f(x)$ are all disjoint from those involved in the linear combination of $f(-x)$. The fact that $\langle u_A,u_B \rangle = 0$ for every disjoint sets $A,B \in \calF$ yields that the vectors $f(x)$ and $f(-x)$ are orthogonal, and by $f(x) = f(-x)$ they must be equal to the zero vector.
We claim now that there is no $A \in \calF$ with $y_A \subseteq H(x)$. To see this, assume in contradiction that $A_1, \ldots, A_m$ are the sets with this property ($m \geq 1$), and let $j \in [2t]$ be the least coordinate in which at least one of the vectors $\phi_{t}(u_{A_1}),\ldots,\phi_{t}(u_{A_m})$ is nonzero. It follows that the $j$th coordinate of $\phi_t(f(x))$ is positive in contradiction to $f(x)$ being the zero vector. By the same reasoning, there is no $A \in \calF$ with $y_A \subseteq H(-x)$.
Finally, let $X$ denote the set of indices $i \in [d]$ for which $y_i$ does not belong to $H(x)$ nor to $H(-x)$. By the assumption of general position, $|X| \leq 2t$. We color the elements of $[d] \setminus X$ as follows: If $y_i \in H(x)$ then $i$ is colored red, and if $y_i \in H(-x)$ then $i$ is colored blue. Since no set $A \in \calF$ satisfies $y_A \subseteq H(x)$ or $y_A \subseteq H(-x)$, we get a proper $2$-coloring of the hypergraph $([d] \setminus X, \{ A \in \calF \mid A \cap X = \emptyset \})$. This implies that ${\mathop{\mathrm{cd}}}_2(\calF) \leq 2t$, as required.
Finally, we prove our lower bound on the minrank parameter over the real field (recall Definition \[def:minrank\]). We start with the following lemma. Here, a real matrix is said to be non-negative if all of its entries are non-negative.
\[lemma:minrank\_nonneg\] Let $\calF$ be a set system such that $\emptyset \notin \calF$, and let $M$ be a real non-negative matrix that represents the graph $\overline{K(\calF)}$ over $\R$. Then, ${{\mathop{\mathrm{rank}}}}_\R(M) \geq \frac{1}{2} \cdot {\mathop{\mathrm{cd}}}_2(\calF)$.
Let $\calF$ be a set system of size $n$ with ground set $[d]$ such that $\emptyset \notin \calF$, let $M \in \R^{n \times n}$ be a non-negative matrix that represents the graph $\overline{K(\calF)}$ over $\R$, and put $t = {{\mathop{\mathrm{rank}}}}_\R(M)$. Write $M = M_1^T \cdot M_2$ for matrices $M_1,M_2 \in \R^{t \times n}$. For every set $A \in \calF$, let $u_A$ and $v_A$ be the $t$-dimensional columns associated with $A$ in $M_1$ and $M_2$ respectively, and let $w_A = u_A \circ v_A$ be the $2t$-dimensional concatenation of $u_A$ and $v_A$. Since $M$ represents $\overline{K(\calF)}$, we have $\langle u_A,v_A \rangle \neq 0$ for every $A \in \calF$, and $\langle u_A,v_B \rangle = 0$ for every disjoint sets $A,B \in \calF$. By the assumption that $M$ is non-negative, we also have $\langle u_A,v_B \rangle \geq 0$ for all $A,B \in \calF$.
Let $y_1,\ldots,y_d \in \S^{2t}$ be $d$ points in a general position. As before, for a set $A \subseteq [d]$ denote $y_A = \{ y_i \mid i \in A \}$. Define a function $f:\S^{2t} \rightarrow \R^{2t}$ by $$f(x) = \sum_{A \in \calF}{w_A \cdot \prod_{j \in A}{\max(\langle x, y_j\rangle,0)}}.$$ Observe that for every $x \in \S^{2t}$, $f(x)$ is a linear combination with positive coefficients of the vectors $w_A$ such that $y_A \subseteq H(x)$. The function $f$ is clearly continuous, hence by Theorem \[thm:BU\] there exists $x \in \S^{2t}$ such that $f(x) = f(-x)$. For this $x$, denote $f(x) = f(-x) = w = w_1 \circ w_2$ where $w_1,w_2 \in \R^t$. By $f(x)=w$ we get that $w_1$ is a linear combination of the vectors $u_A$ with $y_A \subseteq H(x)$, and by $f(-x)=w$ we get that $w_2$ is a linear combination of the vectors $v_A$ with $y_A \subseteq H(-x)$. However, the sets $A$ with $y_A \subseteq H(x)$ are all disjoint form the sets $A$ with $y_A \subseteq H(-x)$, hence the fact that $\langle u_A,v_B \rangle = 0$ for every disjoint sets $A,B \in \calF$ yields that the vectors $w_1$ and $w_2$ are orthogonal.
We claim now that there is no $A \in \calF$ with $y_A \subseteq H(x)$. To see this, assume in contradiction that $A_1, \ldots, A_m$ are the sets with this property ($m \geq 1$). By the definition of $f$ we can write $w = f(x)= \sum_{i=1}^{m}{c_i \cdot w_{A_i}}$ for some positive coefficients $c_i>0$. However, this implies that $$\langle w_1,w_2\rangle = \langle \sum_{i=1}^{m}{c_i \cdot u_{A_i}} ~, \sum_{i=1}^{m}{c_i \cdot v_{A_i}}\rangle = \sum_{1 \leq i,j \leq m}{c_i c_j \cdot \langle u_{A_i},v_{A_j} \rangle} > 0,$$ where the inequality holds since $\langle u_{A_i},v_{A_j} \rangle \geq 0$ for all pairs $i,j$ and $\langle u_{A_i},v_{A_i} \rangle > 0$ for every $i$. This is in contradiction to the fact that the vectors $w_1$ and $w_2$ are orthogonal. By the same reasoning, there is no $A \in \calF$ with $y_A \subseteq H(-x)$.
Finally, let $X$ denote the set of indices $i \in [d]$ for which $y_i$ does not belong to $H(x)$ nor to $H(-x)$. By the assumption of general position, $|X| \leq 2t$. We color the elements of $[d] \setminus X$ as follows: If $y_i \in H(x)$ then $i$ is colored red, and if $y_i \in H(-x)$ then $i$ is colored blue. Since no set $A \in \calF$ satisfies $y_A \subseteq H(x)$ or $y_A \subseteq H(-x)$, we get a proper $2$-coloring of the hypergraph $([d] \setminus X, \{ A \in \calF \mid A \cap X = \emptyset \})$. This implies that ${\mathop{\mathrm{cd}}}_2(\calF) \leq 2t$, as required.
Equipped with Lemma \[lemma:minrank\_nonneg\], we are ready to complete the proof of Theorem \[thm:cd\].
Let $\calF$ be a set system of size $n$ such that $\emptyset \notin \calF$, and let $M$ be an $n \times n$ matrix that represents the graph $\overline{K(\calF)}$ over $\R$. Consider the $n \times n$ matrix $M'$ defined by $M'_{i,j} = M_{i,j}^2$ for all $i,j$. It is well known and easy to check that $M'$ is a principal sub-matrix of the tensor product $M \otimes M$ of $M$ with itself, hence $${{\mathop{\mathrm{rank}}}}_\R (M') \leq {{\mathop{\mathrm{rank}}}}_\R (M \otimes M) = {{\mathop{\mathrm{rank}}}}_\R (M)^2.$$ The non-negative matrix $M'$ represents $\overline{K(\calF)}$ since it has the same zero pattern as $M$, so we can apply Lemma \[lemma:minrank\_nonneg\] to obtain that $${{\mathop{\mathrm{rank}}}}_\R(M) \geq \sqrt{{{\mathop{\mathrm{rank}}}}_\R(M')} \geq \sqrt{\frac{{{\mathop{\mathrm{cd}}}}_2(\calF)}{2}},$$ completing the proof.
### The Kneser Graph {#sec:Kneser}
Recall that for integers $d \geq 2s$, the Kneser graph $K(d,s)$ is the graph whose vertices are all the $s$-subsets of $[d]$, where two sets are adjacent if they are disjoint. We need the following simple claim (see, e.g., [@MatousekBook Section 3.4]).
\[claim:cd2\_Kneser\] For integers $d \geq 2s$, let $\calF$ be the collection of all $s$-subsets of $[d]$. Then, ${\mathop{\mathrm{cd}}}_2(\calF) = d-2s+2$.
Let $X \subseteq [d]$ be an arbitrary set of size $d-2s+2$, and consider an arbitrary balanced $2$-coloring of the $2s-2$ elements of $[d] \setminus X$. Clearly, no $s$-subset of $[d] \setminus X$ is monochromatic, hence ${\mathop{\mathrm{cd}}}_2(\calF) \leq d-2s+2$. For the other direction, notice that for every $X \subseteq [d]$ of size at most $d-2s+1$ there are at least $2s-1$ elements in $[d] \setminus X$, hence every $2$-coloring of $[d] \setminus X$ includes a monochromatic $s$-subset. This implies that ${\mathop{\mathrm{cd}}}_2(\calF) \geq d-2s+2$ and completes the proof.
The following corollary summarizes our bounds for the Kneser graph.
\[cor:xiKneser\] For every integers $d \geq 2s$,
1. \[itm:Kneser\] $\xi_\R(\overline{K(d,s)}) = d-2s+2$,
2. $\xi_\C(\overline{K(d,s)}) \geq (d-2s+2)/2$, and
3. ${{\mathop{\mathrm{minrk}}}}_\R(\overline{K(d,s)}) \geq \sqrt{(d-2s+2)/2}$.
Notice that $K(d,s)$ is the graph $K(\calF)$ where $\calF$ is the collection of all $s$-subsets of $[d]$. The three lower bounds follow directly by combining Theorem \[thm:cd\] with Claim \[claim:cd2\_Kneser\]. The matching upper bound in Item \[itm:Kneser\] follows by $\xi_\R(\overline{K(d,s)}) \leq \chi(K(d,s)) = d-2s+2$.
Proof of Theorem \[thm:geo\]
----------------------------
We prove below Item \[thm:geom\_itm:1\] of Theorem \[thm:geo\]. The other two items follow similarly, using ideas from the proofs of Items \[thm:cd\_itm:2\] and \[thm:cd\_itm:3\] of Theorem \[thm:cd\]. To avoid repetitions, we omit the details.
Let $\calF$ be a set system with ground set $[d]$ such that $\emptyset \notin \calF$, and let $y_1,\ldots,y_d \in \S^{t-2}$ be the points given in the theorem. Put $t' = \xi_\R(\overline{K(\calF)})$. Then there exists an assignment of a nonzero vector $u_A \in \R^{t'}$ to every set $A \in \calF$, such that $\langle u_A,u_B \rangle = 0$ for every disjoint sets $A,B \in \calF$. It can be assumed without loss of generality that the first nonzero coordinate in every vector $u_A$ is positive (otherwise replace $u_A$ by $-u_A$).
Define a function $f:\S^{t-2} \rightarrow \R^{t'}$ by $$f(x) = \sum_{A \in \calF}{u_A \cdot \prod_{j \in A}{\max(\langle x, y_j\rangle,0)}}.$$ For any $x \in \S^{t-2}$, let $C_x$ be the collection of sets $A \in \calF$ such that $y_A \subseteq H(x)$, where, as before, $y_A = \{ y_i \mid i \in A \}$. By the assumption on the points $y_1,\ldots,y_d$ we have $|C_x| \geq 1$. Observe that $f(x)$ is a linear combination with positive coefficients of the vectors $u_A$ with $A \in C_x$. Letting $j \in [t']$ be the least coordinate in which at least one of the vectors of $\{ u_{A} \mid A \in C_x \}$ is nonzero, it follows that the $j$th coordinate of $f(x)$ is positive, hence $f(x)$ is nonzero. Further, by $H(x) \cap H(-x) = \emptyset$ we get that the sets of $C_x$ are all disjoint from those of $C_{-x}$, hence the fact that $\langle u_A,u_B \rangle = 0$ for every disjoint sets $A,B \in \calF$ yields that the vectors $f(x)$ and $f(-x)$ are orthogonal.
Consider the function $g:\S^{t-2} \rightarrow \S^{t'-1}$ defined by $$g(x) = \frac{f(x)}{\|f(x)\|}.$$ Note that $g$ is well defined as $f(x)$ is nonzero for every $x \in \S^{t-2}$. Consider also the function $\widetilde{g} : \S^{t-2} \rightarrow \R^{t'-1}$ that maps every $x \in \S^{t-2}$ to the projection of $g(x)$ to its last $t'-1$ coordinates (i.e., all of its coordinates besides the first one). We claim that there is no $x \in \S^{t-2}$ such that $\widetilde{g}(x)=\widetilde{g}(-x)$. To see this, notice that $g(x)$ is a unit vector whose first entry is non-negative, so the projection of $g(x)$ to its last $t'-1$ coordinates fully determines $g(x)$. This implies that if there exists an $x \in \S^{t-2}$ satisfying $\widetilde{g}(x)=\widetilde{g}(-x)$ then this $x$ also satisfies $g(x)=g(-x)$, in contradiction to the orthogonality of $f(x)$ and $f(-x)$. Since $\widetilde{g}$ is continuous we can apply Theorem \[thm:BU\] to derive that $t'-1 > t-2$ which implies that $t' \geq t$ and completes the proof.
### The Schrijver Graph
We say that a set $A \subseteq [d]$ is [*stable*]{} if it does not contain two consecutive elements modulo $d$ (that is, if $i \in A$ then $i+1 \notin A$, and if $d \in A$ then $1 \notin A$). In other words, a stable subset of $[d]$ is an independent set in the cycle $C_d$ with the numbering from $1$ to $d$ along the cycle. Recall that for $d \geq 2s$, the Schrijver graph $S(d,s)$ is the graph whose vertices are all the stable $s$-subsets of $[d]$, where two sets are adjacent if they are disjoint.
We need the following strengthening of a lemma of Gale [@Gale56] proved by Schrijver in [@SchrijverKneser78]. See [@MatousekBook Section 3.5] for a nice proof by Ziegler based on the moment curve.
\[lemma:Gale\] For every integers $d \geq 2s$, there exist points $y_1,\ldots,y_d \in \S^{d-2s}$ such that every open hemisphere of $\S^{d-2s}$ contains the points of $\{ y_i \mid i \in A \}$ for some stable $s$-subset $A$ of $[d]$.
For $d \geq 2s$, consider the collection $\calF$ of all stable $s$-subsets of $[d]$, and notice that $S(d,s)$ is the graph $K(\calF)$. By Lemma \[lemma:Gale\], $\calF$ satisfies the condition of Theorem \[thm:geo\] for $t=d-2s+2$. This directly implies the following corollary which summarizes our bounds for the Schrijver graph.
\[cor:xiSchrijver\] For every integers $d \geq 2s$,
1. \[itm:Kneser\] $\xi_\R(\overline{S(d,s)}) = d-2s+2$,
2. $\xi_\C(\overline{S(d,s)}) \geq (d-2s+2)/2$, and
3. ${{\mathop{\mathrm{minrk}}}}_\R(\overline{S(d,s)}) \geq \sqrt{(d-2s+2)/2}$.
We note that the bounds that Theorem \[thm:cd\] implies for the Schrijver graph $S(d,s)$ are weaker than the bounds obtained above. Indeed, it is easy to check that the set system $\calF$ that corresponds to the graph $S(d,s)$ satisfies ${\mathop{\mathrm{cd}}}_2(\calF) = d-4s+4$. For a discussion comparing the bounds derived from Theorems \[thm:cd\] and \[thm:geo\], see [@MatousekZ04 Section 6].
### The Borsuk Graph {#sec:Borsuk}
For $0< \alpha <2$ and an integer $d$, the Borsuk graph $B(d,\alpha)$ is defined as the (infinite) graph on the vertex set $\S^{d-1}$ where two points $y,y' \in \S^{d-1}$ are adjacent if $\|y-y'\| \geq \alpha$. It is known that the Borsuk-Ulam theorem implies that $\chi(B(d,\alpha)) \geq d+1$ for every $0< \alpha <2$ and $d$, and that this bound is tight whenever $\alpha \geq \sqrt{2(d+1)/d}$ (see, e.g., [@Lovasz83]). We apply here the proof technique of Theorem \[thm:geo\] to obtain the same bound on the orthogonality dimension over $\R$ of the complement of $B(d,\alpha)$. We first prove the following.
\[thm:BorsukGraph\] For $0< {{\varepsilon}}<1$ and an integer $d$, let $V \subseteq \S^{d-1}$ be a finite collection of points in $\S^{d-1}$ such that for every $x \in \S^{d-1}$ there exists $y \in V$ for which $\|x-y\| < {{\varepsilon}}$. Let $G$ be the graph on the vertex set $V$ where two points $y,y' \in V$ are adjacent if $\|y-y'\| \geq 2-2{{\varepsilon}}$. Then, $\xi_\R(\overline{G}) \geq d+1$.
For a graph $G$ as in the theorem, put $t = \xi_\R(\overline{G})$. Then there exists an assignment of a nonzero vector $u_y \in \R^t$ to every point $y \in V$, such that $\langle u_y,u_{y'} \rangle = 0$ for every points $y,y' \in V$ satisfying $\|y-y'\| \geq 2-2{{\varepsilon}}$. It can be assumed without loss of generality that the first nonzero coordinate in every vector $u_y$ is positive (otherwise replace $u_y$ by $-u_y$).
Define a function $f:\S^{d-1} \rightarrow \R^t$ by $$f(x) = \sum_{y \in V}{u_y \cdot \max ( {{\varepsilon}}- \|x-y\| ,0 )}.$$ For an $x \in \S^{d-1}$, let $C_x \subseteq V$ be the set of all points $y \in V$ such that $\|x-y\| < {{\varepsilon}}$. By the assumption on $V$ we have $|C_x| \geq 1$. Observe that $f(x)$ is a linear combination with positive coefficients of the vectors $u_y$ with $y \in C_x$. Letting $j \in [t]$ be the least coordinate in which at least one of the vectors of $\{ u_{y} \mid y \in C_x \}$ is nonzero, using the assumption that the first nonzero coordinate of every $u_{y}$ is positive, it follows that the $j$th coordinate of $f(x)$ is positive, hence $f(x)$ is nonzero. Further, since the distance between $x$ and $-x$ is $2$, for every $y \in C_x$ and $y' \in C_{-x}$ we have $\|y-y'\| \geq 2-2{{\varepsilon}}$, and thus $\langle u_y, u_{y'} \rangle = 0$. This implies that the vectors $f(x)$ and $f(-x)$ are orthogonal.
Consider the function $g:\S^{d-1} \rightarrow \S^{t-1}$ defined by $g(x) = \frac{f(x)}{\|f(x)\|}$. Note that $g$ is well defined as $f(x)$ is nonzero for every $x \in \S^{d-1}$. Consider also the function $\widetilde{g} : \S^{d-1} \rightarrow \R^{t-1}$ that maps every $x \in \S^{d-1}$ to the projection of $g(x)$ to its last $t-1$ coordinates. We claim that there is no $x \in \S^{d-1}$ such that $\widetilde{g}(x)=\widetilde{g}(-x)$. To see this, notice that $g(x)$ is a unit vector whose first entry is non-negative, so the projection of $g(x)$ to its last $t-1$ coordinates fully determines $g(x)$. This implies that if there exists an $x \in \S^{d-1}$ satisfying $\widetilde{g}(x)=\widetilde{g}(-x)$ then this $x$ also satisfies $g(x)=g(-x)$, in contradiction to the orthogonality of $f(x)$ and $f(-x)$. Since $\widetilde{g}$ is continuous we can apply Theorem \[thm:BU\] to derive that $t-1 > d-1$ which implies that $t \geq d+1$ and completes the proof.
\[cor:BorsukGraph\] For every $0< \alpha <2$ and an integer $d$, $\xi_\R(\overline{B(d,\alpha)}) \geq d+1$.
For $0< \alpha <2$ and an integer $d$, let $V$ be a maximal collection of points in $\S^{d-1}$ with pairwise distances at least ${{\varepsilon}}= 1-\frac{\alpha}{2}$. Observe that $V$ is finite and that for every $x \in \S^{d-1}$ there exists $y \in V$ for which $\|x-y\| < {{\varepsilon}}$. The graph $G$ associated with $V$ and ${{\varepsilon}}$ in Theorem \[thm:BorsukGraph\] is a subgraph of $B(d,\alpha)$, hence $\xi_\R(\overline{B(d,\alpha)}) \geq \xi_\R(\overline{G}) \geq d+1$, and we are done.
Fractional Upper Bounds on the Shannon Capacity {#sec:Shannon4}
===============================================
Let $f$ be a real-valued non-negative function on graphs. As explained in Section \[sec:Shannon\], if $f$ is a sub-multiplicative upper bound on the independence number then it forms an upper bound on the Shannon capacity. A fractional variant $f^*$ of $f$ was introduced in [@HuTS17], and it was shown there that if $f$ is a sub-multiplicative upper bound on the independence number then so is $f^*$, hence $f^*$ also forms an upper bound on the Shannon capacity. For a graph $G$ on the vertex $V$, the definition of $f^*(G)$ is given in , and by duality it is equal to the value of the following linear program. $$\begin{aligned}
\label{Dual}
\begin{split}
\textup{minimize } &\sum_{S \subseteq V} q(S) \cdot f(G[S]) \\
\textup{subject to } &\sum_{S: x \in S}q(S) \ge 1 \textup{~~~ for each vertex $x \in V$},\\
&q(S)\ge 0 \textup{~~~ for each set $S \subseteq V$}.
\end{split}\end{aligned}$$ Note that every graph $G$ satisfies $f^*(G) \leq f(G)$, as follows by taking $q(S)=1$ for $S=V$ and $q(S)=0$ otherwise. We study here the integrality gap of fractional upper bounds $f^*$ on the Shannon capacity, namely, the largest possible ratio $f(G)/f^*(G)$ over all $n$-vertex graphs $G$.
Upper Bound
-----------
We prove now Theorem \[thm:fractional\_upper\], which claims that for sub-additive functions $f$ on graphs, the integrality gap of $f^*$ is at most logarithmic in the number of vertices. Recall that $f$ is sub-additive if for every graph $G$ on the vertex set $V$ and every sets $S_1$ and $S_2$ such that $V = S_1 \cup S_2$, it satisfies $f(G) \leq f(G[S_1]) + f(G[S_2])$.
We use the dual definition of $f^*$ given in . Let $G=(V,E)$ be an $n$-vertex graph, and let $q$ be an optimal solution of . Denote $Q = \sum_{S \subseteq V}{q(S)} \geq 1$, and let $D$ be the distribution over the subsets of $V$ that assigns to every set $S \subseteq V$ the probability $\frac{q(S)}{Q}$. For, say, $t = \lceil Q \cdot \ln (3n) \rceil$, let $S_1,\ldots,S_t$ be $t$ random subsets of $V$ chosen independently from the distribution $D$.
We first claim that the probability that the sets $S_1,\ldots,S_t$ do not cover the entire vertex set $V$ is at most $1/3$. Indeed, for every $x \in V$ and $i \in [t]$ the probability that $x \in S_i$ is $$\sum_{S:x \in S}{\frac{q(S)}{Q}} = \frac{1}{Q} \cdot \sum_{S:x \in S}{q(S)} \geq \frac{1}{Q},$$ where the inequality holds since $q$ is a feasible solution of . Hence, the probability that $x$ does not belong to any of the sets $S_1,\ldots,S_t$ is at most $$\Big (1-\frac{1}{Q} \Big )^t \leq e^{-t/Q} \leq \frac{1}{3n}.$$ By the union bound, the probability that there exists $x \in V$ such that $x \notin \cup_{i=1}^{t}{S_i}$ is at most $1/3$.
We next consider the expectation of the sum $\sum_{i=1}^{t}{f(G[S_i])}$. For every $i \in [t]$ we have $${\mathop{{\mathbb{E}}}}[f(G[S_i])] = \sum_{S \subseteq V} {\frac{q(S)}{Q} \cdot f(G[S])} = \frac{f^*(G)}{Q},$$ where the second equality holds since $q$ is an optimal solution of . By linearity of expectation, $${\mathop{{\mathbb{E}}}}\Big [\sum_{i=1}^{t}{f(G[S_i])} \Big ] = t \cdot \frac{f^*(G)}{Q} = \lceil Q \cdot \ln (3n) \rceil \cdot \frac{f^*(G)}{Q} \leq 2 \cdot \ln (3n) \cdot f^*(G).$$ By Markov’s inequality, the probability that $\sum_{i=1}^{t}{f(G[S_i])} \geq 6 \cdot \ln (3n) \cdot f^*(G)$ is at most $1/3$.
Finally, by the union bound, there are sets $S_1,\ldots,S_t$ that cover the vertex set $V$ and satisfy $\sum_{i=1}^{t}{f(G[S_i])} \leq 6 \cdot \ln (3n) \cdot f^*(G)$. For these sets the sub-additivity of $f$ implies that $$f(G) \leq \sum_{i=1}^{t}{f(G[S_i])} \leq 6 \cdot \ln (3n) \cdot f^*(G),$$ and we are done.
Lower Bounds
------------
We turn to show that the ratio between $f$ and $f^*$ is unbounded for several upper bounds $f$ on the Shannon capacity. In fact, we present an explicit family of graphs for which $f^*$ is bounded from above by a constant whereas $f$ is arbitrarily large. We consider here the graph parameters $\xi_{\mathbb{F}}$ and ${{\mathop{\mathrm{minrk}}}}_{\mathbb{F}}$, which, as mentioned before, are sub-additive and sub-multiplicative upper bounds on the independence number for every field ${\mathbb{F}}$.
We need the well-studied notion of fractional chromatic number of graphs. For a graph $G$ on the vertex set $V$, let $\calI(G)$ denote the collection of all independent sets of $G$. The [*fractional chromatic number*]{} of $G$, denoted by $\chi_f(G)$, is defined as the value of the following linear program. $$\begin{aligned}
\label{fractional_chi}
\begin{split}
\textup{minimize } &\sum_{I \in \calI(G)} q(I) \\
\textup{subject to } &\sum_{I \in \calI(G): x \in I}q(I) \ge 1 \textup{~~~ for each vertex $x \in V$},\\
&q(I)\ge 0 \textup{~~~ for each set $I \in \calI(G)$}.
\end{split}\end{aligned}$$ The following claim, given in [@HuTS17], follows directly from and .
\[claim:f\*chi\_f\] Let $f$ be a function on graphs satisfying $f(G)=1$ whenever $G$ is complete. Then for every graph $G$, $f^*(G) \leq \chi_f(\overline{G})$. In particular, for every graph $G$ and every field ${\mathbb{F}}$, $${{\mathop{\mathrm{minrk}}}}_{\mathbb{F}}^*(G) \leq \xi_{\mathbb{F}}^*(G) \leq \chi_f(\overline{G}).$$
For integers $d \geq 2s$, consider the Kneser graph $K(d,s)$ (see Section \[sec:Kneser\]). The number of vertices in $K(d,s)$ is $\binom{d}{s}$. By the Erd[ő]{}s-Ko-Rado theorem [@ErdosKoRado61] its independence number is $\alpha(K(d,s)) = \binom{d-1}{s-1}$, and as already mentioned, its chromatic number is $\chi(K(d,s)) = d-2s+2$ [@LovaszKneser]. It is known that every vertex-transitive graph $G=(V,E)$ (that is, a graph whose automorphism group is transitive), satisfies $\chi_f(G) = \frac{|V|}{\alpha(G)}$ (see, e.g., [@GraphTheoryBook]). Hence, $$\begin{aligned}
\label{eq:fracKneser}
\chi_f(K(d,s)) = \frac{\binom{d}{s}}{\binom{d-1}{s-1}} = \frac{d}{s}.\end{aligned}$$
Now we are ready to derive the following theorem, which confirms Theorem \[thm:IntroIntegrality\].
\[thm:AllBounds\] For every fixed ${{\varepsilon}}>0$, there exists an explicit family of $n$-vertex graphs $G$ such that
1. \[itm:1\] ${{\mathop{\mathrm{minrk}}}}^*_{\mathbb{F}}(G) \leq \xi^*_{\mathbb{F}}(G) \leq 2+{{\varepsilon}}$ for every field ${\mathbb{F}}$,
2. \[itm:2\] $\xi_\R (G)= \Theta(\log n)$,
3. \[itm:3\] $\xi_\C (G)= \Theta(\log n)$,
4. \[itm:4\] ${{\mathop{\mathrm{minrk}}}}_\R (G) \geq \Omega( \sqrt{ \log n})$, and
5. \[itm:5\] $\xi_{\mathbb{F}}(G) \geq {{\mathop{\mathrm{minrk}}}}_{\mathbb{F}}(G) \geq \Omega( \log \log n)$ for every fixed finite field ${\mathbb{F}}$.
For a fixed ${{\varepsilon}}>0$, let $d$ and $s$ be two integers satisfying $d = (2+{{\varepsilon}}) \cdot s$. Let $G$ be the complement of the Kneser graph $K(d,s)$. We show that $G$ satisfies the assertion of the theorem. By , we have $$\chi_f(\overline{G}) = \chi_f(K(d,s)) = \frac{d}{s} = 2+{{\varepsilon}},$$ hence Item \[itm:1\] of the theorem follows from Claim \[claim:f\*chi\_f\]. The graph $G$ has $n = \binom{d}{s} = 2^{\Theta(s)}$ vertices, so we have $d-2s+2 = {{\varepsilon}}\cdot s +2 = \Theta(\log n)$. For the upper bounds in Items \[itm:2\] and \[itm:3\] notice that for every field ${\mathbb{F}}$, $\xi_{\mathbb{F}}(G) \leq \chi(\overline{G}) = d-2s+2 = \Theta(\log n)$. The lower bounds in Items \[itm:2\], \[itm:3\], and \[itm:4\] follow from Corollary \[cor:xiKneser\]. Finally, Item \[itm:5\] follows from Claim \[claim:minrankF\].
Acknowledgments {#acknowledgments .unnumbered}
===============
We would like to thank Itzhak Tamo for useful conversations and for his comments on an early version of the paper. We are also grateful to Florian Frick and Günter M. Ziegler for helpful discussions.
[^1]: School of Computer Science, The Academic College of Tel Aviv-Yaffo, Tel Aviv 61083, Israel.
[^2]: Note that the orthogonality dimension parameter (also known as orthogonality rank) is sometimes defined in the quantum communication complexity literature as the orthogonality dimension of the complement graph, namely, the definition requires vectors associated with [*adjacent*]{} vertices to be orthogonal. In this paper we have decided to follow the definition commonly used in the information theory literature.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'This paper studies the energy efficiency of the cloud radio access network (C-RAN), specifically focusing on two fundamental and different downlink transmission strategies, namely the data-sharing strategy and the compression strategy. In the data-sharing strategy, the backhaul links connecting the central processor (CP) and the base-stations (BSs) are used to carry user messages – each user’s messages are sent to multiple BSs; the BSs locally form the beamforming vectors then cooperatively transmit the messages to the user. In the compression strategy, the user messages are precoded centrally at the CP, which forwards a compressed version of the analog beamformed signals to the BSs for cooperative transmission. This paper compares the energy efficiencies of the two strategies by formulating an optimization problem of minimizing the total network power consumption subject to user target rate constraints, where the total network power includes the BS transmission power, BS activation power, and load-dependent backhaul power. To tackle the discrete and nonconvex nature of the optimization problems, we utilize the techniques of reweighted $\ell_1$ minimization and successive convex approximation to devise provably convergent algorithms. Our main finding is that both the optimized data-sharing and compression strategies in C-RAN achieve much higher energy efficiency as compared to the non-optimized coordinated multi-point transmission, but their comparative effectiveness in energy saving depends on the user target rate. At low user target rate, data-sharing consumes less total power than compression, however, as the user target rate increases, the backhaul power consumption for data-sharing increases significantly leading to better energy efficiency of compression at the high user rate regime.'
author:
- '[^1]'
bibliography:
- 'IEEEabrv.bib'
- 'myref.bib'
title: Energy Efficiency of Downlink Transmission Strategies for Cloud Radio Access Networks
---
Cloud radio access network (C-RAN), data-sharing strategy, compression strategy, energy efficiency, power minimization, base-station activation, base-station clustering, beamforming, backhaul power.
Introduction
============
deployment of small cells and cooperative communications are recognized as two promising technologies to meet the ever increasing demand of data traffic for future wireless networks [@Rost14]. However, both technologies come at the cost of increase in energy consumption because of the additional energy needed to support the increasing number of base-station (BS) sites and the substantially increased backhaul between the BSs for cooperation. The excessive energy consumption of wireless networks not only has an ecological impact in terms of carbon footprint but also has an economical impact on the operational cost to the mobile operators. Thus, the compelling call for improvement of *spectrum efficiency* in the fifth-generation (5G) wireless network needs to be accompanied by a call for improvement of *energy efficiency* to the same extent.
Cloud radio access network (C-RAN) is an emerging network architecture that shows significant promises in improving both the spectrum efficiency and the energy efficiency of current wireless networks [@CRAN]. In C-RAN, the BSs are connected to a central processor (CP) through backhaul links. The benefits of the C-RAN architecture in energy saving are several-fold. First, under the C-RAN architecture, most of the baseband signal processing in traditional BSs can be migrated to the cloud computing center so that the conventional high-cost high-power BSs can be replaced by low-cost low-power radio remote heads (RRHs). Second, the existence of CP also allows for the joint precoding of user messages for interference mitigation. With less interference generated, the transmit power at the BSs can therefore be reduced. Third, as on average (and especially during non-peak time) a significant portion of network resources can be idle [@Correia10], the CP can perform joint resource allocation among the BSs to allocate resources on demand and put idle BSs into sleep mode for energy saving [@Frenger11].
The above-mentioned benefits of C-RAN in energy saving are concerned with the BS side. However, the additional energy consumption due to the increased backhaul between the CP and the BSs also needs to be taken into account [@Tombaz11]. In this paper, we investigate the potential of C-RAN in improving energy efficiency of the communication aspect of the network by considering the energy consumption due to BS activation, transmission, and backhaul provisioning. The backhaul energy consumption depends on the backhaul rate, which further depends on the interface between the CP and the BSs. In this paper, we investigate two fundamental and different transmission strategies for the downlink C-RAN. In the *data-sharing strategy*, the CP uses the backhaul links to share user messages to a cluster of cooperating BSs. The backhaul cost of the data-sharing strategy depends on the number of BSs that the user messages need to be delivered to: larger cluster size leads to larger cooperative gain, but also higher backhaul rate. In an alternative strategy called the *compression strategy*, the CP performs joint precoding of the user messages centrally then forwards a compressed version of the precoded signals to the BSs. The backhaul cost of the compression strategy depends on the resolution of the compressed signals: higher-resolution leads to better beamformers, but also larger backhaul rate.
This paper aims to quantify the energy saving of C-RAN while accounting for both the BS and the backhaul energy consumptions, and specifically to answer the question of between the data-sharing strategy and the compression strategy, which one is more energy efficient? The answer to this question is nontrivial as there are three factors that can lead to energy reduction: decrease in BS transmit power, turning-off of the BSs, and reduction in backhaul rate. These three factors are interrelated. For example, it may be beneficial to keep more BSs active and to allocate higher backhaul rate in order to allow for better cooperation among the BSs so that more interference can be mitigated. This leads to less transmit power consumption, but it can also lead to higher BS and backhaul power consumption. This paper intends to capture such interplay using an optimization framework. Specifically, we propose a joint design of the BS transmit power, BS activation and backhaul by minimizing network-wide power under given user rate constraints for both the data-sharing strategy and the compression strategy. The resulting optimization problems are nonconvex in nature and are highly nontrivial to solve globally. This paper approximates the problems using reweighted $\ell_1$ minimization technique [@Candes08] and successive convex approximation technique, and devises efficient algorithms with convergence guarantee. We identify operating regimes where one strategy is superior to the other, and show that overall optimized C-RAN transmission can lead to more energy efficient network operation than the non-optimized coordinated multi-point (CoMP) transmission.
Related Work
------------
The potential of C-RAN in improving the performance for future wireless networks has attracted considerable attentions recently. In the uplink, [@Lei13] and [@Yuhan14] show that with the capability of jointly decoding user messages in the CP, the throughput of traditional cellular networks can be significantly improved. A similar conclusion also has been drawn in the downlink. In particular, [@Park13] proposes a joint design of beamforming and multivariate compression to maximize the weighted sum rate for the compression strategy, while [@hong12] and [@BinbinSparseBFJnal] consider joint beamforming and BS clustering design to maximize the weighted sum rate for data-sharing.
This paper focuses on the energy efficiency of C-RAN in the downlink. Several metrics have been proposed in the literature to measure the energy efficiency of a network. For example, the area power consumption metric (watts$/$unit area) is proposed in [@Richter09] to evaluate the energy efficiency of networks of different cell site densities. Another widely adopted measurement in energy efficiency is bits per joule metric, which has been studied in [@Sarkiss12] under a simplified single-user-two-BSs model from an information-theoretical point of view and also studied in [@Schober12; @Peng14; @Huq14] for orthogonal frequency division multiple access (OFDMA) based cooperative networks from the practical system design point of view.
In this paper, we formulate the problem of minimizing the total required power for downlink C-RAN in order to provide a given set of quality-of-service (QoS) targets for the scheduled users. Such an optimization problem can also be thought of as minimizing watts per bit (or maximizing the bits per watt) at given user service rates. In this domain, most of the previous works are restricted to the data-sharing strategy [@Shi13; @Rui14; @Han14]. Specifically, [@Shi13] proposes a joint BS selection and beamforming design algorithm to minimize the total power consumption in the downlink, while [@Rui14] generalizes to a joint downlink and uplink total power minimization problem. Both [@Shi13] and [@Rui14] take advantage of the fact that if a BS is not selected to serve any user at the current time slot, it can be put into low-power sleep mode for energy saving purpose. In contrast, [@Han14] assumes fixed BS association but exploits the delay tolerance of the users to improve the energy efficiency in CoMP transmission. In delay-tolerant applications, BSs can aggregate the user messages and transmit them with high rate during a short time frame while remain idle for the rest of the time slots under power-saving sleep mode. Fast deactivation/activation of hardware power-consuming components achieve significant energy reductions [@Frenger11; @Kimmo13].
In this paper, we adopt a similar energy saving perspective as in [@Shi13] but consider in addition the compression strategy. Compression strategy differs from data-sharing strategy in the way that the backhaul is utilized. We adopt the model proposed in [@Fehske10] to model the power consumption of backhaul links as a linear function of backhaul rates. This is in contrast to [@Shi13], where the backhaul power is modeled as a step function with only two levels of power consumption depending on whether the backhaul link is active or not.
In addition to the backhaul power, we also consider the BS power consumption by adopting the model proposed in [@Auer11], which approximates the power consumption of a BS as a piecewise linear function of transmit power. In such model, BS sleep mode corresponds to a constant but lower power consumption with zero transmit power. BS active mode corresponds to a higher constant power plus a nonzero transmit power. The overall framework of this paper is a joint optimization of BS transmit power, BS activation and backhaul rate for both the compression and the data-sharing strategies.
From the optimization perspective, the total power consumption for the data-sharing strategy involves a sum of weighted nonconvex $\ell_0$-norms, which is highly nontrivial to optimize globally. Instead, we adopt the reweighted $\ell_1$ technique [@Candes08] to approximate the nonconvex total power into a convex weighted sum of transmit power, where the weights are iteratively updated in a way to reduce not only the number of active BSs but also the backhaul rate. Such technique has also been applied to minimize the total backhaul rate in [@Zhao12], and to optimize the tradeoff between the total transmit power and total backhaul rate in [@binbin13]. On a related note, the discrete $\ell_0$-norm can also be approximated using other tractable continuous functions such as Gaussian-like function in [@Zhuang14] and exponential function in [@Vu14]. It has been reported recently in [@ZhouTaoChen] that those approximation methods show similar effectiveness in inducing sparsity.
Further, the mathematical expression of the total power consumption for the compression strategy involves a difference of two logarithmic functions, which is also nonconvex. We propose to approximate the first logarithmic function using the successive convex approximation technique, which transforms the objective function into a convex form. The adopted reweighted $\ell_1$ minimization technique and successive convex approximation technique in this paper are related to the majorization-minimization (MM) algorithm [@Bharath11], which deals with an optimization problem with nonconvex objective function by successively solving a sequence of optimization problems with approximate objective functions. This paper utilizes the known sufficient conditions of convergence for the MM algorithm in the literature [@Meisam13] to show the convergence of the proposed algorithms for both the data-sharing and the compression strategies.
Finally, we mention that the data-sharing and compression strategies considered in this paper are not the only possibilities for the downlink of C-RAN. There is a potential to combine these two strategies by sending directly the messages of only the strong users to the BSs and compressing the rest [@Patil14]. Also, reverse compute-and-forward strategy that accounts for the lattice nature of the transmitted message is also possible [@HongCaire13]. However, such strategy is difficult to optimize because of the need in choosing the right integer zero-forcing precoding coefficients at the CP so that the effective noise, caused by the non-integer penalty due to practical channels, at each user is minimized.
Main Contributions
------------------
This paper considers energy-efficient design of the data-sharing strategy and the compression strategy for downlink C-RAN by formulating a problem of minimizing the total network power consumption subject to user rate constraints. The first contribution of this paper is the modeling of both the BS power and the backhaul power consumption in the network. The BS power consumption model includes a low-power sleep mode, while the backhaul power consumption is modeled as a linear function of backhaul traffic rate.
For the data-sharing strategy, we propose a novel application of reweighted $\ell_1$ minimization technique to approximate the nonconvex BS activation power and backhaul power. Such approximation technique reduces the nonconvex optimization problem to a conventional convex transmit power minimization problem, which can be solved efficiently using the uplink-downlink duality approach or through transformation as second-order cone programming (SOCP). Moreover, we adopt a reweighting function that enables us to connect the reweighted $\ell_1$ minimization technique with the MM algorithm. This connection allows us to prove the convergence behavior of the proposed algorithm for the data-sharing strategy.
For the compression strategy, in addition to the reweighted $\ell_1$ approximation to the BS activation power as in the data-sharing strategy, we propose a successive convex approximation to the backhaul power, which is in a nonconvex form as a difference of two logarithmic functions. The proposed successive convex approximation technique and the reweighted $\ell_1$ approximation technique can be combined together. The combined algorithm falls into the class of the MM algorithms and has convergence guarantee.
Through simulations, we show that optimized data-sharing and compression strategies in C-RAN can bring much improved energy efficiency as compared to the non-optimized CoMP transmission. However, the comparative energy saving of data-sharing versus compression depends on the user target rates. The energy efficiency of the data-sharing strategy is superior to that of the compression strategy in the low-rate regime. However, the backhaul power consumption of the data-sharing strategy increases significantly with the user rate. Thus, in high user rate regime, the compression strategy may be preferred from an energy saving perspective.
Paper Organization and Notations
--------------------------------
The remainder of this paper is organized as follows. Section \[sec:SystemModel\] introduces the system model and power consumption model considered throughout this paper. Section \[sec:data\_sharing\] considers the total power minimization under the data-sharing transmission strategy, while Section \[sec:compression\] considers the compression strategy. Simulation results are presented in Section \[sec:simulations\] and conclusions are drawn in Section \[sec:conclusion\].
Throughout this paper, lower-case letters (e.g. $x$) and lower-case bold letters (e.g. $\mathbf{x}$) denote scalars and column vectors respectively. We use $\mathbb{C}$ to represent complex domain. The transpose, conjugate transpose and $\ell_p$-norm of a vector are denoted as $(\cdot)^{T}$, $(\cdot)^H$ and $\Vert\cdot\Vert_p$ respectively. The expectation of a random variable is denoted as $\mathsf{E} \left [ \cdot \right]$. Calligraphy letters are used to denote sets, while $|\cdot|$ stands for either the size of a set or the absolute value of a scalar, depending on the context.
System and Power Consumption Model {#sec:SystemModel}
==================================
In this section, we describe the overall system model and power consumption model for the downlink C-RAN considered throughout this paper.
System Model
------------
Consider a downlink C-RAN with $L$ BSs serving $K$ users. All the BSs are connected to a CP via backhaul[^2] links and each user receives a single independent data stream from the BSs. All the user messages are assumed to be available at the CP and are jointly processed before being forwarded to the BSs through the backhaul links. We assume that the CP has access to global channel state information (CSI) but point out that such assumption can be relaxed so that only the CSI from the neighboring BSs of each user is needed in the CP.
To simplify notations and ease analysis, we assume that the BSs and the users are equipped with a single antenna each, although the proposed algorithms in this paper can be easily generalized to the case of multi-antenna BSs as discussed later in the paper. Let $x_l \in \mathbb{C}$ denote the transmit signal at BS $l$, we can write the received signal $y_k \in \mathbb{C}$ at user $k$ as $$\label{eq:yk_general}
y_k = \mathbf{h}_k^{H} \mathbf{x} + n_k, \quad k \in \mathcal{K} = \left\{1, 2, \cdots, K\right\}$$ where $\mathbf{x} = \left[x_1, x_2, \cdots, x_L\right]^{T}$ is the vector of transmit signals across all the $L$ BSs and $\mathbf{h}_k \in \mathbb{C}^{L \times 1}$ is the vector of channel gains from all the $L$ BSs to user $k$. The received noise $n_k$ is modeled as a complex Gaussian random variable with zero mean and variance $\sigma^2$. Each user decodes its own message $s_k \in \mathbb{C}$ from the received signal $y_k$.
In this paper, we investigate two fundamental but different transmission strategies, the data-sharing strategy and the compression strategy, for the downlink C-RAN for delivering the message $s_k$ to user $k$ via the transmit signal $\mathbf{x}$ from the BSs. In particular, we compare the potential of these two strategies in improving the energy efficiency. Before discussing the details of the two strategies, we first describe the power consumption model adopted in this paper.
Power Consumption Model
-----------------------
Traditional cellular network transmission strategy design typically only considers transmit power at each BS, which is written as $$\label{eq:TxPower}
P_{l, tx} = \mathsf{E} \left[ \vert x_l \vert^{2} \right] \leq P_l, \quad l \in \mathcal{L} = \left\{1, 2, \cdots, L\right\},$$ where $P_l$ is the transmit power budget available at BS $l$. However, a full characterization of power consumption at a BS should also consider the efficiency of the power amplifier and other power-consuming components such as baseband unit, cooling system, etc. In addition, the power consumption of backhaul links connecting the BSs to the CP also needs to be taken into account for the specific C-RAN architecture considered in this paper. In the following, we describe the power consumption model adopted in this paper for the BSs and the backhaul links respectively.
### Base-Station Power Consumption
The characteristic of power-consuming components in a BS depends on the BS design. We adopt the following unified power consumption model proposed in [@Auer11], which is applicable for different types of BSs. This model approximates the BS power consumption as a piecewise linear function of the transmit power $P_{l, tx}$: $$\label{eq:bs_power}
P_{l}^{BS} = \left \{
\begin{array}{ll}
\eta_l P_{l, tx} + P_{l, active}, & \text{if} ~ 0 < P_{l, tx} \leq P_l \\
P_{l, sleep}, & \text{if} ~ P_{l, tx} = 0
\end{array}
\right. \hspace{-3mm}, ~ l \in \mathcal{L}$$ where $\eta_l > 0$ is a constant reflecting the power amplifier efficiency, feeder loss and other loss factors due to power supply and cooling for BS $l$, $P_{l, tx}$ is the transmit power defined in and $P_{l, active}$ is the minimum power required to support BS $l$ with non-zero transmit power. If BS $l$ has nothing to transmit, it can be put into sleep mode with low power consumption $P_{l, sleep}$. Typically, $P_{l, sleep} < P_{l, active}$ so that it is beneficial to turn BSs into sleep mode, whenever possible, for energy-saving purpose.
### Backhaul Power Consumption
In C-RAN, the BSs are connected to the CP with the backhaul links. The power consumption due to backhaul links varies with different backhaul technologies. In this paper, we model the backhaul as a set of communication channels, each with capacity $C_l$ and power dissipation $P_{l,max}^{BH}$, and write the backhaul power consumption as $$\label{eq:bkhaul_power}
P_l^{BH} = \frac{R_l^{BH}}{C_l} P_{l,max}^{BH} = \rho_l R_l^{BH}, \quad l \in \mathcal{L}$$ where $\rho_l = P_{l,max}^{BH} / C_l$ is a constant scaling factor and $R_l^{BH}$ is the backhaul traffic between BS $l$ and the CP. This model has been used in [@Fehske10] for microwave backhaul links and can also be generalized to other backhaul technologies, such as passive optical network, fiber-based Ethernet, etc., as mentioned in [@wu2012green]. Note that [@Shi13] also considers the sleep mode capability for backhaul links. We point out that such consideration can be unified with $P_{l, active}$ and $P_{l, sleep}$ in the BS power consumption model .
### Total Power Consumption
Based on the above BS power consumption model and backhaul power consumption model , we can write the total power consumption $P_{total}$ for C-RAN as $$\begin{aligned}
\label{eq:total_power}
P_{total}
& = \sum_{l \in \mathcal{L}} \left( P_{l}^{BS} + P_l^{BH} \right)\nonumber \\
& = \sum_{l \in \mathcal{L}} \bigg( \eta_l P_{l, tx} + \mathbbm{1}_{\left\{ P_{l, tx} \right\}} \left( P_{l, active} - P_{l, sleep} \right) \nonumber \\
& \hspace{4.5cm} + P_{l, sleep} + \rho_l R_l^{BH} \bigg) \nonumber \\
& = \sum_{l \in \mathcal{L}} \bigg( \eta_l P_{l, tx} + \mathbbm{1}_{\left\{ P_{l, tx} \right\}} P_{l, \Delta} + \rho_l R_l^{BH} \bigg) \nonumber \\
& \hspace{4.5cm} + \underbrace{\sum_{l \in \mathcal{L}} P_{l, sleep}}_{\text{constant}}\end{aligned}$$ where $\mathbbm{1}_{\left\{ \cdot \right\}}$ is the indicator function defined as $$\label{eq:indicator}
\mathbbm{1}_{\left\{x\right\}} = \left \{
\begin{array}{l}
1, \quad \text{if} ~ x > 0 \\
0, \quad \text{otherwise}
\end{array}
\right. ,$$ and $P_{l, \Delta} = P_{l, active} - P_{l, sleep}$ is the difference between the minimum active BS power consumption and the sleep mode BS power consumption.
As we can see from , there are three possibilities in improving the energy efficiency of C-RAN: reducing the transmit power, putting BSs into sleep mode, and decreasing the backhaul traffic. However, these three aspects cannot be realized simultaneously: deactivating more BSs means reduced capability for interference mitigation among the active BSs, which leads to higher transmit power in order to maintain the QoS for the users; higher backhaul rate can allow for more user information being shared among the BSs so that the BSs can better cooperate to mitigate interference, thus less transmit power may be needed. A joint design is necessary in order to balance the roles of transmit power, BS activation and backhaul traffic rate in achieving energy efficiency. In the following, we describe the general problem formulation considered in this paper for such joint design used in both the data-sharing and the compression strategies.
Energy Efficiency Maximization
------------------------------
This paper aims to understand the energy efficiency for downlink C-RAN, which can be defined as the ratio of the achievable sum rate and the sum power consumption, i.e. $\frac{\sum_{k} R_k}{P_{total}}$ where $R_k$ is the data rate for user $k$ determined by the specific transmission strategy and $P_{total}$ is the total consumed power defined in . Towards this end, this paper takes the similar approach as in [@Han14] to fix the service rates of scheduled users and consider the minimization of total power consumption: $$\begin{aligned}
\label{prob:power_min}
\operatorname*{minimize}& \quad P_{total} \\
\operatorname{subject \text{ } to}& \quad R_k \geq r_k, \quad \forall k \in \mathcal{K} \nonumber \\
& \quad \mathsf{E} \left[ \vert x_l \vert^{2} \right] \leq P_l, \quad \forall l \in \mathcal{L} \nonumber \end{aligned}$$ where $r_k$ is the fixed target rate for user $k$. The solution to the above problem gives us the energy efficiency $\frac{\sum_{k} R_k}{P_{total}}$ of the system at the operating point $\left(r_1, r_2, \cdots, r_K \right)$. To maximize energy efficiency, we need to further search over all operating points. For the rest of the paper, we study and compare the minimum required total power for different transmission strategies under the same operating point $\left(r_1, r_2, \cdots, r_K \right)$ in the downlink of C-RAN. Note that problem implicitly assumes fixed user scheduling. There also exists a possibility of doing joint user scheduling and power minimization by considering a problem of minimizing the total power consumption across *multiple time slots* subject to a minimum target for each user’s *average* rate. Such problem is considerably more complicated.
Data-Sharing versus Compression
-------------------------------
Data-sharing and compression are two fundamentally different transmission strategies for the downlink of C-RAN for delivering data to the users. These two strategies correspond to alternative functional splits in C-RAN. In the data-sharing transmission strategy, the CP routes each scheduled user’s intended message to a cluster of BSs through the backhaul links; the cluster of BSs then cooperatively serve that user through joint beamforming. In contrast, in the compression strategy, the precoding operation is implemented centrally at the CP, which then forwards a compressed version of the analog beamformed signal to the BSs through the backhaul/fronthaul links. The BSs then simply transmit the compressed beamforming signals to the users [@Park13; @Patil14; @PratikEUSIPCO].
The data-sharing strategy differs from the compression strategy in backhaul utilization. In data-sharing, the backhaul rate is a function of the user message rate and the BS cluster size, while in the compression strategy the backhaul cost is determined by the compression resolution. Intuitively, as the user target rate and BS cluster size increase, the backhaul rate for the data-sharing strategy would increase significantly, leading to high energy consumption. However, in the low user rate regime where the BS cluster size is small, data-sharing can be more efficient than compression as the latter suffers from quantization noise. Therefore, there exists a tradeoff between data-sharing and compression in terms of backhaul rate and energy efficiency at different user target rate operating points. In the following two sections, we describe in details the data-sharing strategy and the compression strategy, and propose corresponding algorithms to find the minimum required total power for each strategy.
Throughout this paper, we primarily account for the energy consumption due to *communications* in either the backhaul or the transmission front-end at the BSs, rather than the energy consumption due to *computing*. There is significant additional energy saving due to migrating signal processing from the BSs to the cloud computer center in the C-RAN architecture. We refer the readers to [@Chen14].
Data-Sharing Strategy {#sec:data_sharing}
=====================
\[tc\]\[Bc\]\[1\][Central Processor]{} \[tc\]\[Bc\]\[1\][$s_1, s_2$]{} \[tc\]\[cc\]\[1\][$s_1$]{} \[tc\]\[cc\]\[1\][$s_2$]{} \[tl\]\[tl\]\[0.7\][$R_1^{BH} = R_1$]{} \[tl\]\[tl\]\[0.7\][$R_2^{BH} = $]{} \[tl\]\[tl\]\[0.7\][$R_1 + R_2$]{} \[tl\]\[tl\]\[0.7\][$R_3^{BH} = R_2$]{} \[tc\]\[cl\]\[0.8\][$x_1 = w_{11}s_1$]{} \[tc\]\[cl\]\[0.8\][$x_2 = w_{21}s_1$]{} \[tc\]\[tl\]\[0.8\][$+ w_{22}s_2$]{} \[tc\]\[cl\]\[0.8\][$x_3 = w_{32}s_2$]{} \[tc\]\[tc\]\[0.9\][BS $1$]{} \[tc\]\[tc\]\[0.9\][BS $2$]{} \[tc\]\[tc\]\[0.9\][BS $3$]{} \[tc\]\[tc\]\[0.9\][User $1$]{} \[tc\]\[tc\]\[0.9\][User $2$]{}
In this section, we study the minimum total power required for the data-sharing strategy in order to support the given scheduled users at guaranteed service rates.
Problem Formulation
-------------------
Consider the data-sharing transmission strategy for the downlink of C-RAN as illustrated in Fig. \[fig:DataSharingModel\], where the each user’s message is shared among a cluster of serving BSs. Let $w_{lk} \in \mathbb{C}$ be the beamforming coefficient for BS $l$ to serve user $k$. If BS $l$ is not part of user $k$’s serving cluster, $w_{lk}$ is set to be zero. The transmit signal $x_l$ at BS $l$ can be written as $x_l = \sum_{k \in \mathcal{K}} w_{lk} s_k$. We model the user messages $s_k$’s as independent and identically distributed complex Gaussian random variables with zero mean and unit variance. The transmit power $P_{l, tx}$ formulated in can be written as $$\label{eq:Tx_power_data_sharing}
P_{l, tx} = \sum_{k \in \mathcal{K}} \left\vert w_{lk} \right\vert^{2}, \quad l \in \mathcal{L}.$$
Substituting $x_l = \sum_{k \in \mathcal{K}} w_{lk} s_k$ into , the received signal $y_k$ at user $k$ is $$\label{eq:y_k_data_sharing}
y_k = \mathbf{h}_k^{H} \mathbf{w}_k s_k + \sum_{j \neq k}\mathbf{h}_k^{H} \mathbf{w}_j s_j + n_k, \quad k \in \mathcal{K},$$ where $\mathbf{w}_k = \left[w_{1k}, w_{2k}, \cdots, w_{Lk} \right]^{T}$ is the network beamformer for user $k$. Based on , the received signal-to-interference-plus-noise ratio (SINR) at user $k$ can be expressed as $$\text{SINR}_k =
\frac{\left\vert\mathbf{h}_k^{H}\mathbf{w}_k \right\vert^2}{\sum_{j \neq
k}\left\vert\mathbf{h}_k^{H}\mathbf{w}_j\right\vert^2 + \sigma^2}, \quad k \in \mathcal{K}$$ and the achievable rate for user $k$ is then $$\label{eq:R_k_data_sharing}
R_k = \log_2\left(1+\frac{\text{SINR}_k}{\Gamma_m}\right), \quad k \in \mathcal{K},$$ where $\Gamma_m$ stands for the signal-to-noise ratio (SNR) gap due to practical modulation scheme.
For the data-sharing strategy, if user $k$ is served by BS $l$, then the CP needs to send user $k$’s message $s_k$, along with the beamforming coefficient $w_{lk}$, to BS $l$ through the backhaul link. In this paper, we assume that the channels are slow varying and ignore the backhaul required for sharing CSI and beamformers, and only consider the backhaul capacity consumption due to data-sharing. Hence, the backhaul rate for BS $l$, $R_{l}^{BH}$, is the accumulated data rates of those users served by BS $l$, which can be formulated as $$\label{eq:bkhaul_data_sharing}
R_{l}^{BH} = \sum_{k \in \mathcal{K}} \mathbbm{1}_{\left\{ \left\vert w_{lk} \right\vert^{2} \right\}} R_k, \quad l \in \mathcal{L},$$ where $\mathbbm{1}_{\left\{ \left\vert w_{lk} \right\vert^{2} \right\}}$ is the indicator function defined in and indicates whether or not BS $l$ serves user $k$.
Substituting and into , the total power minimization problem can be formulated for the data-sharing strategy as
\[prob1:data\_sharing\] $$\begin{aligned}
\displaystyle \operatorname*{minimize}_{\left\{w_{lk} \right\}} & \quad \sum_{l\in\mathcal{L}}
\Bigg( \eta_l \sum_{k \in \mathcal{K}} \left\vert w_{lk} \right\vert^{2} +
\mathbbm{1}_{\left\{ \sum_{k \in \mathcal{K}} \left\vert w_{lk} \right\vert^{2} \right\}} P_{l, \Delta} \nonumber \\
& \hspace{2.5cm} + \rho_l \sum_{k\in \mathcal{K}} \mathbbm{1}_{\left\{ \left\vert w_{lk} \right\vert^{2} \right\}} R_k \Bigg)
\label{obj_data_sharing1} \\
\operatorname{subject \text{ } to}& \quad R_k = \log_2\left(1+\frac{\text{SINR}_k}{\Gamma_m}\right) \geq r_k, ~ k \in \mathcal{K} \label{rate_const} \\
&\quad \sum_{k \in \mathcal{K}} \left\vert w_{lk} \right\vert^{2} \leq P_l, \quad l \in \mathcal{L}.\end{aligned}$$
Note that the $\sum_{l \in \mathcal{L}} P_{l, sleep}$ term in is a constant and has been dropped in the objective function . It is easy to see that the minimum rate constraint is met with equality at the optimal point. Hence, problem can be equivalently formulated as
\[prob:data\_sharing\] $$\begin{aligned}
\displaystyle \operatorname*{minimize}_{\left\{w_{lk} \right\}} & \quad \sum_{l\in\mathcal{L}}
\Bigg( \eta_l \sum_{k \in \mathcal{K}} \left\vert w_{lk} \right\vert^{2} +
\mathbbm{1}_{\left\{ \sum_{k \in \mathcal{K}} \left\vert w_{lk} \right\vert^{2} \right\}} P_{l, \Delta} \nonumber \\
& \hspace{2.5cm} + \rho_l \sum_{k\in \mathcal{K}} \mathbbm{1}_{\left\{ \left\vert w_{lk} \right\vert^{2} \right\}} r_k \Bigg)
\label{obj_data_sharing} \\
\operatorname{subject \text{ } to}& \quad \text{SINR}_k \geq \gamma_k, \quad k \in \mathcal{K} \label{sinr_const} \\
& \quad \sum_{k \in \mathcal{K}} \left\vert w_{lk} \right\vert^{2} \leq P_l, \quad l \in \mathcal{L} \label{Per_BS_Power}\end{aligned}$$
where the variable $R_k$ in is replaced by the target rate $r_k$ in and $\gamma_k = \Gamma_m \left(2^{r_k} - 1\right)$ in . The new SINR constraint is also met with equality at the optimality. However, we keep as an inequality constraint, so that it can be reformulated as a convex second-order cone (SOC) constraint [@wiesel2006]. Note that problem is equivalent to problem in the sense that they have the same optimal solutions and the same feasibility region.
Note that the above optimization is over the beamforming coefficients and also implicitly over the BS cluster for each user. The overall optimization problem aims to choose the optimal cluster of serving BSs for each scheduled user for minimizing the total power consumption while satisfying the user QoS constraints. Due to the indicator functions in the objective function , problem is nonconvex (discrete), so finding its global optimum solution is challenging. In the following, we propose to approximate the nonconvex indicator function using reweighted convex $\ell_1$-norm and show that with a particular reweighting function the proposed algorithm always converges[^3].
Proposed Algorithm
------------------
We make an observation that the indicator function defined in is equivalent to the $\ell_0$-norm of a scalar. The $\ell_0$-norm of a vector is defined as the number of nonzero entries in the vector, so it reduces to the indicator function in the scalar case. In compressive sensing literature [@Candes08], nonconvex $\ell_0$-norm minimization problem can be approximated as convex reweighted $\ell_1$ minimization problem. We take advantage of this technique and propose to approximate the indicator functions in the objective function as $$\label{wgt_l}
\mathbbm{1}_{\left\{ \sum_{k \in \mathcal{K}} \left\vert w_{lk} \right\vert^{2} \right\}}
= \left\Vert \sum_{k \in \mathcal{K}} \left\vert w_{lk} \right\vert^{2} \right\Vert_0
\approx \mu_l \sum_{k \in \mathcal{K}} \left\vert w_{lk} \right\vert^{2}$$ $$\label{wgt_lk}
\mathbbm{1}_{\left\{ \left\vert w_{lk} \right\vert^{2} \right\}}
= \left\Vert \left\vert w_{lk} \right\vert^{2} \right\Vert_0
\approx \nu_{lk} \left\vert w_{lk} \right\vert^{2}$$ with weights $\mu_l$ and $\nu_{lk}$ iteratively updated according to $$\begin{aligned}
\label{wgt_update_data}
\mu_l = f\left( \sum_{k \in \mathcal{K}} \left\vert w_{lk} \right\vert^{2}, \tau_1 \right)
& = \frac{c_1}{\sum_{k \in \mathcal{K}} \left\vert w_{lk} \right\vert^{2} + \tau_1} ~ , \nonumber \\
\nu_{lk} = f\left(\left\vert w_{lk} \right\vert^{2}, \tau_2\right)
& = \frac{c_2}{\left\vert w_{lk} \right\vert^{2} + \tau_2} $$ where $\left\{ w_{lk} \right\}$ is the beamformer from the previous iteration, $\tau_1 > 0$ and $\tau_2 > 0$ are some constant regularization factors, and $c_1, c_2$ are constants.
Note that in the above iterative updates of $\mu_l$ and $\nu_{lk}$, the BSs with small transmit power, $\sum_{k \in \mathcal{K}} \left\vert w_{lk} \right\vert^{2}$, or small transmit power to user $k$, $\left\vert w_{lk} \right\vert^{2}$, at current iteration are given larger weights $\mu_l$ or $\nu_{lk}$ in the next iteration. This further decreases $\sum_{k \in \mathcal{K}} \left\vert w_{lk} \right\vert^{2}$ or $\left\vert w_{lk} \right\vert^{2}$ in the next iteration, and eventually forces BS $l$ toward sleep mode (i.e., $\sum_{k \in \mathcal{K}} \left\vert w_{lk} \right\vert^{2} = 0$) or to be removed from user $k$’s serving cluster (i.e., $\left\vert w_{lk} \right\vert^{2} = 0$). The weight $\mu_l$ has the effect of putting appropriate BSs to sleep mode, while $\nu_{lk}$ has the effect of determining the BS cluster size for user $k$, which in turn affects the backhaul capacity consumption of user $k$.
The resulting optimization problem after the $\ell_1$-norm approximation is formulated as follows: $$\begin{aligned}
\label{prob:data_sharing_approx}
\displaystyle \operatorname*{minimize}_{\left\{w_{lk} \right\}} & \quad \sum_{l\in\mathcal{L}} \sum_{k\in \mathcal{K}}
\alpha_{lk} \left\vert w_{lk} \right\vert^{2} \\
\operatorname{subject \text{ } to}& \quad \eqref{sinr_const}, ~~ \eqref{Per_BS_Power} \nonumber\end{aligned}$$ where $\alpha_{lk} = \eta_l + \mu_l P_{l, \Delta} + \rho_l \nu_{lk} r_k$. Problem is a weighted sum transmit power minimization problem, which can be solved efficiently through the uplink-downlink duality approach [@dahrouj10] or by transforming it into an SOCP problem [@wiesel2006]. We now summarize the proposed algorithm to solve the total power minimization problem for the data-sharing strategy in Algorithm \[alg:data\_sharing\].
[**Initialization**]{}: Set the initial values for $\left\{\mu_{l}, \nu_{lk} \right\}$ according to with the $\left\{ w_{lk} \right\}$ chosen as a feasible point of problem ;\
[**Repeat**]{}:
1. Fix $\left\{\mu_{l}, \nu_{lk}\right\}$, find the optimal $\left\{ w_{lk} \right\}$ by solving problem using the uplink-downlink duality approach [@dahrouj10] or by transforming it into an SOCP problem [@wiesel2006];
2. Update $\left\{\mu_{l}, \nu_{lk} \right\}$ according to .
[**Until**]{} convergence
Note that a similar problem as to is considered in our previous work [@binbin13], where we formulate the problem as a tradeoff between the BS transmit power and the backhaul capacity. This paper considers a more realistic BS power consumption model with sleep mode capability, and also accounts for backhaul power consumption. The considered problem in this paper can also be thought of as providing a tradeoff between the per-BS power consumption and the per-BS backhaul capacity consumption, where the tradeoff constant $\rho_l$ is specifically chosen according to the backhaul power consumption model .
Convergence Analysis {#sec:Data_Converge}
--------------------
Algorithm \[alg:data\_sharing\] relies on the reweighting heuristic to deactivate BSs and reduce the BS cluster size for energy saving purpose. To establish the convergence proof for Algorithm \[alg:data\_sharing\] under arbitrary reweighting function is challenging, however, we show in the following that if the reweighting function is chosen as $$\label{eq:reweight}
f\left(x, \tau\right) = \frac{1}{\left(x + \tau\right) \ln \left( 1 + \tau^{-1} \right) } ~,$$ i.e. the constants in are chosen as $c_1 = \frac{1}{\ln \left( 1 + \tau_1^{-1} \right)}, c_2 = \frac{1}{\ln \left( 1 + \tau_2^{-1} \right)}$, Algorithm \[alg:data\_sharing\] can be seen as a special case of the MM algorithms [@Bharath11] and is guaranteed to converge.
\[thm:1\] Starting with any initial point, the sequence $\left\{ w_{lk}^{(n)} \right\}_{n=1}^{\infty}$ generated by Algorithm \[alg:data\_sharing\] with the reweighting function chosen as is guaranteed to converge.
See Appendix \[apdx:a\].
Finally, we point out that the choice of the reweighting function is not unique. There exist other reweighting functions that may work well in different problem setups [@Candes08]. Recently, [@ZhouTaoChen] has experimented with other approximation functions to the $\ell_0$-norm, e.g. exponential function and arc-tangent function, in addition to the logarithmic function used in this paper, and observed similar effectiveness of these functions in inducing sparsity.
Complexity Analysis
-------------------
Algorithm \[alg:data\_sharing\] is an iterative procedure between updating the weights $\left\{\mu_{l}, \nu_{lk}\right\}$ and solving the weighted transmit power minimization problem . The problem can be formulated as an SOCP and solved using the interior-point method, e.g. using the convex optimization solver [@CVX]. The total number of variables in problem is $LK$ and the total number of SOC constraints is $\left(L + K \right)$. The complexity order for solving such a problem through interior-point method is given as $O\left( \left(L + K \right) \left(LK\right)^3 \right)$ [@boyd]. Assuming that Algorithm \[alg:data\_sharing\] requires a total number of $T_1$ weight updates, the overall complexity order for Algorithm \[alg:data\_sharing\] is then $O\left( T_1\left(L + K \right) \left(LK\right)^3\right)$.
Note that in the above complexity order, $K$ is the number of scheduled users, which is comparable to the number of active BSs in the network. In addition, instead of considering all the $L$ BSs in the entire network, we can set the nearest $L_c < L$ BSs around each scheduled user as its candidate serving BS cluster. This further reduces the computational complexity for Algorithm \[alg:data\_sharing\] with negligible performance loss.
Generalization to the Multi-Antenna System
------------------------------------------
Algorithm \[alg:data\_sharing\] can be readily generalized to the case with multiple transmit antennas at each BS. In such case, one only needs to replace the beamforming coefficient $w_{lk}$ with the beamforming vector $\mathbf{w}_{lk} \in \mathbb{C}^{N_l \times 1}$ from BS $l$ to user $k$, where $N_l$ is the number of antennas at BS $l$. The rest of the optimization parameters are straightforward extensions based on $\mathbf{w}_{lk}$ [@binbin13].
Algorithm \[alg:data\_sharing\] can also be applied to the case with multiple receive antennas at each user but with fixed receive beamformer. In this case, the only change is to replace the channel gain vector $\mathbf{h}_{k}$ with the effective channel gain $\tilde{\mathbf{h}}_{k} = \mathbf{H}_{k} \mathbf{u}_{k}$, where $\mathbf{H}_{k} \in \mathbb{C}^{L \times M_k}$ and $\mathbf{u}_{k} \in \mathbb{C}^{M_k \times 1}$ are the channel matrix seen by user $k$ and the receive beamformer at user $k$, $M_k$ is the number receive antennas at user $k$. However, the joint design of transmit beamformer and receive beamformer for the multiple receive antennas case is more complicated. One possible way is to iteratively design the transmit beamformer assuming fixed receive beamformer and update the receiver as the optimal minimum mean square error (MMSE) beamformer.
\[tc\]\[Bc\]\[1\][Central Processor]{} \[tc\]\[Bc\]\[0.9\][$\hat{x}_l = w_{l1}s_1 + w_{l2}s_2$]{} \[tc\]\[Bc\]\[0.9\][$l = 1, 2, 3$]{} \[tc\]\[tc\]\[0.9\][$x_1$]{} \[tc\]\[tc\]\[0.9\][$x_2$]{} \[tc\]\[tc\]\[0.9\][$x_3$]{} \[tl\]\[tc\]\[0.8\][$R_1^{BH}$]{} \[tl\]\[tc\]\[0.8\][$R_2^{BH}$]{} \[tl\]\[tc\]\[0.8\][$R_3^{BH}$]{} \[tc\]\[cl\]\[0.8\][$x_1 = \hat{x}_1 + e_1$]{} \[tc\]\[cl\]\[0.8\][$x_2 = \hat{x}_2 + e_2$]{} \[tc\]\[cl\]\[0.8\][$x_3 = \hat{x}_3 + e_3$]{} \[tc\]\[tc\]\[0.8\][BS $1$]{} \[tc\]\[tc\]\[0.8\][BS $2$]{} \[tc\]\[tc\]\[0.8\][BS $3$]{} \[tc\]\[tc\]\[0.8\][User $1$]{} \[tc\]\[tc\]\[0.8\][User $2$]{}
Compression Strategy {#sec:compression}
====================
In this section, we aim to minimize the total power consumption for downlink C-RAN under the compression strategy.
Problem Formulation
-------------------
Consider the compression transmission strategy for downlink C-RAN as illustrated in Fig. \[fig:CompressionModel\]. Let $\hat{x}_l = \sum_{k \in \mathcal{K}} w_{lk} s_k$ denote the beamformed signal formed in the CP for BS $l$. The CP compresses $\hat{x}_l$ into $x_l$ and sends $x_l$ to BS $l$. In this paper, we assume that each $\hat{x}_l$ is compressed independently[^4] and model the compression procedure as the following forward test channel: $$\label{eq:compressX}
x_l = \hat{x}_l + e_l, \quad l \in \mathcal{L},$$ where $e_l \in \mathbb{C}$ is the quantization noise independent of $\hat{x}_l$ and is assumed to be Gaussian distributed with zero mean and variance $q_l^{2}$. Substituting to , the transmit power at BS $l$ under the compression strategy can be written as $$\label{eq:Tx_power_compression}
P_{l, tx} = \sum_{k \in \mathcal{K}} \left\vert w_{lk} \right\vert^{2} + q_l^{2}, \quad l \in \mathcal{L}.$$ Comparing with , we can see that different from the data-sharing strategy, the BS transmit power in the compression strategy involves a quantization noise power in addition to the beamforming power.
Substituting into , the received signal $y_k$ at user $k$ under the compression strategy can be written as $$\label{eq:y_k_compression}
y_k = \mathbf{h}_k^{H} \mathbf{w}_k s_k + \sum_{j \neq k}\mathbf{h}_k^{H} \mathbf{w}_j s_j + \mathbf{h}_k^{H} \mathbf{e} + n_k, \quad k \in \mathcal{K},$$ where $\mathbf{e} = \left[e_1, e_2, \cdots, e_L \right]^{T}$ is the quantization noise vector transmitted from all the $L$ BSs. As we can see, besides the inter-user interference and background noise, each user now also receives an additional quantization noise term $\mathbf{h}_k^{H} \mathbf{e}$ from the BSs. The user received SINR is expressed as $$\label{eq:sinr_compression}
\text{SINR}_k =
\frac{\left\vert\mathbf{h}_k^{H}\mathbf{w}_k \right\vert^2}{\sum_{j \neq
k}\left\vert\mathbf{h}_k^{H}\mathbf{w}_j\right\vert^2 +
\sum_{l \in \mathcal{L}} \left\vert h_{lk} q_l \right\vert^2 + \sigma^2}, \quad k \in \mathcal{K}.$$
The backhaul capacity consumption for the compression strategy is related to the level of quantization noise $q_l^2$: lower quantization noise requires higher backhaul rate. Under the forward test channel , the achievable compression rate is the mutual information between $x$ and $\hat{x}$, according to rate-distortion theory [@EIT], given as $\log_2 \left( 1 + \frac{\sum_{k \in \mathcal{K}} \left\vert w_{lk} \right\vert^{2}}{q_l^2} \right)$, where $\sum_{k \in \mathcal{K}} \left\vert w_{lk} \right\vert^{2}$ is the power of the signal to be compressed, i.e. $\hat{x}_l$. However, practical quantizer may be far from the theoretically ideal quantizer. Similar to [@PratikEUSIPCO], we introduce a notion of gap to rate-distortion limit, denote as $\Gamma_q > 1$, to account for the loss due to practical quantizer and formulate the backhaul capacity consumption for BS $l$ as $$\label{eq:bkhaul_compression}
R_{l}^{BH} = \log_2 \left( 1 + \frac{\Gamma_q \sum_{k \in \mathcal{K}} \left\vert w_{lk} \right\vert^{2}}{q_l^2} \right), \quad l \in \mathcal{L}.$$
Substituting and into , the total power minimization problem for the compression strategy is formulated as follows
\[prob:compression\] $$\begin{aligned}
& \hspace{-5mm} \displaystyle \min_{\left\{w_{lk}, q_l\right\}} & \hspace{-1mm} \sum_{l\in\mathcal{L}}
\Bigg( \eta_l \left( \sum_{k \in \mathcal{K}} \left\vert w_{lk} \right\vert^{2} + q_l^2 \right)+
\mathbbm{1}_{\left\{ \sum_{k \in \mathcal{K}} \left\vert w_{lk} \right\vert^{2} + q_l^2 \right\}} P_{l, \Delta} \nonumber \\
& \quad &\hspace{1.2cm} + \rho_l \log_2 \left( 1 + \frac{\Gamma_q \sum_{k \in \mathcal{K}} \left\vert w_{lk} \right\vert^{2}}{q_l^2} \right) \Bigg)
\label{obj_compression} \\
& \hspace{-5mm} \operatorname{s.t.}& \hspace{-1mm} \text{SINR}_k \geq \gamma_k, \quad k \in \mathcal{K} \label{sinr_const_compression} \\
& \quad & \hspace{-1mm} \sum_{k \in \mathcal{K}} \left\vert w_{lk} \right\vert^{2} + q_l^2 \leq P_l, \quad l \in \mathcal{L} \label{compression_power_const}\end{aligned}$$
where the SINR in is defined in . Due to the indicator function as well as the backhaul rate expression in , the optimization problem is nonconvex. In the following, we describe the techniques to approximate in a convex form.
Proposed Algorithm
------------------
The difficulties in solving problem lie in both the indicator function and the nonconvex backhaul rate expression in the objective function . For the indicator function, we can utilize the similar technique used in the previous section to approximate it using reweighted $\ell_1$-norm: $$\begin{aligned}
\label{eq:rewgt_compression}
\mathbbm{1}_{\left\{ \sum_{k \in \mathcal{K}} \left\vert w_{lk} \right\vert^{2} + q_l^2 \right\}}
& = \left\Vert \sum_{k \in \mathcal{K}} \left\vert w_{lk} \right\vert^{2} + q_l^2 \right\Vert_0 \nonumber \\
&\approx \beta_l \left( \sum_{k \in \mathcal{K}} \left\vert w_{lk} \right\vert^{2} + q_l^2 \right) \end{aligned}$$ where $\beta_l$ is iteratively updated according to the following reweighting function $$\label{wgt_mu}
\beta_l = f\left(\sum_{k \in \mathcal{K}} \left\vert w_{lk} \right\vert^{2} + q_l^2 , \tau_3\right)
= \frac{c_3}{\sum_{k \in \mathcal{K}} \left\vert w_{lk} \right\vert^{2} + q_l^2 + \tau_3}$$ where $\left\{w_{lk}, q_l \right\}$ come from the previous iteration, $\tau_3 > 0$ is some constant regularization factor, and $c_3$ is a constant.
For the backhaul rate , we can express it as a difference of two logarithmic functions: $\log_2 \left( q_l^2 + \Gamma_q \sum_{k \in \mathcal{K}} \left\vert w_{lk} \right\vert^{2} \right) - 2 \rho_l \log_2 q_l$. Although the second term $- 2 \rho_l \log_2 q_l$ is convex in $q_l$, the first term is still nonconvex. To deal with the nonconvexity of the backhaul rate, we propose to successively approximate the first logarithmic function using the following inequality $$\begin{aligned}
\label{ineq:compression}
&\log_2 \left( q_l^2 + \Gamma_q \sum_{k \in \mathcal{K}} \left\vert w_{lk} \right\vert^{2} \right) \nonumber \\
& \leq ~ \log_2 \lambda_l + \frac{ q_l^2 + \Gamma_q \sum_{k \in \mathcal{K}} \left\vert w_{lk} \right\vert^{2}}{\lambda_l \ln 2} -
\frac{1}{\ln 2} \end{aligned}$$ due to the concavity of $\log_2(x)$. The above inequality achieves equality if and only if $$\label{lambda_update}
\lambda_l = q_l^2 + \Gamma_q \sum_{k \in \mathcal{K}} \left\vert w_{lk} \right\vert^{2}.$$ The right-hand side of is a convex quadratic function in $\left\{ w_{lk}, q_l \right\}$ for fixed $\lambda_l$. This fact motivates us to successively solve the problem with $\log_2 \left( q_l^2 + \Gamma_q \sum_{k \in \mathcal{K}} \left\vert w_{lk} \right\vert^{2} \right)$ replaced by the right-hand side of for fixed $\lambda_l$, then to iteratively update $\lambda_l$ according to .
Combining the above described $\ell_1$-norm reweighting and successive convex approximation techniques, we get the resulting optimization problem under fixed $\beta_l$ and $\lambda_l$ as $$\begin{aligned}
\label{prob:approx_compression1}
\displaystyle \operatorname*{minimize}_{\left\{w_{lk}, q_l\right\}} & \quad \sum_{l\in\mathcal{L}} \sum_{k\in\mathcal{K}} \phi_{l} \left\vert w_{lk} \right\vert^{2} + \sum_{l\in\mathcal{L}} \left( \psi_{l} q_l^2 - 2 \rho_l \log_2 q_l\right) \nonumber \\
& \hspace{1.5cm} + \underbrace{\sum_{l\in\mathcal{L}} \rho_l \left( \log_2 \lambda_l - \frac{1}{\ln 2} \right)}_{\text{constant}} \\
\operatorname{subject \text{ } to}& \quad \eqref{sinr_const_compression}, ~~ \eqref{compression_power_const} \nonumber\end{aligned}$$ where $\phi_{l} = \eta_l + \beta_l P_{l, \Delta} + \frac{\rho_l \Gamma_q}{\lambda_l \ln 2}$ and $\psi_{l} = \eta_l + \beta_l P_{l, \Delta} + \frac{\rho_l}{\lambda_l \ln 2}$. Similar to the SINR constraint in , the constraint can also be equivalently reformulated as an SOC constraint. Thus, problem is a convex optimization problem and can be solved efficiently using standard convex optimization solver, e.g. [@CVX], with polynomial complexity.
[**Initialization**]{}: Set the initial values for $\left\{ \beta_l \right\}$ and $\left\{ \lambda_l \right\}$ according to and respectively with $\left\{ w_{lk}, q_l \right\}$ chosen as a feasible point of problem ;\
[**Repeat**]{}:
1. Fix $\left\{ \beta_l, \lambda_l \right\}$, find the optimal $\left\{ w_{lk}, q_l \right\}$ by solving the convex optimization problem ;
2. Update $\left\{ \beta_l \right\}$ and $\left\{ \lambda_l \right\}$ according to and respectively.
[**Until**]{} convergence
We summarize the proposed algorithm for solving problem in Algorithm \[alg:compression\], which admits guaranteed convergence property as stated in the following theorem.
\[thm:2\] Starting with any initial point, the sequence $\left\{ w_{lk}^{(n)}, q_l^{(n)} \right\}_{n=1}^{\infty}$ generated by Algorithm \[alg:compression\] with the reweighting function in chosen as is guaranteed to converge.
See Appdendix \[apdx:b\]
Algorithm \[alg:compression\] shows a similar computational complexity as Algorithm \[alg:data\_sharing\] but with additional $L$ quantization noise variables to be optimized in each iteration. Assuming that Algorithm \[alg:compression\] converges in $T_2$ iterations, its complexity order is then given as $O\left( T_2\left(L + K \right) \left(LK + L\right)^3 \right)$.
Generalization to the Multi-Antenna System
------------------------------------------
Algorithm \[alg:compression\] can be readily applied to the scenario where multiple transmit antennas are available at the BSs assuming that the CP performs independent compression for each antenna. Joint compression among the antennas may improve the performance but results in a different optimization problem. Similar to Algorithm \[alg:data\_sharing\], generalization of Algorithm \[alg:compression\] to multiple receive antennas at the user side is also straightforward if the receive beamformer is assumed to be fixed, however, joint design of transmit and receive beamformer is by no means trivial and requires additional efforts.
Numerical Evaluation of Energy Efficiency {#sec:simulations}
=========================================
--------------------------------------------------- -------------------------------
Cellular Hexagonal
Layout $7$-cell wrapped-around
Channel bandwidth $10$ MHz
Distance between cells $0.8$ km
Number of RRHs$/$cell $4$
Number of antennas$/$(RRH, user) $(1, 1)$
Maximum transmit power for RRH $P_l$ $20$ Watts
Active mode power for RRH $P_{l, active}$ $84$ Watts
Sleep mode power for RRH $P_{l, sleep}$ $56$ Watts
Slope of transmit power $\eta_l$ $2.8$
Backhaul link capacity $C_l$ $100$ Mbps
Maximum backhaul power $P_{l,max}^{BH}$ $50$ Watts
Antenna gain $15$ dBi
Background noise $-169$ dBm/Hz
Path loss from RRH to user $128.1+ 37.6 \log_{10}(d)$
Log-normal shadowing $8$ dB
Rayleigh small scale fading $0$ dB
SNR gap $\Gamma_m$ $0$ dB
Gap to rate-distortion limit $\Gamma_q$ $4.3$ dB
Reweighting parameters ($\tau_1, \tau_2, \tau_3$) $(10^{-5}, 10^{-8}, 10^{-5})$
--------------------------------------------------- -------------------------------
: Simulation Parameters.[]{data-label="table:system-parameter"}
\[cc\]\[cc\]\[0.8\][km]{} \[cc\]\[cc\]\[0.8\][km]{} \[cr\]\[cr\]\[0.7\][$0$]{} \[cr\]\[cr\]\[0.7\][$0.5$]{} \[cr\]\[cr\]\[0.7\][$1$]{} \[cr\]\[cr\]\[0.7\][$1.5$]{} \[cr\]\[cr\]\[0.7\][$-0.5$]{} \[cr\]\[cr\]\[0.7\][$-1$]{} \[cc\]\[cc\]\[0.7\][$-1.5$]{} \[cr\]\[cr\]\[0.7\][$0$]{} \[cr\]\[cr\]\[0.7\][$0.5$]{} \[cr\]\[cr\]\[0.7\][$1$]{} \[cr\]\[cr\]\[0.7\][$1.5$]{} \[cr\]\[cr\]\[0.7\][$-0.5$]{} \[cr\]\[cr\]\[0.7\][$-1$]{} \[cc\]\[cc\]\[0.7\][$-1.5$]{}
![A cellular topology with $7$ cells and $4$ RRHs each cell, where each dot represents a RRH.[]{data-label="fig:topology"}](topology.eps){width="45.00000%"}
In this section, we evaluate the energy efficiency of the data-sharing and compression strategies for downlink C-RAN using the proposed algorithms for a $7$-cell network with wrapped-around topology. Each cell here refers to a geographic area with $4$ RRHs as shown in Fig. \[fig:topology\]. Equivalently, the network consists of $28$ BSs. The out-of-cell interference combined with background noise is set as $-150$ dBm$/$Hz. The gap to rate-distortion limit is set as $\Gamma_q = 4.3$ dB corresponding to the uncoded fixed rate uniform scalar quantizer [@Gray98]. For simplicity, the SNR gap is set to be $\Gamma_m = 0$ dB. The parameters in the BS power consumption model are taken from [@Auer11] while the parameters for the backhaul power model are from [@Fehske10]. All the parameters related to the simulations are listed in Table \[table:system-parameter\].
\[tc\]\[Bc\]\[0.8\][Iteration ]{} \[Bc\]\[tc\]\[0.8\][Number of Active BSs]{} \[Bl\]\[Bl\]\[0.65\][Data-Sharing (Alg.1) w/ $1$ user$/$cell]{} \[Bl\]\[Bl\]\[0.65\][Data-Sharing (Alg.1) w/ $2$ users$/$cell]{} \[Bl\]\[Bl\]\[0.65\][Data-Sharing (Alg.1) w/ $3$ users$/$cell]{} \[Bl\]\[Bl\]\[0.65\][Compression (Alg.2) w/ $1$ user$/$cell]{} \[Bl\]\[Bl\]\[0.65\][Compression (Alg.2) w/ $2$ users$/$cell]{} \[Bl\]\[Bl\]\[0.65\][Compression (Alg.2) w/ $3$ users$/$cell]{} \[cc\]\[cr\]\[0.7\][$10$]{} \[cc\]\[cc\]\[0.7\][$20$]{} \[cc\]\[cc\]\[0.7\][$30$]{} \[cc\]\[cc\]\[0.7\][$40$]{} \[cc\]\[cc\]\[0.7\][$50$]{} \[cc\]\[cc\]\[0.7\][$60$]{} \[cc\]\[cc\]\[0.7\][$70$]{} \[rc\]\[rc\]\[0.7\][$5$]{} \[rc\]\[rc\]\[0.7\][$10$]{} \[rc\]\[rc\]\[0.7\][$15$]{} \[rc\]\[rc\]\[0.7\][$20$]{} \[rc\]\[rc\]\[0.7\][$25$]{} \[rc\]\[rc\]\[0.7\][$30$]{} \[rc\]\[rc\]\[0.7\][$35$]{} ![Trajectories of number of active BSs under $r_k = 20$ Mbps target rate for each user.[]{data-label="fig:DataSharingActiveBSs"}](Per_Iteration_Active_BS.eps "fig:"){width="45.00000%"}
\[tc\]\[Bc\]\[0.8\][Iteration]{} \[Bc\]\[tc\]\[0.7\][Objective Value of Problem $/$]{} \[Bl\]\[Bl\]\[0.65\][Data-Sharing (Alg.1) w/ $1$ user$/$cell]{} \[Bl\]\[Bl\]\[0.65\][Data-Sharing (Alg.1) w/ $2$ users$/$cell]{} \[Bl\]\[Bl\]\[0.65\][Data-Sharing (Alg.1) w/ $3$ users$/$cell]{} \[Bl\]\[Bl\]\[0.65\][Compression (Alg.2) w/ $1$ user$/$cell]{} \[Bl\]\[Bl\]\[0.65\][Compression (Alg.2) w/ $2$ users$/$cell]{} \[Bl\]\[Bl\]\[0.65\][Compression (Alg.2) w/ $3$ users$/$cell]{} \[cc\]\[cr\]\[0.7\][$10$]{} \[cc\]\[cc\]\[0.7\][$20$]{} \[cc\]\[cc\]\[0.7\][$30$]{} \[cc\]\[cc\]\[0.7\][$40$]{} \[cc\]\[cc\]\[0.7\][$50$]{} \[cc\]\[cc\]\[0.7\][$60$]{} \[cc\]\[cc\]\[0.7\][$70$]{} \[cc\]\[cc\]\[0.7\][$80$]{} \[rc\]\[rc\]\[0.7\][$0$]{} \[rc\]\[rc\]\[0.7\][$400$]{} \[rc\]\[rc\]\[0.7\][$800$]{} \[rc\]\[rc\]\[0.7\][$1200$]{} \[rc\]\[rc\]\[0.7\][$1600$]{} \[rc\]\[rc\]\[0.7\][$2000$]{} ![Convergence behavior of proposed algorithms under $r_k = 20$ Mbps target rate for each user.[]{data-label="fig:compressionObj"}](Per_Iteration_Obj.eps "fig:"){width="45.00000%"}
We first evaluate the effectiveness of the proposed $\ell_1$-norm reweighting technique in turning off BSs. We plot the number of active BSs remained in each iteration for both data-sharing (Algorithm \[alg:data\_sharing\]) and compression (Algorithm \[alg:compression\]) strategies in Fig. \[fig:DataSharingActiveBSs\]. The user target rate is set to be $20$ Mbps for every user and different number of scheduled users are simulated. Instead of considering all the $28$ BSs in the entire network as potential serving BSs for each user, we set the initial BS cluster for each user as the strongest $L_c = 14$ BSs to reduce the amount of CSI acquisition for the CP. From Fig. \[fig:DataSharingActiveBSs\] we can see that all the BSs are active at the first iteration, however, the number of active BSs decreases as the iteration goes on. Intuitively, more users are served, more BSs need to remain active. This can be verified from Fig. \[fig:DataSharingActiveBSs\]. We also observe from Fig. \[fig:DataSharingActiveBSs\] that Algorithm \[alg:data\_sharing\] for data-sharing exhibits faster convergence speed than Algorithm \[alg:compression\] for compression, where the former converges within $30$ iterations while the later requires $50$ iterations to converge in the worst case.
We then evaluate the convergence behavior of the proposed algorithms in Fig. \[fig:compressionObj\]. As in Fig. \[fig:DataSharingActiveBSs\], different number of scheduled users are tested and each user’s target rate is set to be $20$ Mbps. As we can see, the objective values monotonically decrease and converge for both the data-sharing (Algorithm \[alg:data\_sharing\]) and the compression (Algorithm \[alg:compression\]) strategies. Similar to Fig. \[fig:DataSharingActiveBSs\], we also observe faster convergence speed for data-sharing than for compression in Fig. \[fig:compressionObj\]. This is possibly due to the fact that compression strategy involves more variables to be optimized.
We now compare the performance of the data-sharing strategy and the compression strategy in terms of power saving in Fig. \[sim:total\_power\]. In addition, we consider two reference schemes. In the first scheme, each user is only served by its strongest BS that is not already associated with another user. This scheme is termed as “Single BS Association”, for which the transmit power for each user can be minimized using the strategy in [@yates1995]. In the second scheme, each user’s message is shared among the $4$ RRHs in its own cell and is cooperatively served by the $4$ RRHs using the coordinated beamforming strategy of [@dahrouj10]. Such scheme is termed as “Per-Cell CoMP”. The “Single BS Association” and “Per-Cell CoMP” are two extreme cases in terms of number of active BSs: the former only has $K$ (number of scheduled users) active BSs while in the latter all the $28$ BSs in the entire network remain active.Each point in Fig. \[sim:total\_power\] is averaged over $100$ channel realizations.
As we can see from Fig. \[sim:total\_power\], “Per-Cell CoMP” consumes the most power since all the BSs are active in this scheme. “Single BS Association” consumes the least power, similar to the data-sharing strategy, but only at low user rate regime because the minimum number of BSs are selected to serve the users in this scheme. However, as the user rate or the number of scheduled users increases, “Single BS Association” becomes infeasible very quickly. For instance, in Fig. \[sim:total\_power\](b) where there are $2$ users per cell, “Single BS Association” can only support each user with $10$ Mbps service rate, while in Fig. \[sim:total\_power\](c) the case of 3 users per cell, “Single BS Association” is not feasible even at $10$ Mbps per user.
\[tc\]\[cc\]\[0.7\][User Target Rate (Mbps)]{} \[tc\]\[cc\]\[0.7\][Total Power Consumption (kWatts)]{} \[Bl\]\[Bl\]\[0.65\][Compression]{} \[Bl\]\[Bl\]\[0.65\][Data-Sharing]{} \[Bl\]\[Bl\]\[0.65\][Per-Cell CoMP]{} \[Bl\]\[Bl\]\[0.65\][Single BS Association]{} \[cc\]\[cc\]\[0.7\][$10$]{} \[cc\]\[cc\]\[0.7\][$20$]{} \[cc\]\[cc\]\[0.7\][$30$]{} \[cc\]\[cc\]\[0.7\][$40$]{} \[cc\]\[cc\]\[0.7\][$50$]{} \[cc\]\[cc\]\[0.7\][$60$]{} \[cc\]\[cc\]\[0.7\][$70$]{} \[cc\]\[cc\]\[0.7\][$1.5$]{} \[cc\]\[cc\]\[0.7\][$2.0$]{} \[cc\]\[cc\]\[0.7\][$2.5$]{} \[cc\]\[cc\]\[0.7\][$3.0$]{} \[cc\]\[cc\]\[0.7\][$3.5$]{} \[cc\]\[cc\]\[0.7\][$4.0$]{} \[cc\]\[cc\]\[0.7\][$4.5$]{} \[cc\]\[cc\]\[0.7\][$5.0$]{}
It is worth noting that “Single BS Association” consumes the same amount of power as data-sharing in the low user target rate regime in Fig. \[sim:total\_power\], as the latter essentially reduces to single BS association at low user rates. However, there is still significant advantage in migrating signal processing to the cloud in a C-RAN as compared to the conventional cellular architecture in term of computation power saving, which is not included in the model in this paper. It is also worth noting that the optimized data-sharing and compression strategies outperform the non-optimized per-Cell CoMP significantly in Fig. \[sim:total\_power\], highlighting the importance of optimization approaches proposed in this paper. Other optimized CoMP schemes with larger cooperation cluster may consume less BS transmission power, however, the overall power consumptions can be still very high if all the BSs remain active.
We also observe from Fig. \[sim:total\_power\] that neither the data-sharing nor the compression strategy dominates the other over the entire user target rate regime. For example in Fig. \[sim:total\_power\](b), although data-sharing consumes less power than compression when the user target rate is below $30$ Mbps, its power consumption increases dramatically with the user rate and eventually crosses over the total power consumed by compression after $40$ Mbps target rate. Similar trend can be observed from Fig. \[sim:total\_power\](a) and \[sim:total\_power\](c). This trend is parallel to the observation made in [@PratikEUSIPCO], in which these two downlink strategies are compared from the utility maximization perspective with limited backhaul constraint. It is observed in [@PratikEUSIPCO] that with low backhaul rate, data-sharing produces higher utility than compression while with high backhaul rate, compression outperforms data-sharing.
\[tc\]\[Bc\]\[0.7\][User Target Rate (Mbps) ]{} \[Bc\]\[tc\]\[0.7\][Power (kWatts)]{} \[Bl\]\[Bl\]\[0.7\][Data-Sharing BS Power]{} \[Bl\]\[Bl\]\[0.7\][Data-Sharing Backhaul Power]{} \[Bl\]\[Bl\]\[0.7\][Compression BS Power]{} \[Bl\]\[Bl\]\[0.7\][Compression Backhaul Power]{} \[cc\]\[cc\]\[0.85\][$10$]{} \[cc\]\[cc\]\[0.85\][$20$]{} \[cc\]\[cc\]\[0.85\][$30$]{} \[cc\]\[cc\]\[0.85\][$40$]{} \[cc\]\[cc\]\[0.85\][$50$]{} \[cc\]\[cc\]\[0.85\][$60$]{} \[cc\]\[cc\]\[0.85\][$70$]{} \[cc\]\[cc\]\[0.85\][$1$]{} \[cc\]\[cc\]\[0.85\][$2$]{} \[cc\]\[cc\]\[0.85\][$3$]{} \[cc\]\[cc\]\[0.85\][$4$]{} \[cc\]\[cc\]\[0.85\][$5$]{} \[cc\]\[cc\]\[0.85\][$6$]{} \[cc\]\[cc\]\[0.85\][$7$]{} \[cc\]\[cc\]\[0.85\][$0$]{}
To investigate further, we plot the individual power consumption of BSs and backhaul links for both the data-sharing and the compression strategies in Fig. \[fig:power\_decomp\] for the case of $2$ users per cell. As seen from Fig. \[fig:power\_decomp\], although the BS power consumptions for data-sharing and compression are similar in each case of the user target rate, the backhaul power consumptions are significantly different and are the determining factor in the choice of strategies. As the user target rate increases, the backhaul power consumption for data-sharing increases significantly and crosses over the compression strategy at around $30$ Mbps. This is because in data-sharing, each user’s message needs to be delivered to each one of its serving BSs through backhaul links. So, the backhaul rate directly depends on both the user target rate and the BS cluster size. Note that as user target rate increases, the size of serving BS cluster also increases. The two factors together contribute to a much higher backhaul rate. In contrast, the backhaul rate of compression strategy depends on the logarithm of the signal-to-quantization-noise ratio, which only increases gradually as user target rate increases. Also, note that in the low user rate regime, data-sharing consumes less backhaul power than compression in Fig. \[fig:power\_decomp\]. This is because it is more efficient to share data directly than to compress when only a few BSs are involved.
In Fig. \[sim:active\_bs\], we compare the percentage of active BSs in the data-sharing strategy versus in the compression strategy. Similar to Fig. \[sim:total\_power\], each point in Fig. \[sim:active\_bs\] is averaged over $100$ channel realizations. As we can see, with higher user target rate and more users to be served, more BSs need to remain active for transmission. Also, from Fig. \[sim:active\_bs\], it is observed that the compression strategy tends to turn off more BSs than the data-sharing strategy.
\[tc\]\[cc\]\[0.6\][User Target Rate (Mbps)]{} \[cc\]\[cc\]\[0.6\][[Fraction of Active BSs]{}]{} \[Bl\]\[Bl\]\[0.65\][Compression]{} \[Bl\]\[Bl\]\[0.65\][Data Sharing]{} \[cc\]\[cr\]\[0.7\][$10$]{} \[cc\]\[cc\]\[0.7\][$20$]{} \[cc\]\[cc\]\[0.7\][$30$]{} \[cc\]\[cc\]\[0.7\][$40$]{} \[cc\]\[cc\]\[0.7\][$50$]{} \[cc\]\[cc\]\[0.7\][$60$]{} \[cc\]\[cc\]\[0.7\][$70$]{} \[cc\]\[cc\]\[0.7\][$0.2$]{} \[cc\]\[cc\]\[0.7\][$0.3$]{} \[cc\]\[cc\]\[0.7\][$0.4$]{} \[cc\]\[cc\]\[0.7\][$0.5$]{} \[cc\]\[cc\]\[0.7\][$0.6$]{} \[cc\]\[cc\]\[0.7\][$0.7$]{} \[cc\]\[cc\]\[0.7\][$0.8$]{} \[cc\]\[cc\]\[0.7\][$0.9$]{} \[cc\]\[cc\]\[0.7\][$0.65$]{} \[cc\]\[cc\]\[0.7\][$0.75$]{} \[cc\]\[cc\]\[0.7\][$0.85$]{} \[cc\]\[cc\]\[0.7\][$0.95$]{} \[cc\]\[cl\]\[0.8\][$1.0$]{}
Conclusion {#sec:conclusion}
==========
This paper compares the energy efficiency between the data-sharing strategy and the compression strategy in downlink C-RAN. We formulate the problem as that of minimizing the total network power consumption subject to user target rate constraints, with both the BS power consumption and the backhaul power consumption taken into account. By taking advantage of the $\ell_1$-norm reweighting technique and successive convex approximation technique, we transform the nonconvex optimization problems into convex form and devise efficient algorithms with provable convergence guarantees.
The main conclusions of this paper are that C-RAN significantly improves the range of feasible user data rates in a wireless cellular network, and that both data-sharing and compression strategies bring much improved energy efficiency to downlink C-RAN as compared to non-optimized CoMP. Moreover, between the data-sharing strategy and the compression strategy, either may be preferred depending on the different target rate regimes: at low user target rate, data-sharing consumes less power, while at high user target rate compression is preferred since the backhaul rate for data-sharing increases significantly as user rate increases.
[Binbin Dai]{} (S’12) received the B.E. degree in Information Science and Engineering from Chien-Shiung Wu Honors College, Southeast University, Nanjing, China, in 2011 and M.A.Sc degree in Electrical and Computer Engineering from University of Toronto, Toronto, Ontario, Canada, in 2014. He is currently working towards the Ph.D degree with the Department of Electrical and Computer Engineering, University of Toronto, Toronto, Ontario, Canada. His research interests include optimization, wireless communications and signal processing.
[Wei Yu]{} (S’97-M’02-SM’08-F’14) received the B.A.Sc. degree in Computer Engineering and Mathematics from the University of Waterloo, Waterloo, Ontario, Canada in 1997 and M.S. and Ph.D. degrees in Electrical Engineering from Stanford University, Stanford, CA, in 1998 and 2002, respectively. Since 2002, he has been with the Electrical and Computer Engineering Department at the University of Toronto, Toronto, Ontario, Canada, where he is now Professor and holds a Canada Research Chair (Tier 1) in Information Theory and Wireless Communications. His main research interests include information theory, optimization, wireless communications and broadband access networks.
Prof. Wei Yu currently serves on the IEEE Information Theory Society Board of Governors (2015-17). He is an IEEE Communications Society Distinguished Lecturer (2015-16). He served as an Associate Editor for <span style="font-variant:small-caps;">IEEE Transactions on Information Theory</span> (2010-2013), as an Editor for <span style="font-variant:small-caps;">IEEE Transactions on Communications</span> (2009-2011), as an Editor for <span style="font-variant:small-caps;">IEEE Transactions on Wireless Communications</span> (2004-2007), and as a Guest Editor for a number of special issues for the <span style="font-variant:small-caps;">IEEE Journal on Selected Areas in Communications</span> and the <span style="font-variant:small-caps;">EURASIP Journal on Applied Signal Processing</span>. He was a Technical Program co-chair of the IEEE Communication Theory Workshop in 2014, and a Technical Program Committee co-chair of the Communication Theory Symposium at the IEEE International Conference on Communications (ICC) in 2012. He was a member of the Signal Processing for Communications and Networking Technical Committee of the IEEE Signal Processing Society (2008-2013). Prof. Wei Yu received a Steacie Memorial Fellowship in 2015, an IEEE Communications Society Best Tutorial Paper Award in 2015, an IEEE ICC Best Paper Award in 2013, an IEEE Signal Processing Society Best Paper Award in 2008, the McCharles Prize for Early Career Research Distinction in 2008, the Early Career Teaching Award from the Faculty of Applied Science and Engineering, University of Toronto in 2007, and an Early Researcher Award from Ontario in 2006. He was named a Highly Cited Researcher by Thomson Reuters in 2014.
[^1]: Manuscript submitted on April 15, 2015; revised on September 15, 2015; accepted on December 11, 2015. This work was supported by Huawei Technologies, Canada, and Natural Sciences and Engineering Research Council (NSERC) of Canada. The authors are with The Edward S. Rogers Sr. Department of Electrical and Computer Engineering, University of Toronto, Toronto, ON M5S 3G4, Canada (e-mails: {bdai, weiyu}@comm.utoronto.ca).
[^2]: This paper refers the link between CP and BSs as *backhaul*, which is appropriate if the data-sharing strategy is used. However, in a C-RAN architecture implementing the compression strategy where the BSs are simply RRHs, the connection between RRH and CP can be referred to more appropriately as *fronthaul*.
[^3]: In fact, it converges to the stationary point solution of an approximation to problem .
[^4]: Correlated compression is also possible and has been considered in [@Park13].
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
author:
- Ben Webster
- Jerry Guan
bibliography:
- 'gen.bib'
---
[11pt]{} [11pt]{} [**]{} [****]{} [.5em]{}
\[section\] \[theorem\][Proposition]{} \[theorem\][Corollary]{} \[theorem\][Lemma]{} \[theorem\][Assumption]{} \[theorem\][Definition]{} \[theorem\][Question]{} \[theorem\][Conjecture]{}
\[theorem\][Remark]{} \[section\]
[.]{}
addtoreset[subeqn]{}[equation]{}
[\^C]{}
[T]{}
[L\_]{}
[q\^[1D]{}]{}
[M\_]{}
\[1\]
[D\^]{}
[D\^]{}
[o]{}
[S]{}
[\_]{}
[,]{}
[,]{}
[Sto[š]{}i[ć]{}]{}
[h]{}
\[3\][\^[\#1]{}\_[\#2,\#3]{}]{}
\[2\][\^[\#1]{}\_[\#2]{}]{}
\[3\][(\#1,\#2,\#3)]{}
[highest weight]{}
[\\]{}
[R\^\_A]{}
[S\_h]{}
[S\_1]{}
\[2\][\_[\#1]{}\^[\#2]{}]{}
\[2\][E\_[\#1]{}\^[\#2]{}]{}
\[1\]
[p]{}
[q]{}
[l]{}
[b\^\_[\_0]{}]{}
\[1\]
\[1\][(\#1 ,)]{}
\[1\][\_[\#1]{}]{}
[R\^(X\_0)\^]{}
\[1\][\^\_[\#1]{}]{}
\[1\][\_[\#1]{}]{}
[\_[0]{}]{}
[X]{}
\[1\][\^[\#1]{}]{}
[U\_q()]{}
[X]{}
[Y]{}
\[2\][\_[\#1]{}\^[\#2]{}]{}
\[2\][\_[\#1]{}\^[\#2]{}]{}
[\^\_[|]{}]{}
[\^L]{}
[\^R]{}
[\_i]{}
[\_i]{}
\[2\][\^R\_[\#1,\#2]{}]{}
\[2\][\^L\_[\#1,\#2]{}]{}
\[2\][\^[\#1]{}\_[\#2]{}]{}
\[2\][\_[\#1,\#2]{}]{}
[\^g]{}
[\^[fd]{}]{}
[\^\_[A]{}]{}
\[2\][B\_[\#1]{}\^[\#2]{}]{}
\[1\][A\_[\#1]{}]{}
[A]{}
\[1\][(\#1)]{}
\[1\][(\#1)]{}
[\_i]{}
[¶]{}
[\^()]{}
[\_[1,]{}]{}
[\_[1,]{}]{}
[u]{}
[m-1]{}
[m]{}
[\_]{}
[\_]{}
[\_]{}
[\_]{}
[K]{}
[C]{}
[H]{}
[Q]{}
[T]{}
\[1\][erv\_[\#1]{}]{}
=1.1
[**[Three perspectives on\
categorical symmetric Howe duality]{}**]{}\
author\
\
author\
[*W*aterloo, ON, Canada ]{}\
[[email protected]]{}
Acknowledgments {#acknowledgments .unnumbered}
---------------
J. G. was supported by NSERC and the University of Waterloo through an Undergraduate Student Research Award. B. W. is supported by an NSERC Discovery Grant. This research was supported in part by Perimeter Institute for Theoretical Physics. Research at Perimeter Institute is supported by the Government of Canada through the Department of Innovation, Science and Economic Development Canada and by the Province of Ontario through the Ministry of Research, Innovation and Science.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
author:
- 'L. Burderi'
- 'T. Di Salvo'
- 'A. Riggio'
- 'M.T. Menna'
- 'G. Lavagetto'
- 'A. Papitto'
- 'R. Iaria'
- 'N. R. Robba'
- 'L. Stella'
date: 'Received 2005 month day; accepted 2005 month day'
title: 'Timing an Accreting Millisecond Pulsar: Measuring the Accretion Torque in '
---
Introduction {#sect:intro}
============
The so-called recycling scenario links two different classes of astronomical objects, namely the millisecond radio pulsars (usually found in binary systems) and the Low Mass X-ray Binaries (hereafter LMXBs), or, at least, a subgroup of them. The leading idea of this scenario is the recycling process itself, during which an old, weakly magnetized, slowly spinning neutron star is accelerated by the accretion of matter and angular momentum from a (Keplerian) accretion disk down to spin periods in the millisecond range. In this way, at the end of the accretion phase, the neutron star rotates so fast that it is resurrected from the radio pulsar graveyard, allowing the radio pulsar phenomenon to occur again despite the weakness of the magnetic field.
Although this scenario was firstly proposed long time ago (see e.g. Bhattacharya & van den Heuvel 1991 for a review), the most embarassing problem was the absence of coherent pulsations in LMXBs. Only recently, the long seeked for millisecond coherent oscillations in LMXBs have been found, thanks to the capabilities (the right combination of high temporal resolution and large collecting area) of the RXTE satellite. In April 1998, a transient LMXB, , was discovered to harbour a millisecond pulsar ($P_{\rm spin}
\simeq 2.5$ ms) in a compact ($P_{\rm orb} \simeq 2$ h) binary system (Wijnands & van der Klis 1998; Chakrabarty & Morgan 1998). We now know seven accreting millisecond pulsars (Wijnands 2005; Morgan et al. 2005); all of them are X-ray transients in very compact systems (orbital period between 40 min and 4 h), the fastest of which ($P_{\rm spin} \simeq 1.7$ ms), , has been discovered in December 2004 (Galloway et al. 2005, hereafter G05). Timing techniques applied to data of various accreting millisecond pulsars, spanning the first few days of their outbursts, allowed an accurate determination of their main orbital parameters. However, only a few attemps have been made to determine the spin period derivative (Chakrabarty et al.2003; Galloway et al. 2002).
In this paper we apply an accurate timing technique to the fastest currently known accreting millisecond pulsar, , in the hope of constraining the predictions of different torque models with good quality experimental data. Our results indicate quite clearly that a net spin up occurs during the December 2004 outburst of (see also Falanga et al. 2005) and that the derived torque is in good agreement with that expected from matter accreting from a Keplerian disk.
The Timing Technique {#sect:data}
====================
In standard timing techniques (see e.g. Blandford & Teukolsky 1976) the predicted arrival time of a given pulse is computed using a first guess of the parameters of the system, and the difference between the experimental and predicted arrival times, namely the residuals, are fitted with a linear multiple regression of the differential corrections to the parameters. This means that the differential correction to orbital parameters, source position in the sky, spin frequency and its derivative, are computed simultaneously. This technique has the obvious advantage to give a self-consistent solution, where all the correlations in the covariance matrix of the system are fully taken into account. However, the convergency of the fit is not always guaranteed and – especially on long temporal baselines – convergence to secondary minima could occur.
On the other hand, if the orbital period is much shorter than the timescale on which the spin period derivative is expected to produce a significant effect, we can demonstrate that a different timing technique is more effective in determining the spin period derivative. This technique relies on the fact that the delays in the arrival times produced by the orbital corrections are effectively decoupled from those caused by the spin evolution. The technique proceedes as follows: in order to obtain the emission times, $t_{\rm em}$, the arrival times of all the events, $t_{\rm arr}$, are firstly reported to the Solar system barycenter adopting the best estimate of the source position in the sky, then corrected for the delays caused by the binary motion using the best estimate of the orbital parameters through the formula: $$\label{eq:corr}
t_{\rm em} \simeq t_{\rm arr} - x \sin \left[ \frac{2\pi}{\Porb}
\left(t_{\rm arr} - T^* \right)\right],$$ where $x = a \sin i /c$ is the projected semimajor axis in light seconds, and $T^*$ is the time of ascending node passage at the begining of the observation. In the following, for simplicity, we use $t$ instead of $t_{\rm
em}$. The differential of this expression with respect to the orbital parameters allows to calculate the uncertainties in the phases, $\sigma_{\phi\,\rm orb}$, caused by the uncertainties in the estimates of the orbital parameters. In a similar way, we have computed the uncertainties in the phase delays, $\sigma_{\phi\,\rm pos}$, caused by the uncertainties on the estimates of the source position in the sky. In this case we can estimate $\dot \nu$ fitting the measured phase variations, while the uncertainties in the adopted orbital parameters and source position will result in a “timing noise” of amplitude $\sigma_{\phi\,\rm par} =
(\sigma_{\phi\,\rm orb}^2 + \sigma_{\phi\,\rm pos}^2)^{1/2}$ (see Burderi et al. 2005 for details).
Observations and Data Analysis {#sect:obs}
==============================
was observed by RXTE between 2004 December 3 and 21. In this paper we report on the data between December 7 and 21 taken from a public ToO. We mainly use data from the PCA for timing anaysis and data from PCA and HEXTE for spectral analysis. The arrival times of all the events were converted to barycentric dynamical times at the solar system barycenter. The position adopted for the source was that of the proposed radio counterpart (which is compatible with that of the proposed optical counterpart, see Rupen et al.2004). We corrected the arrival times of all the events for the delays caused by the binary motion using eq. (\[eq:corr\]) with the orbital parameters given in G05.
Adopting the uncertainties in the estimates of the orbital parameters given in G05, the positional uncertainty of $0.04''$ radius reported by Rupen et al. (2004) we obtain: $\sigma_{\phi\, \rm orb} \la 0.01$, $\sigma_{\phi\,\rm pos} \la 0.006$ where we have maximized $\sin$ and $\cos$ functions with 1, and used $t-T_0 \la 7$ days. Therefore, we expect that the uncertainties in the orbital parameters and source position will cause a “timing noise” not greater than $\sigma_{\phi\,
\rm par} \times P_{\rm spin} \sim 0.02$ ms.
To compute phases of good statistical significance we epoch folded each interval of data in which the pulsation was significantly detected at the spin period given in G05 with respect to the same reference epoch, $T_0$, corresponding to the begining of our observations. The fractional part of the phase was obtained fitting each pulse profile with a sinusoid of fixed period. To compute the associated errors we combined the statistical errors derived from the fit, $\sigma_{\phi\,\rm stat}$, with the errors $\sigma_{\phi\, \rm par}$. In order to derive the differential correction to the spin frequency, $\Delta \nu_0$, and its derivative, $\dot \nu_0$, at the time $T_0$ we have to derive a functional form for the time dependence of the phase delays. We started from the following simple assumptions: i) the bolometric luminosity $L$ is a good tracer of the mass accretion rate $\dot M$ [*via*]{} the relation $L = \zeta
(GM/R) \dot M$, where $\zeta \leq 1$, and $G$, $M$, and $R$ are the gravitational constant and the neutron star mass and radius, respectively. ii) The matter accretes through a Keplerian disk truncated at the magnetospheric radius, $R_{\rm m} \propto \dot M^{-\alpha}$, by its interaction with the (dipolar) magnetic field of the neutron star. At $R_{\rm m}$ the matter is forced to corotate with the magnetic field of the neutron star and is funneled (at least in part) towards the rotating magnetic poles, thus causing the pulsed emission. For standard disk accretion $\alpha =2/7$. Indeed we considered two extreme cases, namely $\alpha = 2/7$ and $\alpha = 0$, since a location of $R_{\rm m}$ independent of $\dot M$ has been proposed (see, [*e.g.*]{}, Rappaport, Fregeau, and Spruit, 2004). iii) The matter accretes onto the neutron star its specific Keplerian angular momentum at $R_{\rm m}$, $\ell = (GMR_{\rm
m})^{1/2}$, thus causing a material torque $\tau_{\dot M} = \ell
\times \dot M$. A firm upper limit to this torque is given by the condition $\tau_{\dot M} \leq \ell_{\rm max} \times \dot M$, with $\ell_{\rm max} =
(GMR_{\rm CO})^{1/2}$, where $R_{\rm CO} = 1.50 \times 10^8 \; m^{1/3}
\nu^{-2/3}$ is the corotation radius (namely the radius at which the Keplerian frequency equals the spin frequency $\nu$ of the neutron star and beyond which accretion is centrifugally inhibited), and $m = M/{\rm M_\odot}$. We [*do not consider*]{} any form of threading of the accretion disk by the magnetic field of the neutron star (see e.g. Rappaport, Fregeau, and Spruit 2004 for a description of the magnetic threading), which implies that the only torque acting during accretion is $\tau_{\dot M}$.
---------------------------------------------------------------------------------------------------------------------------------------
Parameter G05 This work
------------------------------------------------------------------------ ------------------------- ------------------------------------
Projected semimajor axis, $\asini$ (lt-ms) $64.993(2)$ –
Orbital period, $\Porb$ (s) $8844.092(6)$ –
Epoch of ascending node passage, ${T^*}$ (MJD) $53345.1619258 –
(4)$
Eccentricity, $e$ $<2 \times 10^ –
{-4}$ (3 $\sigma$)
Spin frequency, $\nu_0$ (Hz) $598.89213064( $598.89213053(2)$
1)$
Spin frequency derivative, $\dot \nu_0$ (Hz/s) $< 8 \times –
10^{-13} $ (3 $\sigma$)
Spin frequency derivative, $\dot \nu_0$ (Hz/s) ($\alpha=0$) – $1.17(0.16)\times 10^{-12} \; \;$
Spin frequency derivative, $\dot \nu_0$ (Hz/s) ($\alpha=2/7$) – $1.11(0.16)\times 10^{-12} \; \; $
Spin frequency derivative, $\dot \nu_0$ (Hz/s) ($\dot \nu =$ constant) – $0.85(0.11)\times 10^{-12} \; \; $
Epoch of the spin period, ${T_0} $ (MJD) – 53346.184635
---------------------------------------------------------------------------------------------------------------------------------------
: Orbital and spin parameters of .[]{data-label="table:1"}
\
Errors are given at $1 \sigma$ confidence level.
In these hypothesis the spin frequency derivative is $\dot \nu = \ell \,
\dot M/(2\pi \, I) $, where $I$ is the moment of inertia of the neutron star and we have neglected any variation of $I$ caused by accretion. If $\dot M = \dot M(t)$, we have $\dot \nu(t) = (2\pi
\,I)^{-1} \ell_0 \dot M_0 (\dot M(t)/\dot M_0)^{1-\alpha/2}$, where $\ell_0 = (GMR_{\rm m \, 0})^{1/2}$, and $R_{\rm m \,0}$ and $\dot M_0$ are $R_{\rm m}$ and $\dot M$ at $t = T_0$, respectively. For the $\alpha = 0$ case we assumed $\ell_0 = \ell_{\rm max}$.
Since we assumed $\dot M(t) \propto L(t)$, to determine the temporal dependence of $\dot M(t)$ we studied the energy spectra of the source for each continuous interval of data combining PCA and HEXTE data. All the spectra are well fitted with a model consisting of a power law with an exponential cutoff plus thermal emission from a Keplerian accretion disk modified by photoelectric absorption and a Gaussian iron line. In order to derive $L(t)$ for each spectrum we made the simple assumption $L(t) \propto F_{(3-150)}(t)$, which is the unabsorbed flux in the RXTE PCA plus HEXTE energy band $(3-150$ keV). A good fit of $F_{(3-150)}(t)$ [*vs*]{} $t$ between December 7 and 14 ($\Delta t_{\rm obs} \sim 7.3$ days) is given by the expression $F_{(3-150)}(t) = F_{(3-150)} \times [1- (t-T_0)/t_B]$ with $t_B=8.4 \pm 0.1 $ days, where $F_{(3-150)}$ is the unabsorbed flux at $t=T_0$. Therefore we have $\dot \nu(t) = \dot \nu_0 \times
[1-(t-T_0)/t_B]^{1-\alpha/2}$, where the spin frequency derivative at $t=T_0$ is $\dot \nu_0 = (2\pi \,I)^{-1} \, \ell_0 \, \dot M_0$. We have therefore fitted the phase delays with the function: $$\phi = - \phi_0 - \Delta \nu_0\, (t - T_0) -
\frac{1}{2} \dot \nu_{0} (t - T_0)^2 \times
\left[ 1 - \left( \frac{2-\alpha}{6} \right) \times
\frac{(t-T_0)}{t_B} \right] .
\label{eq:phi}$$ Using the best fit value for $\Delta \nu_0$ we computed the improved spin frequency estimate and repeated the same procedure described at the begining of this paragraph, folding at the new estimate of the spin period. The new phases were fitted with eq. (\[eq:phi\]). In this case, $\Delta \nu_0$ was fully compatible with zero. These phases are plotted versus time in Figure \[fig1\] (upper panel) together with the residuals in units of $\sigma$ with respect to eq. (\[eq:phi\]) (lower panel). The best fit estimates of $\nu_0$ and $\dot \nu_0$ are reported in Table 1 for three values of $\alpha$, namely $\alpha = 0$ which correspond to a location of $R_{\rm m}$ independent of the accretion rate (cfr. the model of Rappaport Fregeau, and Spruit 2004 in which $R_{\rm m}
= R_{\rm CO}$ for any $\dot M$), the standard case $\alpha = 2/7$ which corresponds to $R_{\rm m}$ proportional to the Alfvén radius, and $\alpha = 2$ which has been given for comparison purposes and corresponds to a parabolic trend, expected in the case of constant $\dot M$. The statistics is not good enough to distinguish between these three possibilities.
From the best-fit value of the spin frequency derivative $\dot \nu_0$ we can compute the mass accretion rate at $t = T_0$ through the formula: $
\dot M_{-10} = 5.9 \times \dot \nu_{-13}\, I_{45}\,
m^{-2/3} (R_{\rm CO}/R_{\rm m \, 0})^{1/2},
$ where $\dot M_{-10}$ is $\dot M_0$ in units of $10^{-10}\, M_\odot$ yr$^{-1}$, $\dot \nu_{-13}$ is $\dot \nu_0$ in units of $10^{-13}$ s$^{-2}$, and $I_{45}$ is $I$ in units of $10^{45}$ g cm$^2$. In the following we will adopt the FPS equation of state for the neutron star matter for $m = 1.4$ and the spin frequency of which gives $I_{45}= 1.29$ and $R= 1.14 \times 10^6$ cm (see e.g. Cook, Shapiro & Teukolsky 1994). In order to compare the experimental estimate of $\dot M_0$ with the observed X-ray luminosity, we have to derive the bolometric luminosity $L(t)$ from the observed flux $F_{(3-150)}(t)$. To this end we consider the spectral shape at $t=T_0$ in more detail.
The power law is the dominant spectral component. This component presumably originates in an atmosphere of small optical depth just above each polar cap (see e.g. Poutanen & Gierlinski 2003; Gierlinski & Poutanen 2005), thus we neglect, to first order, any effect of the inclination of the emitting region with respect to the observer. On the other hand, we observe a single-peaked pulse profile, which means that we only see the emission from one of the polar caps (e.g. Kulkarni & Romanova 2005). We have therefore multiplied by 2 the unabsorbed flux of the power law in order to take into account the emission of the unseen polar cap. Assuming isotropic emission, we computed $L_{{\rm PL}} \simeq 2 F_{{\rm PL} \; (0,\infty)}
\times 4 \pi d^2 = 1.5^{+0.4}_{-0.3} \times 10^{37} \; d_{\rm 5\,kpc}^2$ erg/s, where $d_{\rm 5\, kpc}$ is the source distance in units of $5$ kpc. The uncertainty on the luminosity has been evaluated propagating the uncertainties on the spectral parameters threated as they were independent on each other.
The second component is the thermal emission from a Shakura-Sunyaev accretion disk. The fraction of the total luminosity that is emitted by the disc is given by the ratio: $0.5 R/R_{\rm m\,0} = 0.34$. In this hypothesis $L_{{\rm BB\; 0}} = 0.39 / (1-0.39) \times L_{{\rm PL} \; 0}
= 9.6 \times 10^{36} \, d_{\rm 5\,kpc}^2$ erg/s. The total bolometric luminosity is therefore $L_0 = 2.46^{+0.94}_{-0.29}
\times 10^{37} \, d_{\rm 5\,kpc}^2$ erg/s. If we compare this luminosity with the mass accretion rate inferred from the timing analysis, we obtain an estimate of the source distance, which is: $d = 9.47_{-2.1}^{+1.2}$ kpc.
Conclusions {#sect:conclusion}
===========
We have analysed RXTE data of the fastest known accreting millisecond pulsar, , during the period $7 - 14$ December, 2004. We report a revised estimate of the spin period and the spin period derivative. The source shows a strong spin-up, which indicates a mass accretion rate of about $8.5 \times 10^{-9} \, M_\odot$ yr$^{-1}$. Comparing the bolometric luminosity of the source as derived from the X-ray spectrum with the mass accretion rate of the source as derived from the timing, we find a good agreement if we place the source at a quite large distance between 7 and 10 kpc. Note that 10 kpc is close to the outer edge of our Galaxy in the direction of . Another possibility is that part of the luminosity of the system is not observed because emitted in other energy bands or because of occultation effects (which may be favoured if indeed the source is highly inclined).
This work was partially supported by the Ministero della Istruzione, della Università e della Ricerca (MIUR).
[99]{} Bhattacharya D., van den Heuvel E. P. J., 1991, Physics Report, 203, 1 Blandford, R., Teukolsky, S. A., 1976, ApJ, 205, 580 Burderi L., et al., 2005, , submitted Chakrabarty, D., Morgan, E. H. 1998, Nature, 394, 346 Chakrabarty, D., et al. 2003, Nature, 424, 42 Cook, G. B., Shapiro, S. L., & Teukolsky, S. A. 1994, ApJ, 424, 823 Falanga M., et al., 2005, , in press Galloway, D. K., Chakrabarty, D., Morgan, E. H., Remillard, R. A., 2002, ApJ, 576, L137 Galloway, D. K., Markwardt, C. B., Morgan, E. H., Chakrabarty, D., Strohmayer, T. E., 2005, ApJ, 622, L45 Gierlinski, M., & Poutanen, J. 2005, MNRAS, 359, 1261 Kulkarni, A. K., & Romanova, M. M. 2005, ApJ, in press Morgan E., Kaaret P., Vanderspeck R., 2005, ATel n. 523 Nowak, M. A., et al. 2004, ATEL n. 369 Poutanen, J., & Gierlinski, M. 2003, MNRAS, 343, 1301 Rappaport, S. A., Fregeau, J. M., Spruit, H., 2004, ApJ, 606, 436 Rupen, M. P., Dhawan, V., Mioduszewski, A. J., 2004, ATel n. 364 Wijnands, R. 2005, to appear in Nova Science Publishers (NY) volume “Pulsars New Research”, astro-ph/0501264 Wijnands, R., van der Klis, M. 1998, Nature, 394, 344
\[lastpage\]
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'We consider a mixed stochastic differential equation ${\ensuremath{\operatorname{d}\!{X_t}}}=a(t,X_t){\ensuremath{\operatorname{d}\!{t}}}+b(t,X_t){\ensuremath{\operatorname{d}\!{W_t}}}+c(t,X_t){\ensuremath{\operatorname{d}\!{B^H_t}}}$ driven by independent multidimensional Wiener process and fractional Brownian motion. Under Hörmander type conditions we show that the distribution of $X_t$ possesses a density with respect to the Lebesgue measure.'
address:
- 'Department of Probability Theory, Statistics and Actuarial Mathematics,Taras Shevchenko National University of Kyiv,Volodymyrska 64, Kyiv 01601, Ukraine andDepartment of Mathematical Economics IIMathematical Institute of the Manhheim UniversityA5, 6D-68131 Mannheim'
- ' Department of Probability Theory, Statistics and Actuarial Mathematics,Taras Shevchenko National University of Kyiv,Volodymyrska 64, Kyiv 01601, Ukraine'
author:
- Taras Shalaiko
- Georgiy Shevchenko
title: Existence of density for solutions of mixed stochastic equations
---
Introduction
============
In this paper we study a so-called mixed stochastic differential equation (SDE) in ${\mathbb R}^d$ $$X_t=X_0+\int_0^t a(s,X_s){\ensuremath{\operatorname{d}\!{s}}}+\int_0^t b(s,X_s){\ensuremath{\operatorname{d}\!{W_s}}}+\int_0^t c(s,X_s){\ensuremath{\operatorname{d}\!{B^H_s}}}
\label{eqIntr}$$ driven by a multidimensional standard Wiener process and a multidimensional fractional Brownian motion (fBm) with Hurst parameter $H\in(1/2,1)$ (see next section for precise definitions). Recently such equations gained a lot of attention thanks to their modeling features. There is already a large literature devoted to them; the few papers we cite here give an extensive overview of existing results. The unique solvability result in the form suitable for our needs is obtained in the paper [@Delay]; although the result is formulated there for equations with delay, it is a fortiori valid for usual equations. The paper [@Integr] contains useful estimates of the solution and results on its integrability. Finally, we mention the paper [@ShSh], where the Malliavin differentiability of the solution is obtained.
The main aim of this article is to provide conditions under which the solution to has a density with respect to the Lebesgue measure. For Itô SDEs, such issues were addressed by many authors, see [@Nualart] and references therein. Existence and regularity of density for SDEs driven by fBm we proved in [@BauHai; @NourdinSimon; @NualSau] under Hörmander type conditions. The recent paper [@Tindel] contains a generalization of these results to equations driven by Gaussian rough paths, in particular, it allows to deduce the existence of a smooth density of the solution to with Stratonovich integral with respect to the Wiener process. However, the machinery used in that article is quite sophisticated, and here we use a more direct approach.
The paper is organised as follows. In Section 2 we introduce our notation, describe the main object and briefly discuss Malliavin calculus of variations for fractional Brownian motion. In Section 3, we prove that the distribution of the solution $X_t, t>0$ possesses density w.r.t. Lebesgue measure under a simplified version of the Hörmander condition. Section 4 contains the result on existence and smoothness of the density under a strong version of the Hörmander condition. The Appendix contains some technical lemmas and the Norris lemma for a mixed SDE.
Preliminaries
=============
Definitions and notation
------------------------
Throughout the paper, ${\left\lvert\cdot\right\rvert}$ will denote the absolute value of a number, the Euclidean norm of a vector, and the operator norm of a matrix. ${\langle}\cdot, \cdot{\rangle}$ stays for the usual scalar product in the Euclidean space. We will use the symbol $C$ to denote a generic constant, whose value is not important and may change from one line to another. We will write a subscript if a constant is relevant or if its value depends on some parameters.
For a matrix $A=(a_{i,j})$ of arbitrary size, we denote by $a_i$ its $i$-th row and by $a_{\cdot,j}$ its $j$-th column.
The classes of continuous and $\theta$-Hölder continuous functions on $[a,b]$ will be denoted respectively by $C[a,b]$ and $C^\theta[a,b]$. For a function $f\colon [a,b]\to{\mathbb R}$ denote by ${\left\lVertf\right\rVert}_{\infty,[a,b]}$ its supremum norm and by $${\left\lVertf\right\rVert}_{\theta,[a,b]}=\sup_{a\leq s<t\leq b}\frac{|f(t)-f(s)|}{|t-s|^\theta}$$ its $\theta$-Hölder seminorm. If there is no ambiguity, we will use the notation ${\left\lVertf\right\rVert}_\infty$ and ${\left\lVertf\right\rVert}_\theta$.
Finally, for a function $h\in C({\mathbb R}^d)$ denote by $\partial_x h = (\frac{\partial}{\partial x_1}h,\dots, \frac{\partial}{\partial x_d}h)$ its gradient and by $\partial^2_{xx} h = (\frac{\partial^2}{\partial x_i\partial x_j}h)_{i,j=1,\dots,d}$ its second derivative matrix.
Main equation and assumptions
-----------------------------
For a fixed time horizon $T>0$, let $\{\Omega, \mathcal F, \mathbf{F}=(\mathcal F_t)_{t\in [0,T]}, \mathsf P\}$ be a standard stochastic basis. Equation is driven by two independent sources of randomness: an $m$-dimensional $\mathbf{F}$-Wiener process $\{W_t=(W^1_t,\ldots,W^m_t), t\in [0,T]\}$ and an $l$-dimensional fBm $\{B^H_t=(B^{H,1}_t,\ldots,B^{H,l}_t), t\ge [0,T]\}$ with Hurst index $H\in(1/2,1)$, i.e. a centered Gaussian process having the covariance $${\mathsf{E}\left[\,B^{H,i}_tB^{H,j}_s\,\right]}=\frac{\delta_{i,j}}{2}(t^{2H}+s^{2H}-|t-s|^{2H}).$$ It is well known that the fBm $B^H$ has a modification with $\gamma$-Hölder continuous path for any $\gamma<H$, in the following we will assume that the process itself is Hölder continuous.
Equation is understood as a system of SDEs on $[0,T]$ $$\begin{aligned}
X^i_t=X^i_0 +\int_0^ ta_i(s,X_s){\ensuremath{\operatorname{d}\!{s}}}+\sum_{j=1}^m\int_0^tb_{i,j}(s,X_s){\ensuremath{\operatorname{d}\!{W}}}^j_s +\sum_{k=1}^l\int_0^tc_{i,k}(s,X_s){\ensuremath{\operatorname{d}\!{B}}}^{H,k}_s,
\end{aligned}$$ $i=1,\ldots,d$, with a non-random initial condition $X_0\in{\mathbb R}^d$. In this equation, the integral w.r.t. $W$ is understood in a usual Itô sense, the one w.r.t. $B^H$ is understood in a pathwise sense, as Young integral. More information on its definition and properties can be found in [@friz-victoir].
The coefficients $a_i,b_{i,j},c_{i,k}\colon [0,T]\times{\mathbb R}^d\to{\mathbb R}^d$, $i=1,\ldots,d,j=1,\ldots, m,k=1,\ldots, l$ are assumed to satisfy the following conditions.
- for all $t\in[0,T]$ $a(t,\cdot),b(t,\cdot)\in C^1({\mathbb R}^d)$, $c(t,\cdot)\in C^2({\mathbb R}^d)$;
- for all $t\in[0,T]$, $x\in {\mathbb R}^d$ $${\left\lverta(t,x)\right\rvert} + {\left\lvertb(t,x)\right\rvert} + {\left\lvertc(t,x)\right\rvert}\le C(1+{\left\lvertx\right\rvert});$$
- for all $t\in[0,T]$, $x\in {\mathbb R}^d$ ${\left\lvert\partial_x c(t,x)\right\rvert}\le C $;
- there exists $\beta>0$ such that for all $t,s\in[0,T]$, $x\in{\mathbb R}^d$ $$|c(t,x)-c(s,x)|\leq C|t-s|^\beta(1+|x|),\quad |\partial_x c(t,x)-\partial_x c(s,x)|\leq C|t-s|^\beta.$$
The continuous differentiability implies that $a,b,\partial_x c$ are locally Lipschitz continuous. Therefore, by [@Delay Theorem 4.1], equation has a unique solution which is Hölder continuous of any order $\theta\in(0,1/2)$.
Ad hoc Malliavin calculus
-------------------------
Here we summarize some facts from the Malliavin calculus of variations, see [@Nualart] for a deeper exposition. Denote by $S[0,T]$ the set the of step functions of the form $f(t) = \sum_{k=1}^{n} c_k {\mathbf{1}_{[a_k,b_k)}}(t)$ defined on $[0,T]$. Let $L^2_H[0,T]$ denote the separable Hilbert space obtained by completing $S[0,T]$ w.r.t. the scalar product $$\begin{gathered}
{\langle}f,g{\rangle}_{L^2_H[0,T]}=\int_0^T\int_0^T f(t)g(s)\phi(t,s){\ensuremath{\operatorname{d}\!{t}}}{\ensuremath{\operatorname{d}\!{s}}},\end{gathered}$$ where $\phi(t,s)=H(2H-1)|t-s|^{2H-2}$.
Consider the product space $$\mathfrak H = \left(L^2_H [0,T]\right)^{ l}\times \left(L^2[0,T]\right)^{m}.$$ It is also a separable Hilbert space with a scalar product $${\langle}f,g{\rangle}_{\mathfrak H} = \sum_{i=1}^l {\langle}f_i,g_i{\rangle}_{L^2_H[0,T]} + \sum_{i=l+1}^{l+m} {\langle}f_i,g_i{\rangle}_{L^2[0,T]}.$$ The map $${\mathcal I}\colon ({\mathbf{1}_{[0,t_1)}},\dots,{\mathbf{1}_{[0,t_l)}},{\mathbf{1}_{[0,s_1)}},\dots,{\mathbf{1}_{[0,s_m)}})\mapsto (B^{H,1}_{t_1},\dots, B^{H,l}_{t_l},W^1_{s_1},\dots, W^m_{s_m})$$ can be extended by linearity to $S[0,T]^{l+m}$. It appears that for $f,g\in S[0,T]^{l+m}$ $${\mathsf{E}\left[\,{\langle}{\mathcal I}(f),{\mathcal I}(g){\rangle}\,\right]} = {\langle}f,g{\rangle}_{\mathfrak H},$$ so ${\mathcal I}$ can be extended to an isometry between $\mathfrak H$ and a subspace of $L^2(\Omega;{\mathbb R}^{m+l})$.
For $\xi = F({\mathcal I}(f_1),\dots ,{\mathcal I}(f_n))$, where $f_1,\dots,f_n\in \mathfrak H$ and $f_i=(f_{i,1},\ldots,f_{i,m+l})$, $i=1,\ldots,n$, $F\colon {\mathbb R}^{n(m+l)} \to {\mathbb R}$ is a continuously differentiable finitely supported function, define the Malliavin derivative $\mathsf D \xi$ as an element of $\mathfrak H$, whose $j$-th coordinate equal to $$\sum_{i=1}^{n} \partial_{(i-1)(l+m)+j} F ({\mathcal I}(f_1),\dots, {\mathcal I}(f_n)) f_{i,j}, j=1,\ldots,l+m.$$ Denote for $p\ge 1$ by $\mathbb D^{1,p}$ the closure of the space of smooth cylindrical random variables with respect to the norm $${\left\lVert\xi\right\rVert}^p_{\mathbb D^{1,p}}={\mathsf{E}\left[\,|\xi|^{p}+{\left\lVert\operatorname{\mathsf{D}}{\xi}\right\rVert}_{\mathfrak H}\,\right]}^{1/p}.$$ $\operatorname{\mathsf{D}}$ is closable in this space and its closure will be denoted likewise. Finally, the Malliavin derivative is a (possibly, generalized) function from $[0,T]$ to ${\mathbb R}^{l+m}$, so we can introduce the notation $$\operatorname{\mathsf{D}}\xi = {\left\{\operatorname{\mathsf{D}}_t\xi = \big(\operatorname{\mathsf{D}}^{H,1}_t \xi,\dots, \operatorname{\mathsf{D}}^{H,l}_t\xi,\operatorname{\mathsf{D}}^{W,1}_t\xi,\dots, \operatorname{\mathsf{D}}^{W,m}_t\xi\big),\ t\in[0,T]\right\}}.$$ We say that $\xi\in \mathbb{D}^{1,p}_{\mathit{loc}}$ if exists a sequence $\{\xi_n(\omega),\Omega_n\}_{n\geq 1}$ such that $\Omega_n\subset \Omega_{n+1}$ for $n\geq 1$, $P\left(\Omega\setminus \left(\bigcup_{n\ge 1}\Omega_n\right)\right)=0$; $\xi_n\in \mathbb D^{1,p}$ and $\xi|_{\Omega_n} = \xi_n|_{\Omega_n}$ for all $n\geq 1$.
For the reader convenience we state here the theorem concerning the Mallivian differentiability of the solution to in the case of SDE with non-homogeneous coefficients. The proof is similar to that of [@ShSh Theorem 2]
\[MalliavinDiff\] Suppose that coefficients $a,b,c$ of satisfy the assumptions
- for all $t\in[0,T]$ $a(t,\cdot),b(t,\cdot)\in C^1({\mathbb R}^d)$, $c(t,\cdot)\in C^2({\mathbb R}^d)$;
- $a,b,\partial_x a, \partial_x b, \partial_x c, \partial^2_{xx}c$ are bounded;
- there exists $\beta>0$ such that for all $t,s\in[0,T]$, $x\in{\mathbb R}^d$ $$|c(t,x)-c(s,x)|\leq C|t-s|^\beta(1+|x|),\quad |\partial_x c(t,x)-\partial_x c(s,x)|\leq C|t-s|^\beta.$$
Then $X_t\in\bigcap_{p\geq 1} \mathbb D^{1,p}$.
Existence of density under simplified Hörmander condition
=========================================================
In this section we prove that a solution to possesses density of a distribution under a quite strong condition, which we call a simplified Hörmander condition. More precisely, we will assume in this section that the coefficients of satisfy $$\label{simpHormander}
\mathsf{span}\{c_{\cdot,k}(0,X_0),b_{\cdot,j}(0,X_0)\mid 1\leq k\leq l,1\leq j\leq m\,\}=
{\mathbb R}^d.$$
The first step to establish the existence of density is to show the (local) Malliavin differentiability of the solution to .
If the coefficients of satisfy the assumptions *A1–A4*, then $X_t\in \bigcap_{p\geq 1} \mathbb{D}^{1,p}_{loc} $.
Define $\Omega_n=\{\omega: {\left\lVertX_\cdot (\omega)\right\rVert}_{\infty,[0,t]}<n\}$, $n\ge 1$. Obviously, $\Omega_n\subset \Omega_{n+1}, n\geq 1$ and, since ${\left\lVertX_\cdot (\omega)\right\rVert}_{\infty,[0,t]}<\infty$ a.s., $\bigcup_{n\geq 1 }\Omega_n=\Omega$. Consider a smooth function $\psi=\psi(x),\,x\in {\mathbb R}$ such that
- for all $x\in {\mathbb R}$ $0\leq \psi(x)\leq 1$;
- $\psi(x)=1,x\in[-1,1]$;
- $\psi(x)=0,\, x\notin[-2,2]$.
For $n\ge 1$, put $\Psi_n=\Psi_n(x_1,\ldots,x_d)=(\int_0^{x_1}\psi(y/n){\ensuremath{\operatorname{d}\!{y}}},\ldots,\int_0^{x_d}\psi(y/n){\ensuremath{\operatorname{d}\!{y}}})$, define $d^{(n)}(s,x) = d(t,\Psi_n(x))$, $d\in {\left\{a,b,c\right\}}$, and let $X^{(n)}$ solve $$X^{(n)}_t=X_0 +\int_0^ ta^{(n)}(s,X^{(n)}_s){\ensuremath{\operatorname{d}\!{s}}}+\int_0^tb^{(n)}(s,X^{(n)}_s){\ensuremath{\operatorname{d}\!{W}}}_s +\int_0^t c^{(n)}(s,X^{(n)}_s){\ensuremath{\operatorname{d}\!{B}}}^{H}_s.
$$ Since the functions $a_n, b_n, c_n$ satisfy assumptions B1–B3, in view of Theorem \[MalliavinDiff\], $X^{(n)}_t\in \bigcap_{p\geq 1} \mathbb D^{1,p}$. It is not hard to see that that $X^{(n)}_t(\omega)=X_t(\omega)$ for $\omega \in \Omega_n$, which concludes the proof.
Now we are to prove the main result of this section.
\[thTOYHorm\] Suppose that the coefficients of satisfy assumptions *A1–A4* and the simplified Hörmander condition . Then for all $t>0$ the law of $X_t$ is absolutely continuous with respect to the Lebesgue measure in ${\mathbb R}^d$.
By the classical condition for existence of density (see e.g. [@Nualart Theorem 2.1.2]) and thanks to the previous theorem, it is enough to verify the non-degeneracy of the Mallivain covariation matrix $M(t)=(M_{i,j}(t))_{i,j=1,\ldots,d}$ with $M_{i,j}(t)={\langle}\operatorname{\mathsf{D}}X^i_t,\operatorname{\mathsf{D}}X^i_t{\rangle}_{\mathcal H}$. Define the matrix-valued process $J_{t,0}=(J_{t,0}(i,j))_{i,j=1,\ldots,d}$ as the solution to $$\label{eqJacobian}
\begin{gathered}
J_{t,0}(i,j)=\delta_{i,j}+\sum_{r=1}^d\bigg[\int_0^t \frac{\partial a_i}{\partial x_r}(s,X_s) J_{s,0}(r,j){\ensuremath{\operatorname{d}\!{s}}} \\ +\sum_{k=1}^m \int_0^t \frac{\partial b_{i,k}}{\partial x_r}(s,X_s)J_{s,0}(r,j){\ensuremath{\operatorname{d}\!{W}}}^{k}_s
+ \sum_{q=1}^l \int_0^t \frac{\partial c_{i,q}}{\partial x_r}(s,X_s)J_{s,0}(r,j) {\ensuremath{\operatorname{d}\!{B}}}^{H,q}_s\bigg].
\end{gathered}$$ where $\delta_{i,j} = {\mathbf{1}_{i=j}}$ is the Kronecker delta. The system above is linear, hence, possesses a unique solution. In view of Lemma \[lemma1\], $J_{t,0}$ is non-degenerate; denoting $J_{t,s}=J_{t,0}J^{-1}_{s,0}$ and applying Lemma \[lemma2\] one can write $$\begin{gathered}
M(t)=\sum_{k=1}^m \int_0^t (J_{\newcommand{\Rho}{\mathrm{P}}t,s}b_{\cdot,k}(s,X_s))(J_{t,s}b_{\cdot,k}(s,X_s))'{\ensuremath{\operatorname{d}\!{s}}}\\
+
\sum_{q=1}^l \int_0^t\int_0^t \varphi_H(s,u)(J_{t,s}c_{\cdot,q}(s,X_s))(J_{t,u}c_{\cdot,q}(u,X_u))'{\ensuremath{\operatorname{d}\!{s}}}{\ensuremath{\operatorname{d}\!{u}}}=J_{t,0}C_t J'_{t,0},
\end{gathered}$$ where $$\begin{gathered}
C_t=\sum_{k=1}^m \int_0^t (J_{s\newcommand{\Rho}{\mathrm{P}},0}^{-1}b_{\cdot,k}(s,X_s))(J_{s,0}^{-1}b_{\cdot,k}(s,X_s))'{\ensuremath{\operatorname{d}\!{s}}} \\
+ \sum_{q=1}^l \int_0^t\int_0^t \varphi_H(s,u)(J_{s,0}^{-1}c_{\cdot,q}(s,X_s))(J^{-1}_{u,0}c_{\cdot,q}(u,X_u))'{\ensuremath{\operatorname{d}\!{s}}}{\ensuremath{\operatorname{d}\!{u}}}.
\end{gathered}$$ Again, due to the invertibility of $J_{t,0}$, $M_t$ is invertible if and only if so is $C_t$. Assuming the contrary, there exists a non-zero vector $v\in {\mathbb R}^d$ such that $v'C_tv=0$. Write $$v'C_tv=\sum_{k=1}^m {\left\lVert{\langle}J_{\cdot,0}b_{\cdot,k}(\cdot,X_\cdot),v {\rangle}\right\rVert}^2_{L^2[0,t]}+\sum_{q=1}^l {\left\lVert{\langle}J_{\cdot,0}c_{\cdot,q}(\cdot,X_\cdot),v {\rangle}\right\rVert}^2_{L^2_H[0,t]}.$$ Since the functions $$\begin{gathered}
s\mapsto {\langle}J^{-1}_{s,0}b_{\cdot,k}(s,X_s,v {\rangle},\, k=1,\ldots,m,\\
s\mapsto {\langle}J^{-1}_{s,0}c_{\cdot,q}(s,X_s),v{\rangle},\, q=1,\ldots,l\end{gathered}$$ are continuous, they must be equal zero for all $s\in [0,t]$. For $s=0$ we get $$\begin{aligned}
&\sum_{i=1}^d b_{i,k}(0,X_0)v_i=0, k=1,\ldots,m;\\
&\sum_{i=1}^d c_{i,q}(0,X_0)v_i=0, q=1,\ldots,l.\end{aligned}$$ This, however, contradicts the assumption . Consequently, $M_t$ is invertible, as required.
Existence of density under strong Hörmander condition
=====================================================
In this section we consider a homogeneous version of : $$\begin{gathered}
X_t=X_0+\int_0^t a(X_s){\ensuremath{\operatorname{d}\!{s}}}+\int_0^t b(X_s){\ensuremath{\operatorname{d}\!{W_s}}}+\int_0^t c(X_s){\ensuremath{\operatorname{d}\!{B^H_s}}}.\label{eqHom}\end{gathered}$$ In this section we assume that Hurst index $H\in (1/2,2/3)$, and some $\theta\in((H-1/2)/(3-4H),1/2)$ is fixed. The role of the restriction $\theta>(H-1/2)/(3-4H)$ will become clear in the proof of the Norris lemma for (Lemma \[Chuck\]). Now we just remark that the expression $(H-1/2)/(3-4H)$ is increasing for $H\in(1/2,3/4)$ and is equal to $1/2$ for $H=2/3$, so the upper bound $H<2/3$ arises naturally.
We impose the following condition on the coefficients of (\[eqHom\]):
- $a,b,c\in C^\infty_b({\mathbb R}^d)$ with all derivatives bounded.
Under this assumption the solution is infinitely differentiable in the Malliavin sense: $X_t\in \bigcap_{k,p=1}^\infty \mathbb D^{k,p}=\mathbb D^{\infty}$, which can be shown similarly to its differentiability under B1–B3.
The aim of this section is to investigate the existence of a density and properties of this density of a distribution of $X_t$ under the strong Hörmander condition, which reads as follows.
Set $V_0=a$, $V_j(\cdot)=b_{\cdot, j}(\cdot)$ for $j=1,\ldots,m$ and $V_{j+m}(\cdot)=c_{\cdot,j}(\cdot)$ for $j=1,\ldots,l$. Using the Lie bracket $[\cdot,\cdot]$, define the set $$\Upsilon_k=\{[V_{i_1},\ldots,[V_{i_{k-1}},V_{i_k}]\ldots], (i_1,\ldots,i_k)\in\{1,\ldots,d\}^k\}.$$ It is said that the vector field $\Upsilon_0=\{V_{j}\}_{j=1,\ldots,m+l}$ satisfies the [Hörmander condition]{} at the point $X_0$, if for some positive integer $n_0$ one has $$\label{Hormander}
\mathsf{span}{\left\{ V(X_0), V\in \bigcup_{k=1}^{n_0} \Upsilon_k\right\}}={\mathbb R}^d.$$ The main result of this section is the following theorem.
Assume that coefficients of (\[eqHom\]) satisfy assumption *C1* and the Hörmander condition . Then the law of $X_t$ for all $t>0$ possesses a smooth density with respect to the Lebesgue measure in ${\mathbb R}^d.$
Using the usual condition for existence of a smooth density (see e.g. [@Nualart Theorem 2.1.4]) and taking into account that all moments of the Jacobian $J_{t,s}$ are finite, it is enough to show that the matrix inverse to the reduced Malliavin covariance matrix of $X_t$ possesses all moments.
Recall from Theorem \[thTOYHorm\] that the reduced Malliavin covariance matrix of the solution to (\[eqHom\]) can be written as $$\begin{gathered}
C(t)=\sum_{k=1}^m \int_0^t(J_{s,0}^{-1} b_{\cdot,k}(X_s))(J_{s,0}^{-1}b_{\cdot,k}(X_s))'{\ensuremath{\operatorname{d}\!{s}}}\\
+ \sum_{q=1}^l \int_0^t\int_0^t \varphi_H(s,u)(J_{s,0}^{-1}c_{\cdot,q}(X_s))(J_{u,0}^{-1}c_{\cdot,q}(X_u))'{\ensuremath{\operatorname{d}\!{s}}}{\ensuremath{\operatorname{d}\!{u}}}.\end{gathered}$$ To simplify the notation, we assume from now that $t=1$. We are to prove that ${\mathsf{E}\left[\,|\det C_t|^{-p}\,\right]}<\infty$ for all $p\geq 1$. Due to [@Nualart Lemma 2.3.1] it suffices to prove that the entries of $C_t$ possess all moments and for any $p\geq 2$ there exists $C_p$ such that for all $\varepsilon>0$ it holds $$\sup_{{\left\lVertv\right\rVert}=1}{\mathsf P\left \{{\langle}v, C_1 v{\rangle}\leq \varepsilon\right\}}\leq C_p \varepsilon^p.$$ Write $$\begin{gathered}
{\langle}v,C_1 v{\rangle}=\sum_{k=1}^m {\left\lVert{\langle}J^{-1}_{\cdot,0}b_{\cdot,k}(X_\cdot),v {\rangle}\right\rVert}^2_{L^2[0,1]}+\sum_{q=1}^l
{\left\lVert{\langle}J^{-1}_{\cdot,0}c_{\cdot,q}(X_\cdot),v{\rangle}\right\rVert}_{L^2_H[0,1]}.\end{gathered}$$ It is well known that ${\left\lVertf\right\rVert}_{L^2_H[0,1]}\le {\left\lVertf\right\rVert}_{L^2[0,1]}$. Therefore, $${\langle}v,C_1 v{\rangle}\geq C\sum_{k=1}^{m+l}{\left\lVertG_k\right\rVert}_{L^2_H[0,1]}, \text{ where \ }G_k={\langle}J^{-1}_{\cdot,0} V_{k}(X_\cdot),v {\rangle}.$$ Applying [@BauHai Lemma 4.4] we get that $${\langle}v,C_1 v{\rangle}\ge C\sum_{k=1}^{m+l}\frac{{\left\lVertG_k\right\rVert}^{2(3+1/\theta)}_\infty}{{\left\lVertG_k\right\rVert}^{2(2+1/\theta)}_\theta}$$ for $\theta>H-1/2$. Thus, $${\mathsf P\left \{{\langle}v,C_1 v{\rangle}\leq \varepsilon\right\}}\leq {\mathsf P\left \{C\sum_{k=1}^{m+l}\frac{{\left\lVertG_k\right\rVert}^{2(3+1/\theta)}_\infty}{{\left\lVertG_k\right\rVert}^{2(2+1/\theta)}_\theta}\leq \varepsilon\right\}}.$$ From [@BauHai Lemma 4.5] and Theorem \[Chuck\] we obtain the following estimate $${\mathsf P\left \{C\sum_{k=1}^{m+l}\frac{{\left\lVertG_k\right\rVert}^{2(3+1/\theta)}_\infty}{{\left\lVertG_k\right\rVert}^{2(2+1/\theta)}_\theta}\leq \varepsilon\right\}}\leq C\varepsilon^p+\min_{k=1,\ldots,m+l}{\mathsf P\left \{{\left\lVert{\langle}v, J_{\cdot,0}V_k(X_\cdot)\right\rVert}_\infty\leq \varepsilon^\alpha\right\}}.$$ Now let $V$ be a bounded vector field with bounded derivatives of all order. The chain rule implies $$\begin{gathered}
J_{t,0}^{-1}V(X_t)=V(X_0)+\int_0^t J_{s,0}^{-1}([V_0,V]+\frac 1 2 \sum_{k=1}^{m+l}[V_k,[V_k,V]])(X_s){\ensuremath{\operatorname{d}\!{s}}}
\\
+ \sum_{k=1}^m \int_0^t J_{s,0}^{-1}[V_k,V](X_s){\ensuremath{\operatorname{d}\!{W_s}}}+\sum_{k=m+1}^{l+m}\int_0^t J_{s,0}^{-1}[V_k,V](X_s){\ensuremath{\operatorname{d}\!{B^H_s}}}.\end{gathered}$$ Thus, applying Theorem \[Chuck\] once more, we obtain $$\begin{gathered}
{\mathsf P\left \{{\left\lVert{\langle}v,J_{\cdot,0}V(X_\cdot){\rangle}\right\rVert}_\infty<\varepsilon\right\}}\leq C \varepsilon^p+
\min_{k=1,\ldots,m+l}{\mathsf P\left \{{\left\lVert{\langle}v,J^{-1}_{\cdot,0}[V_k,V](X_\cdot) {\rangle}\right\rVert}_\infty\leq \varepsilon^\alpha\right\}}.\end{gathered}$$ Let $n_0$ be the integer from the Hörmander condition. Iterating our consideration above, we obtain $$\begin{gathered}
{\mathsf P\left \{{\langle}v, C_1 v{\rangle}\leq \varepsilon\right\}}\leq C\varepsilon^p+
\min_{V\in\bigcup_{k=1}^{n_0} \Upsilon_k} {\mathsf P\left \{{\left\lVert{\langle}v,J^{-1}_{\cdot,0}V(X_\cdot){\rangle}\right\rVert}_\infty\leq \varepsilon^\alpha\right\}}\end{gathered}$$ for all $\varepsilon$ small enough. Since $\{V(x_0), V\in \bigcup_{k=1}^{n_0} \Upsilon_k\}$ spans ${\mathbb R}^d$, there exists $v$ such that ${\langle}v, V(x_0){\rangle}\neq 0$. Hence, there exists $\varepsilon_0(p)$ such that for all $\varepsilon<\varepsilon_0(p)$ the second term vanishes. As a result, $${\mathsf P\left \{{\langle}v, C_1 v{\rangle}\leq \varepsilon\right\}}\leq C_p \varepsilon^p$$ for all $\varepsilon\leq \varepsilon_0(p)$, as required.
Technical lemmas
================
The following two lemmas concern the Jacobian of the flow generated by the solution $X$ to equation . These are quite standard facts, so we just sketch the proofs.
\[lemma1\] Under assumptions *A1–A4* the matrix valued process $J_{t,0}=(J_{t,0}(i,j))_{i,j=1,\ldots,d}$ given by has an inverse $Z_{t,0}=(Z_{t,0}(i,j))_{i,j=1,\ldots,d}$ for all $t>0$. Moreover, ${\left\{Z_{t,0},t\ge 0\right\}}$ satisfies the following system of equations $$\label{eqInv}
\begin{gathered}
Z_{t,0}(i,j)=\delta(i,j)-\sum_{r=1}^d\bigg[\int_0^t \frac{\partial a_r}{x_j}(s,X_s) Z_{s,0}(r,j){\ensuremath{\operatorname{d}\!{s}}}\\
{} - \sum_{k=1}^m \int_0^t \frac{\partial b_{r,k}}{\partial x_j}(s,X_s)Z_{s,0}(i,r){\ensuremath{\operatorname{d}\!{W}}}^{k}_s-
\sum_{q=1}^l \int_0^t \frac{\partial c_{r,q}}{\partial x_j}(s,X_s)Z_{s,0}(i,r) {\ensuremath{\operatorname{d}\!{B}}}^{H,q}_s\\ + \frac12\sum_{u=1}^d\sum_{v=1}^d \int_0^t \frac{\partial b_{r,v}}{\partial x_u}(s,X_s)\frac{\partial b_{u,v}}{\partial x_r}(s,X_s)Z_{s,0}(i,r){\ensuremath{\operatorname{d}\!{s}}}\bigg].
\end{gathered}$$
The equation is linear, thus possesses a unique solution $Z_{t,0}$. So we need to verify that $Z_{t,0}J_{t,0}=J_{t,0}Z_{t,0}=I_d$, the identity matrix. The equality clearly holds for $t=0$. To show it for $t>0$, it is enough to show that the differentials of $Z_{t,0}J_{t,0}$ and of $J_{t,0}Z_{t,0}$ vanish. But this can be routinely checked using the Itô formula.
Denote for $t\geq s$ $J_{t,s}=J_{t,0}J^{-1}_{s,0}$.
\[lemma2\] Under assumptions *A1–A4*, the Malliavin derivatives of the solution to are given by $$\begin{gathered}
\operatorname{\mathsf{D}}^{W,k}_s X_t=J_{t,s} b_{\cdot,k}(s,X_s){\mathbf{1}_{s\le t}},\,k=1,\ldots,m, \label{eqW}\\
\operatorname{\mathsf{D}}^{H,q}_s X_t= J_{t,s} c_{\cdot,q}(s,X_s){\mathbf{1}_{s\le t}},\, q=1,\ldots,l.\end{gathered}$$
The argument is exactly the same for both equations, so we prove only . Evidently, $\operatorname{\mathsf{D}}_s^{W,k} X_t = 0$ for $s>t$, so suppose that $s\le t$. Due to the closedness of the derivative, we can freely differentiate as if the integrals were finite sums, in particular, using the chain rule, we can write for $i=1,\dots,d$ $$\begin{gathered}
\operatorname{\mathsf{D}}_s^{W,k} \int_0^t a_i(u,X_u){\ensuremath{\operatorname{d}\!{u}}} = \int_0^t \operatorname{\mathsf{D}}_s^{W,k} a_i(u,X_u){\ensuremath{\operatorname{d}\!{u}}} =
\sum_{r=1}^{d}\int_s^t \frac{\partial}{\partial x_r}a_i(u,X_u)\operatorname{\mathsf{D}}_s^{W,k}X^r_u {\ensuremath{\operatorname{d}\!{u}}}\end{gathered}$$ and similarly $$\begin{gathered}
\operatorname{\mathsf{D}}_s^{W,k} \int_0^t c_{i,q}(u,X_u){\ensuremath{\operatorname{d}\!{B}}}^{H,q}_u = \sum_{r=1}^{d}\int_s^t \frac{\partial}{\partial x_r}c_{i,q}(u,X_u)\operatorname{\mathsf{D}}_s^{W,k}X^r_u {\ensuremath{\operatorname{d}\!{B}}}^{H,q}_s, q=1,\dots,l,\\
\operatorname{\mathsf{D}}_s^{W,k} \int_0^t b_{i,j}(u,X_u){\ensuremath{\operatorname{d}\!{W}}}^{j}_u = \sum_{r=1}^{d}\int_s^t \frac{\partial}{\partial x_r}b_{i,j}(u,X_u)\operatorname{\mathsf{D}}_s^{W,k}X^r_u {\ensuremath{\operatorname{d}\!{W}}}^j_s, j=1,\dots,m, j\neq k.\end{gathered}$$ To differentiate the integral w.r.t. $W^k$, approximate it by an integral sum and note that we will have an extra term corresponding to the derivative of the increment of $W^k$ on the interval containing $s$. Passing to the limit, we get $$\operatorname{\mathsf{D}}_s^{W,k} \int_0^t b_{i,k}(u,X_u){\ensuremath{\operatorname{d}\!{W}}}^{k}_u = b_{i,k}(s,X_s) + \sum_{r=1}^{d}\int_s^t \frac{\partial}{\partial x_r}b_{i,k}(u,X_u)\operatorname{\mathsf{D}}_s^{W,k}X^r_u {\ensuremath{\operatorname{d}\!{W}}}^k_s.$$ Therefore, we have for $s\le t$ the following linear equation on $\operatorname{\mathsf{D}}_s^{W,k}X_u$: $$\begin{gathered}
\operatorname{\mathsf{D}}^{W,k}_s X_t = b_{\cdot,k}(s,X_s) + \sum_{r=1}^{d}\bigg[
\int_s^t \frac{\partial}{\partial x_r}a(u,X_u)\operatorname{\mathsf{D}}_s^{W,k}X^r_u {\ensuremath{\operatorname{d}\!{u}}}
\\ + \sum_{q=1}^l \int_s^t \frac{\partial}{\partial x_r}c_{\cdot,q}(u,X_u)\operatorname{\mathsf{D}}_s^{W,k}X^r_u {\ensuremath{\operatorname{d}\!{B}}}^{H,q}_s
+ \sum_{j=1}^m
\int_s^t \frac{\partial}{\partial x_r}b_{\cdot,j}(u,X_u)\operatorname{\mathsf{D}}_s^{W,k}X^r_u {\ensuremath{\operatorname{d}\!{W}}}^j_s
\bigg].\end{gathered}$$ On the other hand, from we can write $$\begin{gathered}
J_{t,0}=J_{s,0}+\sum_{r=1}^d\bigg[\int_s^t \frac{\partial a_i}{\partial x_r}(u,X_u) J_{u,0}{\ensuremath{\operatorname{d}\!{u}}} \\ +\sum_{k=1}^m \int_s^t \frac{\partial b_{\cdot,k}}{\partial x_r}(s,X_s)J_{u,0}{\ensuremath{\operatorname{d}\!{W}}}^{k}_u
+ \sum_{q=1}^l \int_s^t \frac{\partial c_{\cdot,q}}{\partial x_r}(s,X_s)J_{u,0} {\ensuremath{\operatorname{d}\!{B}}}^{H,q}_u\bigg],\end{gathered}$$ which, upon multiplying by $J_{s,0}^{-1}b_{\cdot,k}(s,X_s)$ on the right leads to the same equation on $J_{u,s} b_{\cdot,k}(s,X_s)$ as that on $\operatorname{\mathsf{D}}_s^{W,k}X^r_u$. Hence, by uniqueness, we get the desired result.
Further we establish a simple estimate on the Itô integral of a Hölder continuous integrand.
\[estimInt\] Let ${\left\{f(t),t\in [0,T]\right\}}$ be an $\mathbf{F}$-adapted stochastic process such that ${\mathsf{E}\left[\,{\left\lVertf\right\rVert}_\theta^p\,\right]}<\infty $ for all $p\geq 1$, and $0<\delta<\Delta\le T$. Then for all $s,t,u\in[0,T]$ such that $u<s<t$, $t-s<\delta$, $t-u\le \Delta$ it holds $$\left|\int_s^t\big(f(v)-f(u)\big){\ensuremath{\operatorname{d}\!{W_v}}} \right|\leq \Delta^\theta\delta^{1/2}\xi_{\Delta,\delta},$$ where ${\mathsf{E}\left[\, \xi^p_{\Delta,\delta}\,\right]}<C_p{\mathsf{E}\left[\,{\left\lVertf\right\rVert}_\theta^p\,\right]}$ for all $p\ge 1$.
It suffices to establish the required result for $p$ large enough, then one can get deduce it for all $p\ge 1$ with the help of Jensen’s inequlaity.
By the Garsia–Rodemich–Rumsey inequality, we get $$\begin{gathered}
\left|\int_s^t\big(f(v)-f(u)\big){\ensuremath{\operatorname{d}\!{W_v}}} \right|\leq C|t-s|^{1/4}\left(\int_s^t\int_s^t \frac{|\int_x^y \big(f(v)-f(u)\big){\ensuremath{\operatorname{d}\!{W_v}}}|^{8}}{|x-y|^{4}} {\ensuremath{\operatorname{d}\!{x}}}{\ensuremath{\operatorname{d}\!{y}}}\right)^{1/8}\\
\le C\Delta^\theta \delta^{1/2 }\xi_{\Delta,\delta}, \end{gathered}$$ where $$\xi_{\Delta,\delta}=\Delta^{-\theta}\delta^{-1/4}\left(\int_s^t\int_s^t \frac{|\int_x^y \big(f(v)-f(u)\big){\ensuremath{\operatorname{d}\!{W_v}}}|^{8}}{|x-y|^{4}} {\ensuremath{\operatorname{d}\!{x}}}{\ensuremath{\operatorname{d}\!{y}}}\right)^{1/8}.$$ For $p>8$ the Hölder inequality entails that $$\begin{gathered}
{\mathsf{E}\left[\,\xi^p_{\Delta,\delta}\,\right]}\leq \Delta^{-\theta p}\delta^{-p/4}(t-s)^{2(p/8-1)} \int_s^t\int_s^t {\mathsf{E}\left[\,\dfrac{|\int_x^y \big(f(v)-f(u)\big){\ensuremath{\operatorname{d}\!{W_v}}}|^p}{|x-y|^{p/2}}\,\right]}{\ensuremath{\operatorname{d}\!{x}}}{\ensuremath{\operatorname{d}\!{y}}}
\\ \leq C_p\Delta^{-\theta p}\delta^{-2} \int_s^t\int_s^t {\mathsf{E}\left[\,\dfrac{\left(\int_x^y |f(v)-f(u)|^2{\ensuremath{\operatorname{d}\!{v}}}\right)^{p/2}}{|x-y|^{p/2}}\,\right]}{\ensuremath{\operatorname{d}\!{x}}}{\ensuremath{\operatorname{d}\!{y}}}\\
\leq C_p \Delta^{-\theta p}\delta^{-2}{\mathsf{E}\left[\,{\left\lVertf\right\rVert}_\theta^p\,\right]}\Delta^{p\theta} \delta^2= C_p{\mathsf{E}\left[\,{\left\lVertf\right\rVert}_\theta^p\,\right]}.
\end{gathered}$$ Hence, we arrive at the desired statement.
We also need the result concerning the pathwise regularity property of $X$. It establishes certain exponential integrability of the Hölder seminorm of $X$, so it is an interesting result on its own.
Let $\{X_t,t\in[0,T]\}$ be the solution to . Assume that $a,b,c$ satisfy the assumption *C1*. Then $X \in C^\theta[0,T]$ for $\theta \in (0,1/2)$ and ${\mathsf{E}\left[\,\exp{\left\{K{\left\lVertX\right\rVert}_\theta^q\right\}}\,\right]}<\infty$ for all $q\in\big(0,q^*\big)$, $K>0$, where $$q^* = \frac{4H}{2(H+\theta)+1}\wedge \frac{2H+1}{4H}.$$ In particular, ${\mathsf{E}\left[\,{\left\lVertX\right\rVert}^p_\theta\,\right]}<\infty$ for all $p> 0$.
Define for ${\varepsilon}\in(0,T]$ $${\left\lVertX\right\rVert}_{\theta,{\varepsilon}} = \sup_{0\le t-{\varepsilon}\le s<t\le T}\frac{{\left\lvertX_t-X_s\right\rvert}}{(t-s)^{\theta}}.$$ Clearly, ${\left\lVertX\right\rVert}_\theta \le {\left\lVertX\right\rVert}_{\theta,{\varepsilon}} + 2{\varepsilon}^{-\theta}{\left\lVertX\right\rVert}_\infty$. It follows from [@Integr equation (4)] that $${\left\lVertX\right\rVert}_{\theta,{\varepsilon}} \le C_1\left(\big\|I^b\big\|_\theta + \Lambda_\mu\big(1+ {\left\lVertX\right\rVert}_\infty{\varepsilon}^{\mu-\theta}\big) \right),$$ for any ${\varepsilon}\in(0,C_2 \Lambda_\mu^{-1/\mu}] $ where $C_1,C_2$ are some positive constants, $\mu\in(1/2,H)$, $\Lambda_\mu = {\left\lVertB_H\right\rVert}_\mu+1$, $I^b_t = \int_0^t b(X_s)dW_s.$ Therefore, setting ${\varepsilon}= C_2 \Lambda_\mu^{-1/\mu}$, we obtain $${\left\lVertX\right\rVert}_{\theta} \le C_1\left(\big\|I^b\big\|_\theta + \Lambda_\mu+ 2 {\left\lVertX\right\rVert}_\infty\Lambda^{\theta/\mu} \right) \le
C\left(\big\|I^b\big\|_\theta + \Lambda_\mu+ {\left\lVertX\right\rVert}_\infty^{p'} + \Lambda^{q'\theta/\mu} \right),$$ where $p'>1$, and $q' = p'/(p'-1)$ is the exponent conjugate to $p'$. Therefore, $${\left\lVertX\right\rVert}_{\theta}^q \le
C\left(\big\|I^b\big\|_\theta^q + \Lambda_\mu^q+ {\left\lVertX\right\rVert}_\infty^{p'q}+ \Lambda^{qq'\theta/\mu} \right).$$ Evidently, $q^*<1$, so it follows from [@ShSh Lemma 1] that ${\mathsf{E}\Big[\,\exp{\left\{K\big\|I^b\big\|_\theta^q\right\}}\,\Big]}<\infty$ for all $K>0$. Further, $\Lambda_\mu$ is an almost surely finite supremum of a centered Gaussian family, so by Fernique’s theorem ${\mathsf{E}\Big[\,\exp{\left\{K \Lambda_\mu^z\right\}}\,\Big]}<\infty$ for any $K>0$, $z\in(0,2)$. Finally, by [@Integr Corollary 4], ${\mathsf{E}\Big[\,\exp{\left\{K {\left\lVertX\right\rVert}^{z}_\infty\right\}}\,\Big]}<\infty$ for all $K>0, z<4H/(2H+1)$. Now if $p'>1$ is close to $4H q^{-1}(2H+1)^{-1}$ (thanks to the bound on $q$ such choice is possible) and $\mu$ is close to $H$, then $q'$ is close to $4H/(4H-q(2H+1))$, and $qq'\theta/\mu$ is close to $4q\theta/(4H-q(2H+1))$, which is less than $2$. Indeed, the last statement is equivalent to $q(2\theta+2H+1)<4H$, which is true thanks to the restriction on $q$. Thus, we get the desired integrability.
The following result is a Norris type lemma for mixed SDEs. It is a crucial result to prove existence of density under the Hörmander condition. Loosely speaking, this statement says that if $$Y_t=Y_0+\int_0^t a(s){\ensuremath{\operatorname{d}\!{s}}}+\int_0^t b(s){\ensuremath{\operatorname{d}\!{W_s}}}+\int_0^tc(s){\ensuremath{\operatorname{d}\!{B^H_s}}},\label{NorrisEq}$$ ${\left\lVertY\right\rVert}_\infty = {\left\lVertY\right\rVert}_{\infty; [0,T]}$ is small, then ${\left\lVertb\right\rVert}_\infty$ and ${\left\lVertc\right\rVert}_\infty$ can not be large. This means that the integral w.r.t. $W$ and w.r.t. $B^H$ can not compensate each other well. The rigorous formulation is as follows.
\[Chuck\] Assume that $H\in(1/2,2/3)$, $\theta\in\big(\theta_*,1/2\big)$, where $$\theta_* = \frac{H-\frac 12}{3-4H},$$ and that $a,b,c$ in are $\mathbf F$-adapted processes satisfying ${\mathsf{E}\left[\,{\left\lVerta\right\rVert}^p_\infty+{\left\lVertb\right\rVert}^p_\theta+{\left\lVertc\right\rVert}^p_\theta\,\right]}<\infty$ for all $p\geq 1$. Then exists $q>0$ such that for all $p\geq 1, \varepsilon>0$ $${\mathsf P\left \{{\left\lVertY\right\rVert}_\infty<\varepsilon \textrm{ and } {\left\lVertb\right\rVert}_\infty+{\left\lVertc\right\rVert}_\infty>\varepsilon^q\right\}}\leq C_p\varepsilon^p.$$
Here we imitate the proof of in [@BauHai Proposition 3.4]. For notational simplicity, we assume that $T=1$. For some positive integers $M$ and $r$ denote $\Delta = 1/M$, $\delta = \Delta/r$ and define the following uniform partitions of $[0,1]$: $T_N = N\delta$, $N=0,\dots,M$; $t_n=\delta n$, $n=0,\ldots, Mr$. Further, fix some $\breve H\in (1/2,H)$ and write for $N=0,\dots,M-1$, $n= Nr,\dots, (N+1)r-1$ (so that $t_n\in [T_N,T_{N+1})$), $i=1,\ldots, d$ $$\label{eqS}
\begin{gathered}
\big{\langle}c_i(T_N), B^H_{t_{n+1}}-B^H_{t_n} \big{\rangle}+
\big{\langle}b_{i}(T_N), W_{t_{n+1}}-W_{t_n}\big{\rangle}\leq |Y^i_{t_{n+1}}-Y^i_{t_n}|+ \delta{\left\lVerta\right\rVert}_\infty \\
+ {\left\lvert\int_{t_n}^{t_{n+1}}{\langle}b_i(s)-b_i(T_N),{\ensuremath{\operatorname{d}\!{W_s}}}{\rangle}\right\rvert}+{\left\lvert\int_{t_n}^{t_{n+1}}{\langle}c_i(s)-c_i(T_N), {\ensuremath{\operatorname{d}\!{B^H_s}}}{\rangle}\right\rvert}\\ \leq 2{\left\lVertY\right\rVert}_\infty+\delta{\left\lVerta\right\rVert}_\infty +C\Delta^\theta\delta^{\breve{H}}{\left\lVertc\right\rVert}_\theta{\left\lVertB^H\right\rVert}_{\breve{H}}+\Delta^{\theta}\delta^{1/2}\xi_{\Delta,\delta}=:S,
\end{gathered}$$where in the last step we have used the Young–Love inequality (see e.g. [@NualSau Proposition 1]) and [Lemma \[estimInt\]]{}.
For processes $\xi,\zeta$ denote $$\label{V_N}
V_N(\xi,\zeta) = \sum_{n=Nr}^{(N+1)r-1} \left(\xi_{t_{n+1}}-\xi_{t_n}\right)\left(\zeta_{t_{n+1}}-\zeta_{t_n}\right);$$ we remind that the summation is in fact over $t_n\in[T_N,T_{N+1})$. Squaring the both sides of , summing over $n= Nr,\dots, (N+1)r-1$ and then taking the square root we get $$\begin{gathered}
\biggl(\sum_{u,v=1}^m \sum_{v=1}^m b_{i,u}(T_N)b_{i,v}(T_N)V_N(W^u,W^v) + \sum_{u,v=1}^l c_{i,u}(T_N)c_{i,v}(T_N)V_N(B^{H,u},B^{H,v})
\\
+ \sum_{u=1}^m \sum_{v=1}^l b_{i,u}(T_N)c_{i,v}(T_N)V_N(W^u,B^{H,v}) \biggr)^{1/2}\leq C \Delta^{1/2}\delta^{-1/2}\mathit{S}.\end{gathered}$$ Therefore, $$\label{sumincrements}
\begin{gathered}
\sum_{u=1}^m {\left\lvertb_{i,u}(T_N)\right\rvert}V_N(W^u,W^u)^{1/2}+\sum_{v=1}^l {\left\lvertc_{i,v}(T_N)\right\rvert}V_N(B^{H,v},B^{H,v}) \\
\le C\biggl ( \sum_{1\le u<v\le m}|b_{i,u}(T_N)|^{1/2}|b_{i,v}(T_N)|^{1/2}|V_N(W^u,W^v)|^{1/2}\\ +\sum_{1\le u<v\le l}|c_{i,u}(T_N)|^{1/2}|c_{i,v}(T_N)|^{1/2}|V_N(B^{H,u},B^{H,v})|^{1/2}\\
+\sum_{u=1}^m \sum_{v=1}^l |b_{i,u}(T_N)|^{1/2}|c_{i,v}(T_N)|^{1/2}|V_N(W^u,B^{H,v}|^{1/2}+\Delta^{1/2}\delta^{-1/2}\mathit S \biggr).
\end{gathered}$$ Further, for arbitrary $f\in C^\theta[0,1]$, $${\left\lvert\Delta\sum_{N=0}^{M-1}|f(T_N)|-{\left\lVertf\right\rVert}_{L^1[0,1]}\right\rvert}\leq {\left\lVertf\right\rVert}_\theta\Delta^{\theta},$$ which yields $$\begin{gathered}
\sum_{u=1}^m{\left\lVertb_{i,u}\right\rVert}_{L^1[0,1]}
+\sum_{v=1}^l{\left\lVertc_{i,v}\right\rVert}_{L^1[0,1]}
\\
\leq \sum_{u=1}^m\bigg(\Delta^{\theta}{\left\lVertb_{i,u}\right\rVert}_\theta+
\Delta\sum_{N=0}^{M-1} {\left\lvertb_{i,u}(T_N)\right\rvert}\bigg) + \sum_{v=1}^l\bigg(\Delta^{\theta}{\left\lVertc_{i,v}\right\rVert}_\theta+
\Delta\sum_{N=0}^{M-1} {\left\lvertc_{i,v}(T_N)\right\rvert}\bigg)\\
\leq \sum_{u=1}^m\bigg(\Delta^{\theta}{\left\lVertb\right\rVert}_\theta+
\Delta^{1/2}{\left\lVertb\right\rVert}_\infty\sum_{N=0}^{M-1} {\left\lvert\Delta^{1/2} - V_N(W^u,W^u)^{1/2}\right\rvert}
\\+
\Delta^{1/2}\delta^{1/2-H}\sum_{N=0}^{M-1} {\left\lvertb_{i,u}(T_N)\right\rvert} V_N(W^u,W^u)^{1/2}\bigg)
\\
+ \sum_{v=1}^l\bigg(\Delta^{\theta}{\left\lVertc\right\rVert}_\theta+
\Delta^{1/2}\delta^{1/2-H}{\left\lVertc\right\rVert}_\infty\sum_{N=0}^{M-1} {\left\lvert\Delta^{1/2}\delta^{H-1/2} - V_N(B^{H,v},B^{H,v})^{1/2}\right\rvert}\\
+
\Delta^{1/2}\delta^{1/2-H}\sum_{N=0}^{M-1}{\left\lvertc_{i,v}(T_N)\right\rvert}
V_N(B^{H,v},B^{H,v})^{1/2}\bigg).\end{gathered}$$ Therefore, using , we arrive at $$\begin{gathered}
{\left\lVertb\right\rVert}_{L^1[0,1]}+{\left\lVertc\right\rVert}_{L^1[0,1]}\leq C\bigg ( \Delta^{\theta}\big({\left\lVertb\right\rVert}_\theta+ {\left\lVertc\right\rVert}_\theta\big)\\
+\Delta^{1/2}\delta^{1/2-H} {\left\lVertb\right\rVert}_\infty\sum_{N=0}^{M-1}\sum_{u,v=1}^m {\left\lvert\Delta^{1/2}\delta_{u,v} - {\left\lvertV_N(W^u,W^v)\right\rvert}^{1/2}\right\rvert}\\ +\Delta^{1/2}\delta^{1/2-H} {\left\lVertc\right\rVert}_\infty\sum_{N=0}^{M-1}\sum_{u,v=1}^l {\left\lvert\Delta^{1/2}\delta^{H-1/2}\delta_{u,v} - {\left\lvertV_N(B^{H,u},B^{H,v})\right\rvert}^{1/2}\right\rvert}\\
+ \Delta^{1/2}\delta^{1/2-H}{\left\lVertc\right\rVert}^{1/2}_\infty
{\left\lVertb\right\rVert}^{1/2}_\infty
\sum_{N=0}^{M-1}\sum_{u=1}^m\sum_{v=1}^l {\left\lvertV_N(W^u,B^{H,v})\right\rvert}^{1/2}
+\delta^{-H}S\bigg)\\
\le C\bigg ( {\left\lVertb\right\rVert}_\theta+ {\left\lVertc\right\rVert}_\theta+
\Delta^{-1/4}\delta^{3/4-H} {\left\lVertb\right\rVert}_\infty R^W\\ +\Delta^{H-1}\delta^{1-H} {\left\lVertc\right\rVert}_\infty R^B
+ \Delta^{-1/4}\delta^{(1-H)/2}\left({\left\lVertc\right\rVert}_\infty +
{\left\lVertb\right\rVert}_\infty\right)R^{W,B}
+\delta^{-H}S\bigg),\end{gathered}$$ where $$\label{RWB}
\begin{gathered}
R^W = \Delta^{3/4}\delta^{-1/4} \sum_{N=0}^{M-1}\sum_{u,v=1}^m {\left\lvert\Delta^{1/2}\delta_{u,v} - {\left\lvertV_N(W^u,W^v)\right\rvert}^{1/2}\right\rvert},\\
R^B = \Delta^{H-3/2}\delta^{1/2}\sum_{N=0}^{M-1}\sum_{u,v=1}^l {\left\lvert\Delta^{1/2}\delta^{H-1/2}\delta_{u,v} - {\left\lvertV_N(B^{H,u},B^{H,v})\right\rvert}^{1/2}\right\rvert},\\
R^{W,B} = \Delta^{3/4}\delta^{-H/2} \sum_{N=0}^{M-1}\sum_{u=1}^m
\sum_{v=1}^{l} {\left\lvertV_N(W^u,B^{H,v})\right\rvert}^{1/2}.
\end{gathered}$$ Further we use the following interpolation inequality, valid for any $f\in C^\theta[0,1]$ and $\gamma<1$: $${\left\lVertf\right\rVert}_\infty\leq C\big(\gamma {\left\lVertf\right\rVert}_\theta+\gamma^{-1/\theta}{\left\lVertf\right\rVert}_{L^1[0,1]}\big).$$ for any $\gamma\leq1$. Thus, $$\label{bcinfty}
\begin{gathered}
{\left\lVertb\right\rVert}_\infty+{\left\lVertc\right\rVert}_\infty\leq C\big({\left\lVertb\right\rVert}_\theta + {\left\lVertc\right\rVert}_\theta\big)\gamma +
C\gamma^{-1/\theta}\Big(\big({\left\lVertb\right\rVert}_\theta + {\left\lVertc\right\rVert}_\theta\big)\Delta^\theta \\
+ ({\left\lVertb\right\rVert}_\infty + {\left\lVertc\right\rVert}_\infty)\Big[
\Delta^{-1/4}\delta^{3/4-H}R^W + \Delta^{H-1}\delta^{1-H}R^B + \Delta^{-1/4}\delta^{(1-H)/2} R^{W,B}\Big] \\
+ \delta^{-H}{\left\lVertY\right\rVert}_\infty + \delta^{1-H} {\left\lVerta\right\rVert}_\infty + \Delta^\theta \delta^{\breve H-H}{\left\lVertc\right\rVert}_\theta{\left\lVertB^H\right\rVert}_{\breve H} + \Delta^\theta \delta^{1/2-H} \xi_{\Delta,\delta}\Big).
\end{gathered}$$ Now we want to put $$\begin{gathered}
\Delta^{\beta}\sim \varepsilon^\beta, \delta\sim \varepsilon^{\alpha}, \gamma\sim \varepsilon^\eta, \alpha>\beta>0, \eta>0,\end{gathered}$$ so that in the right-hand side of , the exponents of $\varepsilon$ are positive for all terms except ${\left\lVertY\right\rVert}_\infty$. Since $(H-1/2)/\theta\le (3-4H)<1$, it is possible to take $\beta/\alpha\in \big((H-1/2)/\theta, (3-4H)\big)$ so that both $\theta \beta + (1/2-H)\alpha$ and $-\beta/4 + (3/4-H)\alpha$ are positive. Also $(H-1)\beta +(1-H)\alpha = (1-H)(\alpha-\beta)>0$, $-\beta/4 + (1-H)\alpha/2 > -\beta/4 + (3/4-H)\alpha>0$, $\theta \beta + (\breve H - H)\alpha > \theta \beta + (1/2-H)\alpha>0$. Therefore, by choosing $\eta$ small enough we can make all needed exponents positive.
Thus, for some $\kappa>0$ and $C_1>0$ we have $$\begin{gathered}
{\left\lVertb\right\rVert}_\infty+{\left\lVertc\right\rVert}_\infty\leq C_1{\left\lVertY\right\rVert}_\infty {\varepsilon}^{-\lambda} +
C_1{\varepsilon}^\kappa\Big( ({\left\lVertb\right\rVert}_\infty + {\left\lVertc\right\rVert}_\infty)\big[
R^W + R^B + R^{W,B}]\\ + {\left\lVertb\right\rVert}_\theta + {\left\lVertc\right\rVert}_\theta
+ {\left\lVerta\right\rVert}_\infty + {\left\lVertc\right\rVert}_\theta{\left\lVertB^H\right\rVert}_{\breve H} + \xi_{\Delta,\delta}\Big),\end{gathered}$$ where $\lambda= H\alpha+\eta/\theta$. Consequently, for ${\varepsilon}$ small enough $$\begin{gathered}
{\mathsf P\left \{{\left\lVertb\right\rVert}_\infty+{\left\lVertc\right\rVert}_\infty>\varepsilon^{\kappa/2} \text{ and } {\left\lVertY\right\rVert}_\infty<\varepsilon^{\lambda+\kappa}\right\}} \\
\le {\mathsf P\left \{R^W \ge {\varepsilon}^{-\kappa/3}\right\}} + {\mathsf P\left \{R^B \ge {\varepsilon}^{-\kappa/3}\right\}} + {\mathsf P\left \{R^{W,B} \ge {\varepsilon}^{-\kappa/3}\right\}} \\+ {\mathsf P\left \{{\left\lVertb\right\rVert}_\theta + {\left\lVertc\right\rVert}_\theta + {\left\lVertc\right\rVert}_\theta{\left\lVertB^H\right\rVert}_{\breve H} + \xi_{\Delta,\delta}\ge {\varepsilon}^{-\kappa/3}\right\}}.\end{gathered}$$ Now the statement follows by applying Lemmas \[estimInt\] and \[concentrat\] and the Chebyshev inequality.
\[concentrat\] Let $R^W,R^B$ and $R^{W,B}$ be given by and . Then we have for any $h>1$ the following concentration inequalities $$\begin{aligned}
{\mathsf P\left \{R^{W}\geq h\right\}}&\leq \dfrac{C}{\Delta}\exp(-Ch^2),\label{conc1}\\
{\mathsf P\left \{R^B\geq h\right\}}&\leq \dfrac{C}{\Delta}\exp(-Ch^2),\label{conc2}\\
{\mathsf P\left \{R^{W,B}\geq h\right\}}&\leq \dfrac{C}{\Delta}\exp(-Ch^2)\label{conc3}.\end{aligned}$$
By [@BauHai Lemma 3.1] we have for $h>0$ $$\label{WvWv}
{\mathsf P\left \{\big|\Delta^{1/2}-V_N(W^u,W^u)^{1/2}\big|\Delta^{-1/4}\delta^{-1/4}\geq h\right\}}\leq C\exp\left(-C {h^2}\right).$$ Further, let $u\neq v$. Since $W^u$ and $W^v$ are independent, and $W^v$ has independent increments, then conditional on $W^v$, $V_N(W^w,W^u)\Delta^{-1/2}\delta^{-1/2}$ has a centered Gaussian distribution with the variance $ V_N(W^v,W^v)\Delta^{-1} $. Therefore, $$\begin{gathered}
{\mathsf P\left \{|V_N(W^u,W^v)|^{1/2}\Delta^{-1/4}\delta^{-1/4}\geq h\right\}}\\ = {\mathsf{E}\left[\,{\mathsf P\left \{|V_N(W^u,W^v)|\Delta^{-1/2}\delta^{-1/2}\geq h^2\right\}}\big|\ W^v\,\right]}
\le C\,{\mathsf{E}\left[\,\exp{\left\{-\frac{h^4\Delta}{4 V_N(W^v,W^v)}\right\}}\,\right]}\\
\le C\exp{\left\{-\frac{h^4\Delta}{4(\Delta^{1/2}+v)^2}\right\}} + {\mathsf P\left \{\big|\Delta^{1/2}-V_N(W^v,W^v)^{1/2}\big|\ge v\right\}}\\
\le C\exp{\left\{-\frac{h^4\Delta }{8(\Delta+v^2)}\right\}} + C\exp{\left\{-C\frac{v^2}{\Delta^{1/2}\delta^{1/2}}\right\}}, \end{gathered}$$ where we have used . Setting $v^2 = h^2\Delta$ and recalling that $\Delta\ge\delta$ we get $${\mathsf P\left \{|V_N(W^u,W^v)|^{1/2}\Delta^{-1/4}\delta^{-1/4}\geq h\right\}}\leq C\exp\left(-C {h^2}\right).$$ Combining this with , we get $$\begin{gathered}
{\mathsf P\left \{R^W\ge h\right\}}\le \sum_{N=0}^{M-1} \sum_{u,v=1}^m {\mathsf P\left \{\Delta^{-1/4}\delta^{-1/4} \big|\Delta^{1/2}\delta_{u,v}-{\left\lvertV_N(W^u,W^v)\right\rvert}^{1/2}\big|\ge hm^2\right\}}\\
\le \frac{C}{\Delta} \exp{\left\{-Ch^2\right\}}.\end{gathered}$$ Using the inequalities from [@BauHai Lemma 3.2] and repeating the last step, we get .
The estimate is proved similarly to , so we omit some details. Write $$\begin{gathered}
{\mathsf P\left \{|V_N(W^u,B^{H,v})|^{1/2}\Delta^{-1/4}\delta^{-H/2}\geq h\right\}}\\ = {\mathsf{E}\left[\,{\mathsf P\left \{|V_N(W^u,B^{H,v})|\Delta^{-1/2}\delta^{-H}\geq h^2\right\}}\big|\ B^{H,v}\,\right]}
\le C\,{\mathsf{E}\left[\,\exp{\left\{-\frac{h^4\Delta\delta^{2H-1}}{4 V_N(B^{H,v},B^{H,v})}\right\}}\,\right]}\\
\le C\exp{\left\{-\frac{h^4\Delta\delta^{2H-1}}{4(\Delta^{1/2}\delta^{H-1/2}+v)^2}\right\}} + {\mathsf P\left \{\big|\Delta^{1/2}\delta^{H-1/2}-V_N(B^{H,v},B^{H,v})^{1/2}\big|\ge v\right\}}\\
\le C\exp{\left\{-\frac{h^4\Delta\delta^{2H-1} }{8(\Delta\delta^{2H-1}+v^2)}\right\}} + C\exp{\left\{-C\frac{v^2}{4\Delta^{2H-1}\delta}\right\}}.\end{gathered}$$ Setting $v^2 = h^2 \Delta \delta^{2H-1}$ and taking into account that $\Delta\ge \delta$, we arrive at $${\mathsf P\left \{|V_N(W^u,B^{H,v})|^{1/2}\Delta^{-1/4}\delta^{-H/2}\geq h\right\}}\le C\exp{\left\{-Ch^2\right\}}.$$ From here is deduced similarly to .
[1]{} F. Baudoin, M. Hairer, *A version of Hörmander’s theorem for the fractional Brownian motion.* Probab. theory Relat. Fields(2000), 373–395. T. Cass, M. Hairer, C. Litterer, S. Tindel, *Smoothness of the density for solutions to Gaussian Rough Differential Equations.* ArXiv: math.PR/1209.3100 P. Friz, N. Victoir, [Multidimensional stochastic processes as rough paths. Theory and applications.]{} Cambridge University Press, 2010. D. Nualart, *The Malliavin calculus and related topics* Second edition,Springer-Verlag, 2006. D. Nualart, B. Saussereau, *Malliavin calculus for stochastic differential equations driven by a fractional Brownian motion.* Stochastic Process. Appl. **119** (2009), 391–409. I. Nourdin. T. Simon, *On the absolute continuity of one-dimensional SDEs driven by a fractional Brownian motion.* Statist. Probab. Lett. **76** (2006), 907–912. G. Shevchenko, T. Shalaiko *Malliavin regularity of solutions to mixed stochastic differential equations.* Statist. Probab. Lett. **83** (2013), 2638–-2646. G. Shevhcenko, *Integrability of solutions to mixed stochastic differential equations.* Ukrainian Math. Bulletin **10** (2013), 559–574. G. Shevchenko *Mixed stochastic delay differential equations*, Teor. Imovirn. Mat. Stat. **89** (2013), 169–182.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'Using the multi-band $d-p$ model and unrestricted Hartree-Fock approximation we investigate the electronic structure and spin-orbital order in three-dimensional VO$_3$ lattice. The main aim of this investigation is testing if simple $d-p$ model, with partly filled $3d$ orbitals (at vanadium ions) and $2p$ orbitals (at oxygen ions), is capable of reproducing correctly nontrivial coexisting spin-orbital order observed in the vanadium perovskites. We point out that the multi-band $d-p$ model has to include partly filled $e_g$ orbitals at vanadium ions. The results suggest weak self-doping as an important correction beyond the ionic model and reproduce the possible ground states with broken spin-orbital symmetry on vanadium ions: either $C$-type alternating orbital order accompanied by $G$-type antiferromagnetic spin order, or $G$-type alternating orbital order accompanied by $C$-type antiferromagnetic spin order. Both states are experimentally observed and compete with each other in YVO$_3$ while only the latter was observed in LaVO$_3$. Orbital order is induced and stabilized by particular patterns of oxygen distortions arising from the Jahn-Teller effect. In contrast to time-consuming *ab-initio* calculations, the computations using $d-p$ model are very quick and should be regarded as very useful in solid state physics, provided the parameters are selected carefully.'
author:
- Krzysztof Rościszewski
- 'Andrzej M. Oleś'
title: '$d-p$ model and spin-orbital order in the vanadium perovskites'
---
Introduction {#intro}
============
The spin and orbital ordering found in three dimensional (3D) vanadium perovskites is an old but still very interesting problem with many challenges. It was discussed in numerous experimental and theoretical papers, considering undoped [@Miz99; @Kaw94; @Ngu95; @Saw96; @Nak99; @Ren00; @Bla01; @Kha04; @Kha01; @Hor03; @Jo03; @Ren03; @Fan04; @Ulr03; @Ots06; @Ree06; @Sol06; @Ray07; @Hor08; @Mos09; @Mos10; @Fuj10; @Kum17; @Kim18] and doped by charged defects [@Fuj05; @Ave13; @Ave15] vanadium perovskites. On the theoretical side, the first insightful explanation of the alternating orbital (AO) order was given by Mizokawa, Khomskii, and Sawatzky in 1999 [@Miz99]. They studied the competition between two types of spin-orbital order in vanadates within the so-called lattice model. It was claimed that Jahn-Teller (JT) distortions of the lattice [@Kan60] (see Fig. 1) are primarily responsible for the onset of this order. Sizable tilting of the apical axes of octahedra (out of an ideal cubic structure) was assumed to be the main driving factor which distinguishes between low temperature and high temperature order in LaVO$_3$ or YVO$_3$ [@Miz99].
-1.5cm
![Schematic view of [*cooperative and static*]{} $Q_4$ JT distortions involving rotations of octahedra groups (upper panel) and $Q_2$ distortions (lower panel). For description and classification of different JT modes see Ref. [@Kan60]. Red/blue dots denote positions of vanadium/oxygen ions in the $ab$ plane. []{data-label="fig1"}](fig-jt.pdf){width="12.4cm"}
Easy-to-grasp presentation of the spin and orbital order in the ground state as perceived today by experimentalists was presented by Blake *et al.* [@Bla01]. The phase diagram of the vanadium perovskites $R$VO$_3$ [@Fuj10] shows several spin- and/or orbital ordered phases. In the regime of compounds with low values of ionic radii $r_R$ of rare earth ions $R$ as in YVO$_3$, two antiferromagnetic (AF) phases with complementary spin-orbital order appear: (i) $G$-type AF ($G$-AF) order accompanied by $C$-type alternating orbital order ($C$-AO) with staggered orbitals in $ab$ planes and repeated orbitals along the $c$ axis (below the magnetic transition at $T_{\rm N2}=77$ K) and (ii) $C$-type AF ($C$-AF) order accompanied by $G$-type AO ($G$-AO) order for $T_{\rm N2}<T<T_{\rm N1}$, where $T_{\rm N1}=116$ K is the high-temperature magnetic transition [@Fuj10].
-1.5cm
![Schematic view of JT distortions used for Hartree-Fock computations in the low-temperature phase of YVO$_3$. The long bars denote preferred $yz$ or $zx$ orbitals — their cooperative arrangement forms $C$-AO order. Spins are not shown. The numbers shown close to vanadium positions identify the ions (see the corresponding entries in Table II). Horizontal and vertical directions on the figure correspond to $x$ and $y$ axes, respectively; note that the $x,y$ axes are at 45 $\deg$ angle to the crystallographic $a,b$ axes, i.e., our $x$ direction corresponds to crystallographic (1,1,0) direction. The orbital order is repeated in consecutive layers when moving up along the $z$ axis (this coincides with crystallographic $c$ axis). []{data-label="fig2"}](fig11.pdf){width="12.4cm"}
It is well understood now that at zero temperature, i.e., when YVO$_3$ is orthorhombic, the $zx$ and $yz$ orbitals on vanadium ions alternate between two sublattices forming orbital $C$-AO long range order and this order resembles AF spin order in a single $ab$ plane, see Fig. 2, while along the $c$ axis this order is repeated, i.e., there is an analogy to ordinary spin ferromagnetic (FM) order [@Bla01]. At the same time the spins are arranged according to ordinary 3D Néel state ($G$-AF spin order). At intermediate temperatures $T>77$ K (when YVO$_3$ is monoclinic) this order is reversed: the $G$-AO order is accompanied by $C$-AF spin order, see Fig. 3. The magnetic transition at $T=77$ K is triggered by the dimerization in spin-orbital chains which requires spin fluctuations at finite temperature [@Sir08]. Altogether this transition takes place between two types of spin-orbital order along the $c$ axis which follow the complementarity predicted by the Goodenough-Kanamori rules [@Goode].
-1.5cm
![Schematic view of JT distortions used for Hartree-Fock computations in the zero-temperature phase of LaVO$_3$. Here over the first layer 1 the layer 2 is stacked and the orbitals $\{yz,zx\}$ form $G$-AO order, i.e., alternate along the $c$ axis. The meaning of other symbols is the same as in Fig. 2. []{data-label="fig3"}](fig12.pdf){width="12.4cm"}
The purpose of this paper is to investigate the spin-orbital order in vanadium perovskites within the multiband $d$-$p$ model, i.e., to go beyond the usually used picture of a Mott insulator with $S=1$ spins and $t_{2g}$ orbital degrees of freedom or effective degenerate Hubbard model of $t_{2g}$ electrons. The $d-p$ model includes non-zero on-site Coulomb interactions defined both on oxygen and on transition metal ions and takes into account the possibility of finite self-doping, explained below and applied before to ruthenium, iridium, and titanium oxides [@Ros15; @Ros16; @Ros17]. The $d-p$ model was developed in these papers into a realistic method, capable of computationally cheap and fast realistic investigation of the electronic structure of complex transition metal oxides.
Up to now, the on-site Coulomb interactions on oxygen ions are being neglected in the majority of papers (as a simplification — to reduce the computational effort). However, when Coulomb repulsion elements on oxygens are neglected, the true $d-p$ model parameters are replaced by effective parameters. In particular, the ”effective” Hubbard repulsion on vanadium ions $U_d$ is smaller by about 50% than the “true” $U_d$ repulsion [@Ros17]. Also the so-called *self-doping* [@Ros15; @Ros16; @Ros17], see below, is neglected in traditional effective $3d$-electron models where one assumes that a cation (for example La in LaVO$_3$ or Y in YVO$_3$) behaves according to the [*idealized ionic model*]{}, i.e., donates [*all*]{} valence electrons into a VO$_3$ unit (for La these are: two $4s$ electrons and one $3d$ valence electron). However, in reality, this charge transfer is smaller — it is not exactly 3 but $(3-x)$ instead. Strictly speaking, we mean by this statement that the occupation number of valence electrons on La as obtained say by Mulliken (or Bader) population analysis (during a parallel *ab-initio* computation) will amount to some finite value of $x>0$. This redistribution of electron charge is called here *self-doping*.
In the present investigation we use up-to-date estimations of crystal-field splittings, spin-orbit interaction at vanadium ions and JT distortions. The model is used to study possible types of order, and to establish the easy spin-axis. We also extracted from our computations HOMO-LUMO gaps which can serve as an estimation of the band gap.
The paper is organized as follows. We define the model and its parameters in Sec. \[sec:model\]. The numerical method and its caveats are addressed in Sec. \[sec:num\]. The results are presented and discussed in Sec. \[sec:res\]. In Sec. \[sec:summa\] we present the main conclusions and a short summary.
Hamiltonian {#sec:model}
===========
We introduce the multi-band $d-p$ Hamiltonian for VO$_3$ three-dimensional (3D) cluster which includes five $3d$ orbitals at each vanadium ion and three $2p$ orbitals at each oxygen ion, $${\cal H}= H_{\rm kin}+H_{\rm so}+ H_{\rm diag} +H_{\rm int}.
\label{model}$$ where $H_{\rm kin}$ stands for the kinetic energy, $H_{\rm so}$ for spin-orbit coupling, $H_{\rm diag}$ for the diagonal part of kinetic energy (also including local crystal-field splittings), and $H_{\rm int}$ for the intraatomic Coulomb interactions. Optionally one can add JT part $H_{\rm JT}$ and this will be discussed in Sec. \[sec:JT\]. The cluster geometry and precise forms of different terms are standard; for the detailed formulas see Refs. [@Ros15; @Ros16].
The kinetic part of the Hamiltonian is: $$H_{\rm kin}=\sum_{\{i\mu; j,\nu\},\sigma}\left(t_{i,\mu;j,\nu}
c^{\dagger}_{i,\mu,\sigma}c_{j,\nu,\sigma}^{} + H.c.\right),$$ where we employ a general notation, with $c_{j,\nu,\sigma}^{\dagger}$ standing for the creation of an electron at site $j$ in an orbital $\nu$ with up or down spin, $\sigma=\uparrow,\downarrow$. The model includes all five $3d$ orbital states $\nu\in\{xy,yz,zx,x^2-y^2,3z^2-r^2\}$, and three $2p$ oxygen orbital states $\nu\in\{p_x,p_y,p_z\}$. Alternatively, i.e., when choosing a more intuitive notation, we can write $d_{j,\nu,\sigma}^{\dagger}$ for $d$ orbitals, while $p_{j,\nu,\sigma}^{\dagger}$ for $p$ orbitals. The matrix $t_{i,\mu j,\nu}$ is assumed to be non-zero only for nearest neighbor vanadium-oxygen $d-p$ pairs, and for nearest neighbor oxygen-oxygen $p-p$ pairs. The next-nearest hopping elements are neglected. (The nonzero $t_{i,\mu; j,\nu}$ elements are listed in the Appendix of Ref. [@Ros15]; we use the Slater notation [@Sla54]). As a side remark we recall that models taking into account only three $t_{2g}$ orbitals and neglecting remaining two $e_g$ orbitals are not accurate enough [@Yan11].
The spin-orbit part, $H_{\rm so}=\zeta\sum_i\textbf{L}_i\cdot\textbf{S}_i$, is a one-particle operator (scalar product of angular momentum and spin operators at site $i$), and therefore it can be represented in the form similar to the kinetic part $H_{\rm kin}$ [@Miz96; @Pol12; @Mat13; @Du13], $$H_{\rm so}= \sum_i\left\{\sum_{\mu\neq\nu;\sigma,\sigma'}
t^{so}_{\mu,\sigma;\nu,\sigma'}
d^{\dagger}_{i,\mu,\sigma}d_{i,\nu,\sigma'}^{} + \mathrm{H.c.}\right\},
\label{so-part}$$ with $t^{so}_{\mu,\sigma;\nu,\sigma'}$ elements restricted to single vanadium sites. They all depend on spin-orbit strength $\zeta$ ($\zeta=0.026$ eV; this value was adopted from Ref. [@Dai08]) which is weak but it can have influence on the preferred spin direction. For detailed formula and tables listing $t_{i\;\nu,\sigma';\mu,\sigma}$ elements, see Refs. [@Ros15; @Pol12].
The diagonal part $H_{\rm diag}$ depends only on electron number operators. It takes into account the effects of local crystal fields and the difference of reference orbital energies (here we employ the electron notation), $$\Delta=\varepsilon_d-\varepsilon_p,
\label{Delta}$$ between $d$ and $p$ orbitals (for bare orbital energies). We can fix the reference energy $\varepsilon_d=0$ for $d$ orbitals to zero and use only $\Delta=-\varepsilon_p$ as a parameter, thus we write $$\begin{aligned}
H_{\rm diag}=
\sum_{i;\mu=x,y,z;\sigma}
\varepsilon_p^{} p^\dagger_{i,\mu,\sigma}p_{i,\mu,\sigma}^{} \nonumber \\
+\sum_{m;\mu=xy,yz,... ;\sigma}
f^{cr}_{\mu,\sigma} d^\dagger_{m,\mu,\sigma}d_{m,\mu,\sigma}^{}.\end{aligned}$$ The first sum is restricted to oxygen sites, while the second one runs over vanadium sites. The crystal-field splitting strength vector ($f^{cr}_{\mu,\sigma}$) describes the splitting within $t_{2g}$ levels. For example, in YVO$_3$ the $xy$ orbital is lowered by $\sim 0.017$ eV (according to Ref. [@Ots06]). At the same time, the $\{yz,zx\}$ doublet is also split (this was discussed in some papers, most clearly in Refs. [@Jo03; @Hor08]) in accordance with local JT distortion of particular VO$_6$ octahedron. *We assume ad-hock* that either $yz$ is lower than $zx$ orbital which should correspond to O$_4$ square (in $ab$ plane) when distorted from ideal square into elongated along $y$-direction rhombus, or the opposite: $zx$ is lower than $yz$ orbital which should correspond to O$_4$ distorted into elongated along $x$-direction rhombus (compare Figs. 1 and 2). This splitting value should be $0.1-0.2$ eV what is an educated guess (compare with the estimation from ).
The distance between $t_{2g}$ levels and $e_g$ levels is large, $1.5-2.0$ eV [@Mos09; @Mos10; @Reu12]). We do not take into account a possible splitting within $e_g$ levels as from our previous experience with transition metal perovskites we do not expect it to be an important factor.
The on-site Coulomb interactions $H_{\rm int}(d)$ for $d$ orbitals take the form of a degenerate Hubbard model [@Ole83], $$\begin{aligned}
H_{\rm int}(d)&=&
\sum_{m,\mu<\nu}\left(U_d-\frac{5}{2} J^d_{\mu\nu}\right)
n_{m\mu} n_{m\nu}\nonumber\\
&+&U_d \sum_{m\mu} n_{m\mu\uparrow} n_{m\mu\downarrow}
-2 \sum_{m,\mu<\nu}J^d_{\mu\nu}\,\vec{S}_{m\mu}\cdot\vec{S}_{m\nu}
\nonumber\\
&+& \sum_{m,\mu\neq\nu} J^d_{\mu\nu}\,
d^\dagger_{m\mu\uparrow} d^\dagger_{m\mu\downarrow}
d_{m\nu\downarrow}^{}d_{m\nu\uparrow}^{}.
\label{hubbard2-intra}\end{aligned}$$ where $n_{m\mu}=\sum_{\sigma}n_{m\mu\sigma}$ is the electron density operator in orbital $\mu$, $\{\mu,\nu\}$ enumerate different $d$ orbitals, and $J_{d,\mu\nu}$ is the non-trivial tensor of on-site interorbital exchange (Hund’s) elements for $d$ orbitals; $J_{d,\mu\nu}$ has different entries for the $\{\mu,\nu\}$ pairs corresponding to two $t_{2g}$ orbitals ($J_{\rm H}^t$), and for a pair of two $e_g$ orbitals ($J_{\rm H}^e$), and still different for the case of cross-symmetry terms [@Ole05; @Hor07]; all these elements are included and we assume the Racah parameters: $B=0.1$ eV and $C=4B$.
The local Coulomb interactions $H_{\rm int}(p)$ at oxygen sites (for $2p$ orbitals) are analogous, $$\begin{aligned}
H_{\rm int}(p)&=&
\sum_{i,\mu<\nu,\sigma} \left(U_p-\frac{5}{2} J^p_{\rm H}\right)
n_{i\mu} n_{i\nu} \nonumber\\
&+&U_p \sum_{i\mu} n_{i\mu\uparrow} n_{i\mu\downarrow}
-2J^p_{\rm H}\sum_{i,\mu<\nu} \vec{S}_{i\mu}\cdot\vec{S}_{i\nu}
\nonumber\\
&+& J^p_{\rm H}\sum_{i,\mu\neq\nu}
p^\dagger_{i\mu\uparrow}p^\dagger_{i\mu\downarrow}
p_{i\nu\downarrow}^{}p_{i\nu\uparrow}^{},
\label{hubbard3-intra}\end{aligned}$$ where the intraatomic Coulomb repulsion is denoted as $U_p$ and all off-diagonal elements of the tensor $J^p_{\mu\nu}$ are equal (as they connect the orbitals of the same symmetry), i.e., $J^p_{\mu\nu}\equiv J^p_{\rm H}$. (Up to now, as already mentioned above, $H_{\rm int}(p)$ was neglected in the majority of studies, i.e., for simplicity it was being assumed that $U_p=J^p_{\rm H}=0$.)
In the following we use the parameters $U_d$, $J^d_{\mu\nu}$, $U_p$, and $J_{\rm H}^p$ similar to those used before for titanium oxides [@Ros16; @Ros17]; for the hopping integrals we follow the studies by Mizokawa and Fujimori [@Miz95; @Miz96]. The value $U_p\sim 4.0$ eV was previously used in copper oxides [@Hybe92; @Sing13] but in addition in some test computations we considered a larger value $U_p=6$ eV. (This choice, i.e., $U_p=6$ eV is advocated and reasonably explained in Refs. [@Sing13; @park12].) Concerning the parameter $\Delta$ an educated guess is necessary as no information for the vanadium perovskites is available. However, we have found before that in titanium oxides $\Delta=6.5$ eV is reasonable [@Ros16; @Ros17]. Here for vanadium oxides a smaller value should be more appropriate. Old-fashioned computations, such as those reported in the classical textbook of Harrison [@Har05] and shown in tables therein suggest a value lower by 1.5 eV (i.e., $\Delta=5.0$ eV); a still lower value of 4.0 eV was suggested by Bocquet *et al.* and Imada *et al.* [@Ima98] (note that in these papers the parameter $U_p$ enters only indirectly). We have tried all values in the range $4.0<\Delta<6.5$ eV and found that the most interesting and sensible physical results could be obtained for $\Delta=5.0$ eV.
$\zeta $ $U_d$ $J_{\rm H}^t$ $J_{\rm H}^e$ $U_p$ $J_{\rm H}^p$
---------- ------- --------------- --------------- ------- ---------------
0.026 8.0 0.8 0.9 4.4 0.8
: Parameters of the multi-band model (\[model\]) (all in eV) used in the calculations. For the hopping integrals we adopt the values from Refs. [@Miz95; @Miz96], i.e., $(pd\sigma)\;(pd\pi)\;(pp\sigma),\;(pp\pi)=-2.2,\,1.1,\,0.6,\,-0.15$ eV which correspond to V$-$O distances of 2.0 Å$\,\,$ (we use Slater notation [@Sla54]). The charge transfer energy (defined for bare levels) is taken as $\Delta=5.0$ eV.
\[tab:para\]
Our reference system is LaVO$_3$ where the total electron number in the $d-p$ subsystem is $N_e=17+3=20$ per one VO$_3$ unit provided we assume an ideal ionic model with no self-doping ($x=0$), i.e., all three La valence electrons are transferred to VO$_3$ unit. Another possibility is when the self-doping is finite: we consider $x=0.5$ (then the cation La donates not 3 but rather on average 3-0.5=2.5 electrons and $N_e=20-x=19.5$); or the extreme $N_e=19$ when the self-doping is $x=1.0$. Note that in the following for our computations we use only certain discrete numbers for $x$ as the studied cluster is finite and the total electron number must be an even integer; moreover the total electron number should hit some magic number so that the ground state wave function of the studied *small cluster* is close-shell and not an open-shell.
The problem how to fix $x$ is a difficult question. If one wants to be sure what is a precise value of $x$, then the best way would be to perform independent, auxiliary *ab-initio* or local density approximation with Coulomb interaction $U$ (LDA+$U$) computations and extract the electronic population on the cation $R$ (in $R$VO$_3$) analogously like it was done in Ref. [@Ros17]. This is however rather expensive. Without such auxiliary *ab-initio* computations one is left with speculations. It seems that for the case of La or Y cation a safe guess is that $x\in[0.0,0.5]$, i.e., all three, or almost all three $5d^16s^2$ valence electrons are transferred to the vanadium octahedron.
Numerical studies {#sec:num}
=================
Computational problems concerning the Jahn-Teller Hamiltonian {#sec:JT}
-------------------------------------------------------------
The important part of the electronic Hamiltonian in perovskites, namely the influence of JT distortions on the electronic structure rarely can be treated in a satisfactory way during the computations. Let us explain what we mean by this statement. An effective Hamiltonian which describes cooperative JT lattice distortions for octahedra in the vanadium perovskites can be assumed in the complicated form which is quadratic in JT distortions and contains in addition the terms $\propto d_{i\nu\sigma}^\dagger d_{i\mu\sigma}^{}$ coupled linearly with JT distortions, for details and explicit (quite complicated) formula, see for instance Ref. [@Mul10]. JT distortions $\{Q_i\}$ ($i=1,\dots,6$) (used notation is the same as in Ref. [@Kan60]) can be treated as quasi-classical continuous variables. There should be appended (to all $Q's$) an additional (extra) subscript $m$ to distinguish between different octahedra which could have (in principle) different, one from another, distortions. Let us remind that (see Ref. [@Kan60]) $\{Q_4,Q_5,Q_6\}$ modes cause tilting (rotations) of the VO$_6$ octahedron. The $Q_2$ mode causes distortion of squares formed by four oxygens (in $ab$ plane a square undergoes distortion into an elongated rhombus), while the $Q_3$ mode causes differences in apical vanadium-oxygens bond lengths (tetrahedral distortion).
In the course of normal computations (when looking for ground state energy minimum) the search for energy minimum due to electronic degrees of freedom must be supplemented with an extra search for the optimal values of continuous classical degrees of freedom ($Q_i$-modes). Then the Hamiltonian becomes intractable, even so for very small clusters, even so if the cooperative pattern of JT distortions is explicitly assumed. Let us remark that assuming cooperative and static pattern of JT distortions (with a certain amount of symmetry) would mean that instead of $Q_{2m},Q_{3m},\dots$ (a lot of separate sets of $Q_{2m},Q_{3m},...$, one set for each individual octahedron $m$) one can consider a single set of $|Q_2|,|Q_3|,\dots$ and the dependence on the octahedron number $m$ within the lattice is realized through alternating plus/minus signs to individual $Q$’s and changing them according to the assumed global symmetry of the static-cooperative JT distortions. Anyway, even with this great simplification there are at least five extra $\{Q_i\}$ variables which makes looking for ground state energy minimum during HF iterations extra expensive.
To overcame this difficulty most often a semiempirical treatment of JT terms is used: namely one assumes an explicit form and the magnitudes of the lattice distortions, usually suggested by the experiment. Thus the distorted lattice is frozen and we take this as an experimental fact (and do not ask any more about the origin of these distortions). Then computations become more feasible. The $Q_{2m},Q_{3m},\dots$ modes and the JT Hamiltonian do not enter computations anymore — their only role was to deform the lattice and to change V$-$O distances. Instead, one collects all V$-$O and O$-$O bond lengths (as suggested by experiment) and because of modified bond lengths one modifies the matrix of kinetic hopping parameters. In this respect quite popular is the Harrison scaling [@Har05] when the difference in V$-$O bond lengths (versus some reference bond lengths, for example those in hypothetical undistorted crystal of ideal cubic symmetry) causes renormalization of the hopping elements. The second important consequence of changed V$-$O distances is creation of local crystal fields acting upon central V-ions: these will split $yz/zx$ doublets as already discussed above for $H_{\rm diag}$ and $f^{cr}_{\mu,\sigma}$.
To simplify the numerical effort, we performed exactly such computations but only for scenario shown in Fig. 2, i.e., only $Q_2$ distortions were included, while the $Q_4$ distortions were neglected. This choice is purely pragmatic: non-zero $Q_4$’s, $Q_5$’s, and $Q_6$’s significantly increase computational effort by drastically lowering the symmetry and therefore increasing the complexity of kinetic hopping matrix. We emphasize that the $d-p$ model is definitely not an *ab-initio* approach thus it can account only for a qualitative description of generic physical properties; one should not expect that all the physical details will be described properly. Therefore certain simplifications in modeling are not a capital offense. In this respect one can still ask if indeed octahedral tilting and finite $Q_4$ distortions are mandatory for spin-orbital order to emerge. Numerous experimental and theoretical papers addressed directly and indirectly these questions: (i) quoting Ref. [@Yan07] where the proof was given than V$-$O$-$V angles deviating strongly from 90 are not primary a driving force stabilizing $C$-type orbital order in vanadates, or that (ii) orbital fluctuations (at zero temperature) are not strong but in fact almost suppressed [@Ray07]. For a more general discussion of these problems see Ref. [@Fuj10]. We suggest that for the description of the onset of spin-and-orbital order, our simplified scenarios with local crystal fields and with geometries depicted in Figs. 2 and 3 are quite enough and that the apical axes non-zero tiltings influence only the distances between the true HF ground state and other (higher in energy) stable HF states.
To summarize, and at the same time to give an explicit example: In YVO$_3$ we studied the zero-temperature geometry as shown in Fig. 2 with repeating layer 1 (along the $c$ axis): the V$-$O bond lengths were set as 2.042, 1.99, and 1.99 Å$\,\,$ [@Bla01; @Kha04] for a long, a short, and an apical bond, respectively. The Slater integrals were scalled following the Harrison’s rules [@Har05] to fit the experimental V$-$O bond lengths. The changes of O$-$O bond lengths caused by JT distortions were neglected (they are expected to be small and less important). On top of it the values of local crystal field splitting of $yz/zx$ doublet were assumed to be $\pm 0.1$ eV.
Unrestricted Hartree-Fock computations
--------------------------------------
We use the unrestricted HF approximation (with a single determinant wave function) to investigate the model (\[model\]). The technical implementation is the same as that described in Refs. [@Ave13; @Miz95; @Miz96; @Sug13; @Ros15; @Ros16] featuring the averages $\langle d^\dagger_{m,\mu,\uparrow}d_{m,\mu,\uparrow}^{}\rangle$ and $\langle p^\dagger_{i,\mu,\uparrow} p_{i,\mu,\uparrow}^{}\rangle$ (in the HF Hamiltonian) which can be treated as order parameters. At the beginning some initial values (a guess) have to be assigned to them. During HF iterations the order parameters are recalculated self-consistently until convergence. If in the course of computations all the averages $\langle d^\dagger_{m,\mu,\uparrow}d_{m,\mu,\uparrow}^{}\rangle,\dots$ would be treated as independent, convergence (if any) would indeed be too slow. Therefore the common strategy is to employ explicit type of symmetry of the order in the ground state (which lowers the number of order parameters) and to perform HF iterations strictly under this assumption. During present computations the chosen scenarios for the ground state symmetry were those with either: (i) orbital order of $G$-type, or $C$-type, or absent; (ii) spin order $G$-AF, or $C$-AF, or absent; (iii) $x$ or $z$ easy magnetization axis. One remark: the hypothetical ground state symmetries which would violate Goodenough-Kanamori rules [@Goode] were also considered (these are: ground state with $C$-AF spin order and $C$-type orbital order and also ground state with $G$-AF spin order and $G$-type orbital order; during computations we found such states to be locally stable in unrestricted HF for some parameters, but they never became true ground state).
Within each of the above scenarios the number of independent order parameters is lowered but still it is large enough so that the HF convergence is rather poor. This was caused mainly by not imposing any restrictions on order parameters associated with oxygens (no orbital equivalence by symmetry, no assumption on oxygens magnetic properties) and not imposing any symmetry restriction on order parameters associated with vanadium $e_g$ orbitals. We found that imposing any of such restrictions is quite risky as any symmetries and orbital equivalences as could be *a priori* assumed, in fact turn out to be too restrictive and only approximate ones. This happens at least for scenarios shown in Figs. 2 and 3. For computations and a quick scan of the phase diagram we used $2\times 2\times 4$ cluster. (A single HF run on an ordinary desktop can be done in about 10 minutes; bigger $4\times 4\times 4$ clusters require from several hours up to one day).
The simplified and popular remedy for poor HF convergence is the so-called dumping technique. Better remedy is the technique known in quantum chemistry and called level shifting [@Sou73]. It is based on replacing the true HF Hamiltonian by a different Hamiltonian — the one with the identical eigenvectors (one particle eigenfunctions) as the original Hamiltonian and with identical *occupied* eigenenergies. The original eigenenergies of virtual states are however uniformly shifted upwards by a fixed constant value. Thus if we apply the shift say by 5.0 eV, then the gap between the highest occupied molecular orbital and the lowest unoccupied molecular orbital (HOMO-LUMO gap) we obtain will be artificially enlarged exactly by 5.0 eV. (So, it has to be corrected by subtracting from the obtained HOMO-LUMO gap the fixed value of 5.0 eV).
When applying virtual level shifting we can obtain some additional information. Namely when the HOMO-LUMO splitting (after correcting for the shift) is negative, then the single-determinant HF ground state we obtained is not correct (this assumes that sufficient number of different HF starting conditions was tried). One possibility is that the true ground state is conducting, another is that a single-determinant HF wave function breaks down due to very strong electronic correlations and multi-configuration HF method is required.
Results and discussion {#sec:res}
======================
Zero-temperature ground state in LaVO$_3$
-----------------------------------------
The symmetry of LaVO$_3$ at zero temperature is monoclinic [@Zub76; @Ren03] which should correspond to $G$-type orbital order (which is induced, or to say it directly, is enforced by cooperative crystal field splittings of $yz/zx$ doublets). The bond lengths at zero temperature were difficult to find in the literature — following Ref. [@Saw96] we took 2.04 Å$\,\,$/1.98 Å$\,\,$ for long/short V-O distances within $ab$ plane and 1.98 Å$\,\,$ for apical V-O bonds. With this choice we assumed the following local crystal field values: 1.8 eV as the distance between $t_{2g}$ levels and $e_g$ levels, and an ad-hock choice: $\pm 0.10$ eV as splitting between $yz/zx$ orbitals, also $xy$ orbital energy is lowered (due to tetragonal distortion of V$-$O apical bonds) by 0.1 eV.
The experimentally found spin-order is $C$-AF with average magnetic moment $|\langle m\rangle|\in(0.6,0.7)$ and easy magnetization axis $c$ [@Zub76; @Tep00]. The estimations of band gaps are in between 1.1 and 1.8 eV [@Tsv04; @Mos09; @Mos10; @Kum17]; the most popular value is 1.1 eV. Below in Table II we collected the obtained results for hypothetical self-dopings $x=0,\,0.5,\,1.0$ (we remind that we do not know which one of these values is closest to the true one). Some comments about the legend in Table II: the indices $m=1$ and 2 in $\langle n_{1,xy,\uparrow}\rangle,\dots$, etc., stand for two nonequivalent vanadium ions, see Fig. 2. $E_{\rm HF}$ is the HF energy per one VO$_3$ unit, $G$ is the HOMO-LUMO gap, $\langle m\rangle$ is an average magnetic moment per V ion (when expressed in $\mu_B$ it should be twice larger). In Table II the spin-order type with an easy magnetization direction is indicated, and finally $x$ is self-doping level which was fixed during computations.
Now we summarize the results obtained for different electronic filling of VO$_3$ octahedra (self-doping). We start with self-doping $x=0$ which stands for an ideal ionic model. The best HF ground state reproduces the experimental spin-orbital order found in LaVO$_3$, see the $x=0$ column of Table II. However, the next candidate for the HF ground state with $C$-AF spin order parallel with the $x$ axis (see Fig. 2), which corresponds to (1,1,0) crystallographic direction and is only by 0.3 meV energetically higher (not shown). Note that when spin-orbit interaction is neglected the change is here insignificant: instead of 0.3 meV we obtained 0.2 meV energy difference. (A general discussion of the role played by spin-orbit interaction in the vanadium perovskites was presented in Refs. [@Hor03; @Yan04; @Zho07]).
For spin order along the $z$ axis site $m=1$ corresponds to magnetization $m_1\simeq 0.99$ and site $m=2$ to $m_2\simeq -0.99$. We observe that when quantum fluctuations are absent as in our calculation, the magnetization is somewhat reduced due to minority-spin electron density in the occupied $t_{2g}$ orbitals, while this reduction is almost fully compensated by majority-spin electron density in the empty $t_{2g}$ and two $e_g$ orbitals. In this way we arrive at $|\langle m\rangle|\simeq 0.99$ which results from electron delocalization by $d-p$ hybridization. It is remarkable that total electron density in $e_g$ orbitals is close to 0.40 which definitely shows that $e_g$ orbitals contribute to the electronic structure. What concerns the average occupation of $2p$ electrons: (i) on oxygens aligned along the $x$ axis (see Fig. 2) it is 5.80 with average moments either 0.0 or $\pm 0.01$ (changing not randomly but in a regular way); (ii) for oxygens aligned along the $y$ axis the corresponding numbers are 5.86 for the charge and 0.0 or $\pm 0.01$ for the moments; (iii) for oxygens aligned along the $z$ axis (this coincides with the crystallographic $c$ direction) the occupation is 5.80 and no moments are found. The next HF stable state is by 1.3 meV higher than the true ground state — it has $G$-AF spin order parallel to the $z$-axis. Note that this state violates Goodenough-Kanamori rules [@Goode].
[lccc]{} $x$ & 0.0 & 0.5 & 1.0\
\
& $G$-AO & $G$-AO &\
& $C$-AF$_z$ & $C$-AF$_z$ & $C$-AF$_x$\
$E_{\rm HF}$ (eV) & 32.555 & 26.573 & 20.673\
$G$ (eV) & 3.92 & 1.99 & 2.89\
$|\langle m\rangle|$ & 0.99 & 0.77 & 0.52\
\
& $C$-AO & $C$-AO &\
& $G$-AF$_z$ & $G$-AF$_z$ & $C$-AF$_x$\
$E_{\rm HF}$ (eV) & 32.734 & 26.802 & 20.935\
$G$ (eV) & 3.99 & 2.02 & 2.86\
$|\langle m\rangle|$ & 0.99 & 0.77 & 0.52\
\
$\langle n_{1,xy,\uparrow} \rangle $ &1.00 & 0.86 & 0.54\
$\langle n_{1,xy,\downarrow} \rangle $ &0.05 & 0.06 & 0.54\
$\langle n_{1,yz,\uparrow} \rangle $ &1.00 & 0.60 &0.12\
$\langle n_{1,yz,\downarrow} \rangle $ &0.04 & 0.06 &0.12\
$\langle n_{1,zx,\uparrow} \rangle $ &0.09 & 0.23 & 0.12\
$\langle n_{1,zx,\downarrow} \rangle $ &0.06 & 0.08 & 0.12\
$\langle n_{1,x^2-y^2,\uparrow} \rangle $ &0.10 & 0.12 &0.13\
$\langle n_{1,x^2-y^2,\downarrow} \rangle $ &0.08 & 0.10 &0.13\
$\langle n_{1,3z^2-r^2,\uparrow}\rangle $ &0.13 & 0.16 &0.18\
$\langle n_{1,3z^2-r^2,\downarrow}\rangle $ & 0.10 & 0.13 &0.18\
$\langle n_{2,xy,\uparrow} \rangle $ &0.05 & 0.06 &0.54\
$\langle n_{2,xy,\downarrow} \rangle $ &1.00 & 0.86 &0.54\
$\langle n_{2,yz,\uparrow} \rangle $ &0.06 & 0.08 &0.12\
$\langle n_{2,yz,\downarrow} \rangle $ &0.09 & 0.23 &0.12\
$\langle n_{2,zx,\uparrow} \rangle $ &0.04 & 0.06 & 0.12\
$\langle n_{2,zx,\downarrow} \rangle $ & 1.00 & 0.60 &0.12\
$\langle n_{2,x^2-y^2,\uparrow} \rangle $ & 0.08 & 0.10 & 0.13\
$\langle n_{2,x^2-y^2,\downarrow} \rangle $ & 0.10 & 0.12 &0.13\
$\langle n_{2,3z^2-r^2,\uparrow}\rangle $ & 0.10 & 0.13 & 0.18\
$\langle n_{2,3z^2-r^2,\downarrow}\rangle $ &0.13 & 0.16 &0.18\
The states with different spin order are almost degenerate. Most probably a more complicated geometry featuring sizable octahedral axes tilting should account for bigger differences, such as those reported in Ref. [@Miz99]. Summarizing, spin-orbital order for $x=0$ is *ideally reproduced* with respect to present paradigm of spin-orbital order in vanadates [@Bla01] but average spin and band-gap we obtained do not agree too well with the experimental values.
Consider now doping $x=0.5$: The best HF ground state we obtained here also reproduces correctly the experimental spin-orbital order found in LaVO$_3$, see the third column of Table II. The next candidate for the ground state is the one with $C$-AF spin order but this time aligned along the $x$ axis (see Fig. 2). Actually, it is by 2.0 meV higher; note that the spin-orbit interaction is here more important and responsible for so large energy difference; when this interaction is absent one finds instead the energy difference of 0.6 meV. Magnetization of $|\langle m\rangle|\simeq 0.77$ corresponds better to the experiment — one finds here definitely weaker magnetization contributions from two occupied $t_{2g}$ orbitals but a larger magnetization in the third $t_{2g}$ orbital. Altogether, electron density in $t_{2g}$ orbitals is lower than that at $x=0$, but at the same time the $e_g$-electron density (but not magnetization) is somewhat enhanced.
The oxygen electron occupations indicate charge delocalization by $d-p$ hybridization in presence of spin-orbit coupling: (i) for oxygens along the $x$ axis electron density is 5.73 while magnetic moments are $\pm 0.01$; (ii) for oxygens along the $y$ axis electron densities of 5.57 and 5.59 are accompanied by $\pm 0.01$ moments (arranged with a suitable regularity, both spins and the tiny charge modulation); (iii) for oxygens along the $z$ axis (this coincides with the crystallographic $c$ axis) occupations are 5.81 with zero moments, or 5.76 with $\pm 0.02$ moments — again both spins and tiny charge density wave are arranged with a suitable regularity. Note that average spin value of 0.77 and the HOMO-LUMO gap of 1.99 eV fit rather well to the experimental results. Thus we suggest that for LaVO$_3$ the self-doping is $x\approx 0.5$ and that the entries from the third column in Table II are a rather faithful description of the experimental situation.
At this point we would like to make a short digression and explain in a more transparent way why a weak ($x=0.5$) self-doping effect is important in LaVO$_3$. It is true that the spin and orbital order for $x=0$ and $x=0.5$ are qualitatively identical. However, the average magnetization (per V ion) is $\sim 1.0$ for the pure ionic model $x=0$ and this is unrealistic. At the same time for $x=0.5$ the average computed magnetization value drops to 0.77 — this is more realistic and quite close to the experimental value. We conclude that self-doping reduces the order parameter by including the covalency effect.
There is also a second argument: the band gap we computed for $x=0.5$ is much closer to the experimental value that band gap we computed for $x=0$. It is well known that Hartree-Fock computations tend to overestimate band gaps. And indeed, for $x=0.5$ we obtained $G\approx 2.0$ eV, while the experimental values indicate $1.1<G<1.8$ eV. However our overestimation of the gap (probably by $\approx$30%) is not that severe as in case of $x=0$ where we obtain $G\approx 4.0$ eV. These two facts clearly suggest that including weak self-doping effect is important for realistic modelling of the vanadium perovskites.
For large self-doping $x=1.0$ orbital order disappears. Only $xy$ orbitals are occupied by approximately one electron, while all the densities in all other ($t_{2g}$ and $e_g$) orbitals are close to 0.25, with somewhat enhanced density of 0.36 in $3z^2-r^2$ orbitals. Note that this large density follows from the delocalization of $2p$ electrons from oxygen ions. The ground state has solely spin $C$-AF order with $x$ easy axis of magnetization. This state contradicts experimental observations and excludes so high self-doping level. No entry in the last column of Table II provides the direct evidence that the spins align indeed along the $x$ axis. To supplement this information we must make another digression. Thus we note that at the $m$-th vanadium ion, $\langle d^\dagger_{m,\mu,\uparrow}d_{m,\mu,\uparrow}^{}\rangle=
\langle d^\dagger_{m,\mu,\downarrow}d_{m,\mu,\downarrow}^{}\rangle$, i.e., the average $z$-th spin component vanishes. Then we inspect the real parts of a subclass of complex order parameters (which we get on convergence from the HF output), namely $\langle d^\dagger_{m,\mu,\uparrow}d_{m,\mu,\downarrow}^{}\rangle$. When the summation over $\mu$ is performed, i.e., if we calculate $Re\big\{\sum_\mu\langle
d^\dagger_{m,\mu,\uparrow}d_{m,\mu,\downarrow}^{}\rangle\big\}$, we obtain the value 0.52 which is just the average spin component along the $x$ direction. The imaginary part of the same sum (here it is zero) corresponds to the average spin component along $y$ direction. This ends our digression.
Zero-temperature ground state in YVO$_3$
-----------------------------------------
The symmetry of YVO$_3$ at zero temperature is orthorhombic [@Bla01; @Ren03]. This corresponds to $C$-AO order accompanied by $G$-AF$_z$ spin order. The bond lengths and average magnetization values were reported in Refs. [@Kaw94; @Nak99; @Bla01; @Ulr03; @Ren03]; band gaps are $1.2-1.6$ eV [@Tsv04; @Kum17].
Our HF results on occupation numbers are virtually the same (two digits accuracy) like those for LaVO$_3$ (shown in Table II). As about spin-order just like it was shown in detail for LaVO$_3$ the $z$ and $x$ easy spin directions are degenerate within 1 meV accuracy (at least for our simplified geometries shown in Figs. 2 and 3). The $T=0$ ground state for YVO$_3$ has $C$-AO order coexisting with $G$-AF$_c$ spin-order and is best reproduced by HF results for self-doping $x\approx 0.5$. For $x=1.0$ we find that the orbital order vanishes.
Zero-temperature ground state in BaVO$_3$
-----------------------------------------
To test how accurately the $d-p$ model works in the vanadium perovskites we decided to test one more completely different case: perovskite quasi-cubic BaVO$_3$ (with V$-$O bonds equal approximately 2.0 Å), which is known to be a conductor [@Nis14] down to $T=0$. This time we cannot use crystal-field splittings as the substance is indeed very close to cubic, all octahedra are undistorted and therefore $t_{2g}$ levels remain unsplit. The other significant difference (with respect to LaVO$_3$) is that Ba cation donates not 3 but 2 electrons into one VO$_3$ unit.
With this input we run our computations only to find that for any doping (including ideal-ionic picture with zero self-doping) and for any starting conditions the obtained HOMO-LUMO gaps (after correcting for virtual level shift) are negative. This is a clear indication that BaVO$_3$ is a conductor in nice agreement with the experimental findings. The same conclusion would be also reached for CaVO$_3$ — though CaVO$_3$ is not quasi-cubic and local crystal fields do split $t_{2g}$ levels. Here the decisive factor is probably not symmetry but the number of electrons transferred from a Ca cation to VO$_3$ unit which is at most 2 (ideal ionic model) or (very likely) much smaller, say within the $(1.0,1.5)$ interval.
Remarks on high temperature ($T>77$ K) ground state of YVO$_3$
--------------------------------------------------------------
First we should clearly state that for $T>0$ K the HF computations of the ground state should not apply directly as we do not know the value of entropy and do not determine the minimum of thermodynamic potential. However just out of curiosity we did them anyway.
The bond lengths and average magnetization values were reported in Refs. [@Kaw94; @Nak99; @Bla01; @Ulr03; @Ren03]; band gaps are $1.2-1.6$ eV [@Tsv04; @Kum17]. The symmetry of YVO$_3$ for $T>77$ K is monoclinic [@Bla01; @Ren03]. Our HF occupation numbers we obtained are very close to those shown in Table II. The symmetry of the obtained ground state is $G$-AO order with $C$-AF$_z$ spin order in accordance with the experiment.
The only disagreement with the experiment is that experimentally [@Ulr03] the easy axis of magnetization is neither in $c$ direction nor it is strictly located in the $ab$ plane; one finds spin components of both types. Such a possibility was not investigated during our computations. However, to be on defensive side, let us remind once more that (like it was shown in detail for LaVO$_3$) the $z$ and $x$ easy spin directions are degenerate within 1.0 meV accuracy (at least for our simplified geometries presented in Figs. 2 and 3).
Summary and conclusions {#sec:summa}
=======================
On some examples we have shown that the $d-p$ model is capable of reproducing spin-orbital order in the vanadium perovskites. The three basic fundamentals leading to non-zero orbital-order are: (i) the electronic configuration of V ions which is close to $V^{3+}$; (ii) non-zero local crystal fields (originating from collective JT deformations) which split $yz/zx$ orbitals; (iii) zero or small self-doping due to cations (i.e., electron donors to the VO$_3$ lattice). With these ingredients orbital order is generic — it comes out correctly for any reasonable Hamiltonian parameter set.
However the question *what kind* of magnetic order accompanies orbital order is more subtle. In particular different spin easy-axis orientations are difficult to find as the states stable in HF (candidates for being true ground state) are almost energetically degenerate. In addition to this problem the stability and the type of dominating magnetic order depends strongly on tiny effects occurring on oxygens: small ($\pm 0.01$) spin modulations and small charge modulations, i.e., $\pm(0.01-0.03)$. If one imposes same additional assumptions (for example the assumption that oxygens in $ab$ planes are nonmagnetic — which may seem to be obvious but which is incorrect) in hope that HF convergence will improve then the order in which the types of magnetic order appear may even come out completely wrong.
The above problem (i.e., how to include tiny magnetization modulations on oxygens) is non-existent for *ab-initio* LDA or LDA with local Coulomb interaction $U$ (LDA+$U$) approaches but at a cost of many-fold increase of computational time and effort. On the other hand, the $d-p$ model is not *ab-initio* and HF computations performed on the $d-p$ model can not reach the level of physical reliability such as the LDA+$U$ does but still for extremely cheap and quick preliminary computations in new perovskite materials with orbital and spin degrees of freedom they are indeed of invaluable help.
Summarizing, the multi-band model considered here reproduces the experimentally observed coexisting $G$-AO and $C$-AF spin order in the vanadium perovskites. We emphasize that the minimal multi-band model for the vanadium perovskites has to include all five $3d$ orbitals on vanadium ions. Electron densities in $e_g$ orbitals are typically even larger than that in the nominally empty third $t_{2g}$ orbital. This redistribution of electron charge follows from rather strong $d-p$ hybridization with two $e_g$ orbitals which contribute to the total electronic charge and magnetization of vanadium ions. Our calculations suggest finite but rather low self-doping of $x=0.5$ in the vanadium perovskites.
We kindly acknowledge support by Narodowe Centrum Nauki (NCN, National Science Center) under Project No. 2016/23/B/ST3/00839.
[\[99\]]{}
T. Mizokawa, D. I. Khomskii, and G. A. Sawatzky, Phys. Rev. B. **60**, 7309 (1999). H. Kawano, H. Yoshizawa, and Y. Ueda, J. Phys. Soc. Jpn. **63**, 2857 (1994). H. C. Nguyen and J. B. Goodenough, Phys. Rev. B **52**, 324 (1995). H. Sawada, N. Hamada, K. Terakura, and T. Asada, Phys. Rev. B **53**, 12742 (1996). H. Nakotte, L. Laughlin, H. Kawanaka, D. N. Argyriou, R. I. Shedon, and Y. Nishihara, J. Appl. Phys. **85**, 4850 (1999). Y. Ren, T. T. M. Palstra, D. I. Khomskii, A. A. Nugroho, A. A. Menovsky, and G. A. Sawatzky, Phys. Rev. B **62**, 6577 (2000). G. R. Blake, T. T. M. Palstra, Y. Ren, A. A. Nugroho, and A. A. Menovsky, Phys. Rev. Lett. **87**, 245501 (2001). R. T. A. Khan, J. Bashir, N. Iqbal, and M. Nasir Khan, Materials Letters **58**, 1737 (2004). G. Khaliullin, P. Horsch, and A. M. Oleś, Phys. Rev. Lett. **86**, 3879 (2001); Phys. Rev. B **70**, 195103 (2004). P. Horsch, G. Khaliullin, and A. M. Oleś, Phys. Rev. Lett. **91**, 257203 (2003). T. Jo, J. Phys. Soc. Jpn. **72**, 155 (2003). Y. Ren, A. A. Nugroho, A. A. Menovsky, J. Strempfer, U. Rütt, F. Iga, T. Takabatake, and C. W. Kimball, Phys. Rev. B **67**, 014107 (2003). Z. Fang and N. Nagaosa, Phys. Rev. Lett. **93**, 176404 (2004). C. Ulrich, G. Khaliullin, J. Sirker, M. Reehuis, M. Ohl, S. Miyasaka, Y. Tokura, and B. Keimer, Phys. Rev. Lett. **91**, 257202 (2003). Y. Otsuka and M. Imada, J. Phys. Soc. Jpn. **75**, 124707 (2006). M. Reehuis, C. Ulrich, P. Pattison, B. Ouladdiaf, M. C. Rheinstädter, M. Ohl, L. P. Regnault, M. Miyasaka, Y. Tokura, and B. Keimer, Phys. Rev. B **73**, 094440 (2006). I. Solovyev, J. Comp. Electronics **10**, 21 (2011).
M. De Raychaudhury, E. Pavarini, and O. K. Andersen, Phys. Rev. Lett. **99**, 126402 (2007). P. Horsch, A. M. Oleś, L. F. Feiner, and G. Khaliullin, Phys. Rev. Lett. **100**, 167205 (2008). R. J. O. Mossanek, M. Abbate, P. T. Fonseca, A. Fujimori, H. Eisaki, S. Uchida, and Y. Tokura, Phys. Rev. B **80**, 195107 (2009). J. Fujioka, T. Yasue, S. Miyasaka, Y. Yamasaki, T. Arima, H. Sagayama, T. Inami, K. Ishii, and Y. Tokura, Phys. Rev. B **82**, 144425 (2010). R. J. O. Mossanek, M. Abbate, T. Yoshida, A. Fujimori, Y. Yoshida, N. Shirakawa, H. Eisaki, S. Kohno, P. T. Fonseca, and F. C. Vicentin, J. Phys. C.: Condens. Matter **22**, 095601 (2010). S. Kumari, S. Paul, and S. Raj, Solid State Commun. **268**, 20 (2017). M. Kim, Phys. Rev. B **97**, 155141 (2018).
J. Fujioka, S. Miyasaka, Y. Tokura, Phys. Rev. B **72**, 024460 (2005); **77**, 144402 (2008).
A. Avella, P. Horsch, and A. M. Oleś, Phys. Rev. B **87**, 045132 (2013).
A. Avella, A. M. Oleś, and P. Horsch, Phys. Rev. Lett. **115**, 206403 (2015); Phys. Rev. B **97**, 155104 (2018). J. Kanamori, J. Appl. Phys. **31**, 14S (1960). J. Sirker, A. Herzog, A. M. Oleś, and P. Horsch, Phys. Rev. Lett. **101**, 157204 (2008). J. B. Goodenough, [*Magnetism and the Chemical Bond*]{} (Interscience, New York, 1963); J. Kanamori, J. Phys. Chem. Solids, **10**, 87 (1959).
K. Rościszewski and A. M. Oleś, Phys. Rev. B **91**, 155137 (2015). K. Rościszewski and A. M. Oleś, Phys. Rev. B **93**, 085106 (2016). K. Rościszewski, P. Piekarz, and A. M. Oleś, Phys. Stat. Sol. B **254**, 1700022 (2017); K. Rościszewski and A. M. Oleś, Acta Phys. Polon. A **133**, 356 (2018).
C. Slater and G. F. Koster, Phys. Rev. **94**, 1498 (1954). J.-Q. Yan, J.-S. Zhou, J. G. Cheng, J. B. Goodenough, Y. Ren, A. Llobet, and R. J. McQueeney, Phys. Rev. B **84**, 214405 (2011). T. Mizokawa and A. Fujimori, Phys. Rev. B **54**, 5368 (1996). L. V. Poluyanov and W. Domcke, J. Chem. Phys. **137**, 114101 (2012).
H. Matsuura and K. Miyake, J. Phys. Soc. Jpn. **82**, 073703 (2013).
L. Du, L. Huang, and X. Dai, Eur. Phys. J. B **86**, 94 (2013).
D. Dai, H. Xiang, and M.-H. Whangbo, J. Comput. Chem. **29**, 2187 (2008).
J. Reul, A. A. Nugroho, T. T. M. Palstra, and M. Grüninger, Phys. Rev. B **86**, 125128 (2012). A. M. Oleś, Phys. Rev. B **28**, 327 (1983).
A. M. Oleś, G. Khaliullin, P. Horsch, and L. F. Feiner, Phys. Rev. B **72**, 214431 (2005).
P. Horsch, *Orbital Physics in Transition-Metal Oxides: Magnetism and Optics*, in: *Handbook of Magnetism and Advanced Magnetic Materials*, edited by H. Kronmüller and S. Parkin, Volume 1: Fundamentals and Theory (J. Wiley & Sons, Hoboken, New York, 2007), p. 164.
T. Mizokawa and A. Fujimori, Phys. Rev. B **51**, 12880 (1995). M. S. Hybertsen, E. P. Stechel, W. M. C. Foulkes, and M. Schlüter, Phys. Rev. B **45**, 10032 (1992). *Electronic Structure and Magnetism of Complex Materials*, Springer Series in Material Science Vol. 54, edited by D. J. Singh and D. A. Papaconstantopoules (Springer Science and Business Media, 2013). B. Magyari-Köpe, S. G. Park, H.-D. Lee, and Y. Nishi, J. Mater. Sci. **47**, 7598 (2012). W. A. Harrison, *Elementary Electronic Structure* (World Scientific, London, 2005).
A. E. Bocquet, T. Mizokawa, K. Morikawa, A. Fujimori, S. R. Barman, K. Maiti, D. D. Sarma, Y. Tokura, and M. Onoda, Phys. Rev. B **53**, 1161 (1996); M. Imada, A. Fujimori, and Y. Tokura, Rev. Mod. Phys. **70**, 1039 (1998). K. A. Müller, *Jahn-Teller Effects in Magnetic Resonance*, in: Properties of Perovskites and other Oxides, edited by K. A. Müller and T. W. Kool (World Scientific, Singapore, 2010). J.-Q. Yan, J.-S. Zhou, J. B. Goodenough, Y. Ren, J. G. Cheng, S. Chang, J. Zarestky, O. Garlea, A. Llobet, H. D. Zhou, Y. Sui, W. H. Su, and R. J. McQueeney, Phys. Rev. Lett. **99**, 197201 (2007). T. Sugimoto, D. Ootsuki, and T. Mizokawa, J. Phys. Soc. Jpn. **82**, 104714 (2013).
V. R. Sounders and I. H. Hillier, Int. J. Quantum Chem. **7**, 699 (1973). V. G. Zubkov, G. V. Bazuev, and G. P. Shveikin, Sov. Phys. Solid State **18**, 1165 (1976); J. Kikuchi, H. Yasuoka, Y. Kokubo, and Y. Ueda, J. Phys. Soc. Jpn. **63**, 3577 (1994). A. E. Teplyakh, A. N. Pirogov, A. Z. Men’shikov, and G. V. Bazuev, Physics of Solid State **42**, 2241 (2000). A. A. Tsvetkov, F. P. Mena, P. H. M. van Loosdrecht, D. van der Marel, Y. Ren, A. A. Nugroho, A. A. Menovsky, I. S. Elfimov, and G. A. Sawatzky, Phys. Rev. B **69**, 075110 (2004). J.-Q. Yan, J.-S. Zhou, and J. B. Goodenough, Phys. Rev. Lett. **93**, 235901 (2004). J.-S. Zhou, J. B. Goodenough, J.-Q. Yan, and Y. Ren, Phys. Rev. Lett. **99**, 156401 (2007). K. Nishimura, I. Yamada, K. Oka, Y. Shimakawa, and M. Azuma, J. Phys. Chem. Solids **75**, 710 (2014).
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'Silicon spin qubits show great promise as a scalable qubit platform for fault-tolerant quantum computing. However, fast high-fidelity readout of charge and spin states, which is required for quantum error correction, has remained elusive. Radio-frequency reflectometry enables rapid high-fidelity readout of GaAs spin qubits, but the relatively large resistances and capacitances of accumulation-mode Si quantum dot devices have made radio-frequency reflectometry challenging in these platforms. In this work, we implement radio-frequency reflectometry in a Si/SiGe quantum dot device with overlapping gates by making minor device-level changes that eliminate these challenges. We demonstrate charge state readout with a fidelity above $99.9\%$ in an integration time of $300~\text{ns}$. We measure the singlet and triplet states of a double quantum dot via both conventional Pauli spin blockade and a charge latching mechanism, and we achieve maximum fidelities of $82.9\%$ and $99.0\%$ in $2.08~\mu$s and $1.6~\mu$s integration times, respectively. We also use radio-frequency reflectometry to perform single-shot readout of single-spin states via spin-selective tunneling in microsecond-scale integration times.'
author:
- 'Elliot J.'
- JJ
- 'John M.'
title: 'Rapid high-fidelity spin state readout in Si/SiGe quantum dots via radio-frequency reflectometry'
---
[^1]
[^2]
Electron spins in gated Si quantum dots are promising qubits because they possess long coherence times, which enable high-fidelity gate operations [@veldhorst2014addressable; @tyryshkin2012electron; @morello2010single; @yoneda2018quantum; @eng2015isotopically; @maune2012coherent; @watson2018programmable; @veldhorst2015two]. In the future, quantum error correction will require a large number of physical qubits and the ability to measure and correct qubits quickly [@fowler2012surface]. The fabrication of Si spin qubits leverages existing commercial technologies, and the production of large numbers of qubits seems within reach. Moreover, current architectures are compatible with one- and two-dimensional qubit arrays [@zajac2016scalable; @Mortemousque2018; @Mukhopadhyay2018]. However, implementing readout methods that are simultaneously fast, high-fidelity, and scalable has been challenging in these systems.
Readout of electron spins in quantum dots is usually performed via spin-to-charge conversion together with an external charge sensor [@field1993; @elzerman2003few; @pla2012single; @zajac2016scalable; @broome2017high; @xue2019benchmarking] or gate-based dispersive sensing techniques [@colless2013dispersive; @pakkiam2018single; @zheng2019rapid; @urdampilleta2019gate; @west2019gate]. Gate-based dispersive sensing does not require an additional charge sensor and is therefore inherently scalable, but it is often less sensitive than charge sensing. Charge sensing is easy to implement, sensitive, and compatible with linear qubit arrays, which have emerged as key elements of near-term spin-based quantum information processors [@zajac2016scalable; @volk2019qubyte; @kandel2019; @qiao2019].
External charge sensors, such as quantum point contacts or quantum dots, can be used for both baseband or radio-frequency (rf) readout. In the former case, high-bandwidth baseband readout can be achieved, but it requires low-noise cryogenic preamplifiers and careful wiring to minimize stray capacitance [@vink2007; @jones2019spin]. In the latter case, rf reflectometry achieves high-bandwidth readout by incorporating the charge sensor into an impedance matching tank circuit [@schoelkopf1998radio; @reilly2007fast; @taskinen2008]. Changes to the electrostatic potential of the charge sensor alter its conductance and therefore generate measurable changes to the reflection coefficient of the circuit. This technique enables fast and high-fidelity readout in GaAs quantum dots [@reilly2007fast; @barthel2009rapid; @barthel2010fast; @higginbotham2014coherent]. It is also easy to implement and enables frequency multiplexing for multi-qubit readout.
Radio-frequency reflectometry has successfully been applied to Si donor-based devices [@villis2014; @hile2015radio]. However, accumulation-mode Si devices present two main challenges to rf reflectometry. First, accumulation-mode devices can incur sizeable capacitances of order $10^{-12}-10^{-11}$ F if large-area accumulation gates are used. This is much larger than typical capacitances in reflectometry circuits, and it can negatively impact the performance of the tank circuit [@taskinen2008]. Second, Si devices often have lower mobilities than GaAs devices. These lower mobilities generate excess resistance in the two-dimensional electron gas (2DEG), diminishing the sensitivity of the charge sensor. Previous work on rf reflectometry in silicon [@angus2008silicon; @yoneda2018quantum] and recent clever circuit modification strategies [@ares2016; @wang2013charge; @volk2019fast] have developed promising approaches to mitigate these challenges.
Here, we implement high-fidelity charge- and spin-state sensing via rf reflectometry in a Si/SiGe quantum-dot device with overlapping gates. We eliminate the problems discussed above by making minor device-level changes. These changes are easy to implement and preserve the scalability of the overlapping-gate architecture. We demonstrate high-fidelity charge and singlet-triplet readout in submicrosecond integration times, and we use rf reflectometry to implement microsecond-scale single-spin readout.
![ **Fig. 1** Device and design. **a** False color scanning electron microscope image of a device with a nominally identical geometry to the one tested. Gates are colored according to their purpose, with accumulation, screening, plunger, and tunneling gates shown in purple, black, blue, and yellow, respectively. The rf reflectometry circuit is connected to an ohmic contact to the 2DEG indicated by a square with an “x” in it. **b** Schematic of the rf reflectometry circuit. The on-chip portion of the circuit is highlighted where the region in pink is the contribution from the 2DEG under the accumulation gate, and represents a distributed network of $R_{S}$ and $c_{g}$ components where $C_g=\sum c_g$ is the total gate capacitance. The region in blue is the contribution from the dot, and $R_d$ is the resistance of the dot. **c** Schematic of the device used showing the underlying 2DEG formed in typical device operation. []{data-label="fig:device"}](f1_v7-eps-converted-to.pdf){width="8.6cm"}
The device in this report (Fig. 1a) is fabricated on an undoped Si/SiGe heterostructure with an 8-nm-thick Si quantum well approximately 50 nm below the surface. Voltages applied to three overlapping layers of aluminum gates are used to confine electrons in up to four quantum dots [@angus2007gate; @zajac2015reconfigurable].
We optimized our device for rf reflectometry through the following empirical design rules. First, the capacitance between the accumulation gate associated with the readout circuit and the 2DEG, $C_G$, should satisfy $C_G < 5 \times 10^{-14}$ F. Second, the total resistance of the path between the doped region associated with the readout circuit and the sensor dot, $R_T$, should satisfy $R_T < 15 R_S$, where $R_S$ is the resistance per square of the 2DEG. Satisfying this condition likely ensures that $R_T \ll R_d$ for typical accumulation gate voltages and densities in Si/SiGe. Here $R_d$ is the resistance of the sensor dot.
To accommodate these design rules, the $n^{++}$ region in the reflectometry circuit extends to within 10 $\mu$m of the sensor dot, which helps to reduce $R_T$ (Fig. 1c). We use a screening gate [@west2019gate], which runs underneath the accumulation gate associated with the reflectometry circuit, RA1, and a mesa etch, to reduce $C_G$ as much as possible. The accumulation gate corresponding to the rf channel has a $12\text{-\ensuremath{\mu}m}^{2}$ area between the screening gate and the dot. The device has a 15-nm-thick Al~2~O~3~ gate-oxide layer and a 30-nm-thick Al~2~O~3~ field-oxide layer. While a thinner gate-oxide layer reduces charge noise [@connors2019], it increases $C_G$. Although a matching capacitor can improve sensitivity of reflectometry in devices with large $C_G$ [@ares2016], the large $R_T$, which is distributed with $C_G$, seemed to prevent success with this approach in our device. See Supplemental Material [@rfrosupp] for further details regarding device design. The device optimization described above represents a relatively simple method to implement rf reflectometry, and it should be widely applicable to most accumulation-mode Si quantum dot devices.
We cool our device in a dilution refrigerator to a base temperature of approximately 50 mK. We tune the gate voltages to form a sensor quantum dot under plunger gate RP1. We apply an rf excitation at $224$ MHz to the ohmic contact corresponding to RA1, which is part of the impedance-matching circuit. The circuit also consists of an 820-nH surface mount inductor and the stray capacitance of the device, $C_{Ext}$ (Fig. 1b). The reflected rf signal is digitized on an Alazar ATS9440 data acquisition card (DAQ). During the tune-up process, we adjust the accumulation gate voltage, which affects both $C_G$ and $R_T$, to optimize the sensitivity of the circuit [@angus2008silicon]. We observe a strong modulation in the reflected signal as the plunger gate sweeps across conductance peaks (Fig. 2a). The Supplemental Material [@rfrosupp] contains further details about the reflectometry circuit.
To perform charge sensing, we set the plunger gate voltage to the side of a conductance peak such that the reflected rf signal is sensitive to small changes in the electrochemical potential of the dot. We then tune the gates on the left side of the device to form a double quantum dot, and acquire charge stability diagrams by measuring the reflected rf signal while varying the voltages on plunger gates LP1 and LP2 (Fig. 2b).
We quantify the charge state readout performance by tuning the left-side double dot to the (1,0) occupancy, where the $(i,j)$ notation refers to the charge configuration with $i(j)$ electrons in the dot under gate LP1(LP2). We adjust the tunnel barrier coupling the dot under LP1 to its reservoir such that the tunneling rate is of order 10 Hz. We set the voltage of the plunger gate directly on the (0,0)-(1,0) transition and acquire a time series of the reflected signal, and we resolve individual charge tunneling events [@zajac2016scalable] (Fig. 2c). We fit a histogram of the data (Fig. 2d) to a function $G(V)=g(V|A_0,\mu_0,\sigma_0)+g(V|A_1,\mu_1,\sigma_1)$ where $$\label{eq:gauss}
g(V|A_i,\mu_i,\sigma_i)=\frac{A_i}{\sqrt{2\pi\sigma_i^2}}e^{-\frac{\left( V-\mu_i \right)^2}{2\sigma_i^2}}$$ is a Gaussian with amplitude, mean, and standard deviation $A_i$, $\mu_i$ and $\sigma_i$, respectively. $i$ indicates the occupation of the dot, and $V$ is the measured voltage.
We define the measurement fidelity associated with occupation $i$ as $$\label{eq:fci}
f_{C,i} = \frac{\int_{V_s}^{V_f}g(V|A_i,\mu_i,\sigma_i)dV}{\int_{-\infty}^{\infty}g(V|A_i,\mu_i,\sigma_i)dV}$$
The integral bounds in Equation \[eq:fci\] are $V_s=-\infty(V_t)$ and $V_f=V_t(\infty)$ for $i=0(1)$, where $V_t$ is the threshold voltage. $V_t$ is chosen to maximize the charge state readout fidelity $$\label{eq:Fc}
F_C = \frac{1}{2}\left(f_{C,0} + f_{C,1} \right).$$
Both $V_t$ and $F_C$ depend on the per-point integration time $T_{int}$. In our device, we achieve a charge state readout fidelity of $F_C=98.8\%$ and signal to noise ratio of $\sfrac{\left| \mu_1-\mu_2\right|}{\frac{\left( \sigma_1+\sigma_2 \right)}{2}}=4.3$ with an integration time as small as $T_{int}=100~$ns (Fig. 2d). By extending the integration time to just $T_{int}=300$ ns, we achieve a charge state readout fidelity $F_C>99.9\%$ \[Fig. 2c\].
![ **Fig. 2** Charge sensing via rf reflectometry. **a** Measurement of conductance peaks of the sensor dot using rf reflectometry. **b** Charge stability diagram of a double quantum dot acquired via rf reflectometry measurement of the sensor dot. A plane has been subtracted to remove cross-talk between the double dot plunger gates and the sensor dot. **c** Plot of $1-F_C$ as a function of integration time $T_{int}$. The inset shows a representative time series with $T_{int}=300~\text{ns}$. **d** Histogram of the time series shown in **c** analyzed with an integration time $T_{int}=100~\text{ns}$ and having a measurement fidelity of $F_C=98.8\%$. []{data-label="fig:chrg"}](f2_v4-eps-converted-to.pdf){width="8.6cm"}
Having demonstrated fast high-fidelity charge sensing, we turn to fast readout of spin states. We observe Pauli spin blockade at the $(4,2)-(3,3)$ transition in this device. (We did not observe spin blockade at the $(1,1)-(0,2)$ transition, likely because of a small valley splitting.) We repeatedly apply a three-step pulse sequence that initializes a random spin state in (3,3) prior to pulsing toward the measurement point, at which point the rf excitation is applied to the sensor dot. When the randomly loaded spin state is a singlet, it can tunnel freely from (3,3) to (4,2). If it is a triplet, it remains blockaded in (3,3) until it either undergoes a spin flip or exchanges electrons with the reservoirs. We vary the position of the measurement point, and plot the average signal acquired at each measurement point in Fig. 3a. A trapezoid indicating the spin blockade region is visible in the (4,2) charge configuration near the interdot transition.
![ **Fig. 3** Singlet-triplet readout via Pauli spin blockade. **a** Average charge sensor signal acquired at each measurement point across many repetitions of a pulse sequence that initializes a random spin state prior to pulsing to the measurement point. A trapezoid near the interdot transition in the (4,2) charge configuration indicates the Pauli spin blockade region. The limits of the color bar have been adjusted such that the signal in the (4,3) and (3,2) charge configurations is saturated so that the spin blockade region is more easily visible. Positions L, R, E, and M, are the ground state initialization, random state initialization, empty, and measure positions, respectively. **b** Difference of the average signals measured at position M after initializing a random state and after initializing a singlet state as a function of measurement time exhibiting a decay with $T_1=11.0~\mu$s. **c** Plot of the singlet-triplet readout fidelity as a function of integration time. A dashed line indicates the integration time at which the maximum fidelity is achieved. **d** Histogram of single shot measurements with an integration time of $T_{int}=2.08~\mu$s and a fidelity of $F = 82.9\%$. The signal acquired on the DAQ is inverted from the signal shown in **a**, resulting in the singlet states having the lower voltage signal. The data shown in **a**, **b**, **c**, and **d** were taken at $B_{ext}=50$ mT, where $B_{ext}$ is the externally-applied in-plane magnetic field. []{data-label="fig:SB"}](f3_v7-eps-converted-to.pdf){width="8.6cm"}
To quantify the singlet-triplet readout fidelity, we perform 10,000 single-shot measurements in which we initialize a random spin state before pulsing to the measurement point in the spin blockade region, and an additional 10,000 measurements in which we instead preferentially initialize a singlet state prior to measurement. The two sequences described above pulse the gates between positions E, R, and M in Figure 3a, and positions E, R, L, and M, respectively. At the measurement point, we acquire a time series of the reflected rf signal for $40~\mu$s for each single-shot measurement. The average difference between these signals as a function of measurement time from 0-40 $\mu$s is shown in Fig. 3b. These data follow a characteristic exponential decay with a relaxation time $T_1=11.0~\mu$s. In fitting this data, we discard the first 2.5 $\mu$s of data at the beginning of each measurement to allow for the circuit to ring up. The value of $T_1$ that we measure is lower than typical spin relaxation times in Si based quantum dots [@zajac2016scalable]. This fast relaxation is likely related to the rf excitation and strong coupling between the dots and reservoirs in this device.
To compute a measurement fidelity, we use Equations (1)-(3) of Ref. [@barthel2009rapid] to fit a histogram of the data from the first pulse sequence discussed above to the sum of two noise broadened peaks with additional terms to account for relaxation [@barthel2009rapid]. We extract the readout fidelity as $F = \frac{1}{2}( F_S +F_T)$, where $F_S$ and $F_T$ are the singlet and triplet readout fidelities [@barthel2009rapid]. We choose the singlet-triplet threshold voltage to maximize the overall fidelity. Despite the enhanced triplet to singlet relaxation, we achieve a maximum fidelity of $F=82.9\%$ in $T_{int}=2.08~\mu$s (Fig. 3d). In this approach, we discard the first 500 ns of data to allow the resonator to ring up. The total measurement time is thus $2.58~\mu$s. In computing the fidelity, we have accounted for spin relaxation during this 500-ns interval.
To improve our readout fidelity, we use a charge latching mechanism [@studenikin2012enhanced; @mason2015metastable; @harvey2018high]. We tune our device such that the tunneling rate between the dot under LP1 and its corresponding reservoir is $\Gamma_1\sim10~\text{MHz}$ and the tunneling rate between the dot under LP2 and its corresponding reservoir is $\Gamma_2\ll\Gamma_1$. In this tuning, we again apply a three-step pulse sequence that loads a random state prior to pulsing to the measurement point, and vary the measurement point (Fig. 4a). The charge latching mechanism allows singlet states to tunnel across to (4,2), but triplet states instead preferentially tunnel to an excited charge state in (4,3). Generally, this technique results in better sensitivity than conventional spin-blockade readout, because the total electron number differs between these states.
We characterize the readout by performing 10,000 single-shot measurements at position M in Fig. 4a after initializing a random state by pulsing to positions E and then R, as well as after initializing a singlet state by pulsing to positions E, R, and then L. We observe a longer decay time of $T_1 = 51.9~\mu$s \[Fig. 4b\], which is likely due to the reduced electron exchange rate with the reservoir connected to the double dot via the slow tunnel barrier. We compute the fidelity as before, but we now subtract an additional mapping error [@harvey2018high] $e_{map}=\frac{1}{T_{int}}\int_{T_0}^{T_0+T_{int}}e^{-\sfrac{t}{T_{L}}}dt$. Here, $T_L \approx 150$ ns is the average tunneling time across the tunnel barrier connecting the dot under LP1 to its reservoir, and $T_0=650 $ ns is the time we discard once the rf excitation is applied to allow the resonator to ring-up and the latching process to take place. $e_{map}$ is the average probability during the integration time that a triplet will not have tunneled to the $(4,3)$ state and will be mistakenly identified as a singlet. Figure 4c shows the fidelity as a function of integration time for this method. We achieve a fidelity $F>98\%$ with an integration time as short as $T_{int}=800~\text{ns}$ \[Fig 4d\], and an improved maximum fidelity of $F=99.0\%$ in $T_{int}=1.65~\mu$s \[Fig 4c\].
![ **Fig. 4** Singlet-triplet readout utilizing a latching mechanism. **a** Average charge sensor signal acquired at each measurement point across many repetitions of a pulse sequence that initializes a random spin state before pulsing to the measurement point. Positions L, R, E, and M, are the ground state initialization, random state initialization, empty, and measure positions, respectively. **b** Difference of the average signals measured at position M after initializing a random state and after initializing a singlet state as a function of measurement time exhibiting a decay with $T_1=51.9~\mu$s. **c** Singlet-triplet readout fidelity as a function of integration time. A dashed line indicates the integration time at which the maximum fidelity is achieved. **d** Histogram of single shot measurements with an integration time of $T_{int}=800~$ns and a fidelity of $F = 98.2\%$. The data shown in **a**, **b**, **c**, and **d** were taken at $B_{ext}=50$ mT. []{data-label="fig:latched"}](f4_v5-eps-converted-to.pdf){width="8.6cm"}
![ **Fig. 5** Single spin readout. **a** Averaged DAQ signal during the single spin readout pulse sequence (pink). The sequence empties the dot, loads an electron with a random spin, and then pulses to the measurement point. Empty, load, and measurement sections of the pulse are separated by dashed gray lines, and are labeled with E, L, and M, respectively, above the plot. A spin bump corresponding to spin up tunneling events is visible. A control pulse that initializes a spin down electron prior to measurement is shown in black. The difference between the control and measurement pulses during the measurement window is shown in the inset. **b** 100 single-shot spin selective tunneling measurements demonstrating individual electron tunneling events. The plunger gate is pulsed to the measurement point at $T=48~\mu$s in the pulse sequence. **c** Two representative single shot traces. The top trace shows a spin-up electron tunneling out and back in as a spin down. The bottom trace shows no spike and indicates a spin-down electron. The traces have been offset for clarity. **d** Histogram of the maximum single-point value of the signal in the range $T=48~\mu$s to $T=66~\mu$s. The data shown in **a**, **b**, **c**, and **d** were taken at $B_{ext}=1.5$ T and a sampling rate of 10 MHz. []{data-label="fig:sst"}](f5_v7-eps-converted-to.pdf){width="8.6cm"}
Having demonstrated fast, high-fidelity singlet-triplet state readout, we now discuss fast single-spin readout via spin-selective tunneling [@elzerman2004single]. Operating the device near the (0,0)-(1,0) transition and with $B_{ext}=1.5$ T, we apply a three step pulse sequence [@elzerman2004single] to plunger gate LP1 which empties and then loads the corresponding dot with a random spin, and then pulses to the measurement point (Fig. 5a). At the measurement point, which is close to the (1,0)-(0,0) transition, a spin-up electron will preferentially tunnel out. Some time later, a spin-down electron will tunnel back in. This brief change in occupancy results in a measurable change in the charge-sensor signal.
We acquire more than 32,000 single-shot measurements using the pulse sequence described above. For each single-shot measurement, we additionally perform a control measurement using a pulse sequence in which we initialize a spin-down electron instead of an electron with a random spin state. Figure 5a shows plots of the average acquired charge sensor signal across all single-shot traces for both the measurement and control pulse sequences. A “spin bump" from the presence of tunneling events corresponding to spin up electrons is visible at the beginning of the measurement window ranging from $T=48-66~\mu$s. The signal from the control pulse shows no spin-bump, as expected. The inset of Fig. 5a shows the difference between the average of the control and measurement pulse sequences in the measurement window. We fit these data to a function of the form $b(T)=A+B(T/\tau)e^{-T/\tau}$, where $A$, $B$, and $\tau$ are a fit parameters, and $T$ is time from the start of the measurement window. We extract a characteristic tunneling time $\tau=3.02~\mu$s. A representative collection of 100 single-shot traces is shown in Figure 5b, and two traces (one spin up and the other spin down) are shown in Fig. 5c.
Figure 5d shows a histogram of the maximum single point value acquired in each single-shot measurement during the measurement window. This histogram shows two distinct peaks corresponding to spin-up and spin-down electrons. The overall acquisition time for each single-shot measurement is only 18 $\mu$s, orders of magnitude faster than usual spin-selective tunneling measurement times [@elzerman2004single]. This increase in speed is enabled by the high bandwidth of the reflectometry circuit.
We have optimized a Si/SiGe quantum dot device with overlapping gates for rf reflectometry by making only modest geometric changes to our device design. The methods we use are applicable to Si devices with and without overlapping gates, can be implemented with relative ease, preserve the scalability of the gate layout, and, importantly, provide the ability to perform rapid high-fidelity charge and spin state readout. We have demonstrated microsecond-scale readout of single-spin states and sub-microsecond singlet-triplet readout. We expect that further improvements are possible via optimization of the dot-reservoir couplings and the sensor dot position. This work presents a feasible solution to achieving rapid and high-fidelity spin-state readout in Si spin qubits that is largely compatible with existing device designs.
During the completion of this manuscript, we became aware of a related result demonstrating similar techniques for the implementation of rf reflectometry in accumulation mode Si devices [@noiri2019radio].
We thank Aaron Mitchell Jones of HRL Laboratories, LLC. for valuable discussions. We thank Lisa F. Edge and Clayton A. Jackson of HRL Laboratories, LLC. for the epitaxial growth of the SiGe material. Research was sponsored by the Army Research Office and was accomplished under Grant Numbers W911NF-16-1-0260 and W911NF-19-1-0167. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the Army Research Office or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation herein. E.J.C. was supported by ARO and LPS through the QuaCGR Fellowship Program.
[48]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\
12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [ ()]{}, [****, ()](https://doi.org/10.1063/1.5025928) [****, ()](https://doi.org/10.1103/PhysRevLett.70.1311) [****, ()](https://doi.org/10.1103/PhysRevB.67.161308) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [ ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [ ()]{}, @noop [****, ()]{} [****, ()](https://doi.org/10.1103/PhysRevApplied.12.014026) [****, ()](https://doi.org/10.1126/science.280.5367.1238) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} [****, ()](https://doi.org/10.1063/1.4883228) @noop [****, ()]{} @noop [****, ()]{} [****, ()](https://doi.org/10.1103/PhysRevApplied.5.034011) @noop [****, ()]{} [****, ()](https://doi.org/10.1021/acs.nanolett.9b02149) @noop [****, ()]{} @noop [****, ()]{} @noop [ ()]{}, @noop @noop [****, ()]{} [****, ()](https://doi.org/10.1103/PhysRevB.92.125434) @noop [****, ()]{} @noop [****, ()]{} @noop [ ()]{}
Supplemental Material {#supplemental-material .unnumbered}
=====================
RF Reflectometry Circuit
------------------------
A full schematic of the circuit diagram used for rf reflectometry measurements is shown in Supplementary Figure 1. The rf carrier is generated at 10 dBm, and the coax through which it travels has a characteristic impedance of $Z_0=50~\Omega$.
{width="8cm"}
Device
------
The device is fabricated on an undoped Si/SiGe heterostructure with an 8-nm-thick Si quantum well approximately 50 nm below the surface and a 4-nm Si cap. We etch away the Si and quantum well underneath the bond pads and under the accumulation gate corresponding to the rf transmission channel area via a dry etch process. The removal of the quantum well under the accumulation gate ensures a minimization of $C_G$, regardless of the presence of or voltage on a screening gate. Source and drain regions near the quantum dots are doped with phosphorus donors. One doped region additionally serves as part of the rf transmission channel, and its geometry is optimized to reduce $R_T$. The device has a 30 nm Al~2~O~3~ field oxide and a 15 nm Al~2~O~3~ gate oxide, both of which are deposited via atomic layer deposition. Three layers of electrostatically isolated overlapping aluminum gates are defined with electron beam lithography and deposited via thermal evaporation [@zajac2015reconfigurable; @angus2007gate]. An appendage of one of the screening gates runs directly underneath the accumulation gate on the non-device side of the doped region overlap. When the device is tuned to form quantum dots, this gate is negatively biased relative to the threshold voltage of the device and it significantly reduces the size of the electron reservoir underneath the accumulation gate. The geometry of the accumulation gate was also optimized such that the area over which it overlaps the electron reservoir is $A= 12\mu\text{m}^2$ and $R_T\sim8\times{}R_{S}$.
Device Design and Performance
-----------------------------
\[fet\] {width="17.2cm"}
Reflection measurements were made with a network analyzer by sending a signal through a room temperature directional coupler into a 4K testing environment. A schematic of the full circuit is shown in Supplementary Figure 2a. Supplementary Figure 2b shows reflection measurements made on a Si/SiGe device not optimized for reflectometry through the ohmic channel. The large $C_{G}$ has a significant impact on the resonance frequency and $R_{T}$ broadens the resonance and causes the reflected signal to be insensitive to changes in $R_{d}$ which is not shown in the figure. Applying a large overdrive voltage to the accumulation gate in order to lower $R_{S}$ was insufficient to improve the device sensitivity with this geometry.
To test the efficacy of the design rules outlined in the main text, we constructed and measured two-layer field effect transistor (FET) devices. The devices are composed of an accumulation gate spanning two ohmic contacts and an electrostatically isolated second depletion gate to vary the channel conductance at fixed accumulation gate voltage and decrease the effective area of accumulation gate. A schematic of these devices is shown in Supplementary Figure 2c. Supplementary Figure 2d shows the response of the reflected signal to the gate voltage in the pinched-off mode and demonstrates the capability of using an accumulation gate for in-situ impedance matching. Supplementary Figure 2e shows the change in the reflected signal as the channel conductance varies with the depletion gate voltage once the accumulation gate has been optimally set, demonstrating good sensitivity of the reflected signal to the channel conductance.
Table I gives relevant design parameters and performance metrics for select devices measured during the design process. In addition to device-level design changes, improvements were made to the signal filtering between 4 QD Device 2 and 4 QD Device 3 in order to decrease noise in the circuit and resulting in a much improved $F_C$ despite a modest improvement in the sensitivity.
[\*5c]{} & & & &\
FET & $\sim130$ & 9 & - & -\
4 QD Device 0 & $\sim5000$ & $\sim30$ & 0 & -\
4 QD Device 1 & 83 & 11 & 5 & -\
4 QD Device 2 & 30 & 8 & 25 & $94.4\%$\
4 QD Device 3 & 12 & 8 & 27 & $>99.9\%$\
[^1]: These authors contributed equally.
[^2]: These authors contributed equally.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'The reflection principle is the statement that if a sentence is provable then it is true. Reflection principles have been studied for first-order theories, but they also play an important role in propositional proof complexity. In this paper we will revisit some results about the reflection principles for propositional proofs systems using a finer scale of reflection principles. We will use the result that proving lower bounds on Resolution proofs is hard in Resolution. This appeared first in the recent article of Atserias and Müller [@atserias-muller] as a key lemma and was generalized and simplified in some spin-off papers [@garlik19; @goos-et-al; @garlik20]. We will also survey some results about arithmetical theories and proof systems associated with them. We will show a connection between a conjecture about proof complexity of finite consistency statements and a statement about proof systems associated with a theory.'
author:
- 'Pavel Pudlák [^1]'
title: |
Reflection principles, propositional proof systems, and theories\
\
[*In memory of Gaisi Takeuti*]{}
---
Introduction
============
This paper is essentially a survey of some well-known results in proof complexity supplemented with some observations. In most cases we will also sketch or give an idea of the proofs. Our aim is to focus on some interesting results rather than giving a complete account of known results. We presuppose knowledge of basic concepts and theorems in proof complexity. An excellent source is Krajíček’s last book [@krajicek-proof], where you can find the necessary definitions and read more about the results mentioned here.
Proof complexity, as we view it, is not only the study of the complexity of propositional proofs, but also the study of first-order theories of arithmetic in connection with questions in computational complexity. In the seminal paper [@cook] Stephen A. Cook defined a theory $PV$ and showed a close connection to the proof system *Extended Resolution,* which is in a well-defined sense equivalent to the more familiar proof system *Extended Frege,* ($EF$). Similar connections have been shown between fragments of bounded arithmetic $T^i_2$, introduced in [@buss], and fragments of the quantified propositional calculus. In this paper we will continue this line of research in connection with the recent progress concerning the computational complexity of proof search in Resolution and some other weak proof systems.
The basic question about proof search is whether short proofs in a propositional proof system $P$ can be efficiently constructed if we know that they exist. More precisely this is the following problem: given a proposition $\phi$ such that there exists a proof of length $m$ in a proof system $P$, can we find a $P$-proof of $\phi$ in polynomial time? If this is possible, we say that $P$ is *automatable*. A weaker property is *weak automatability* where we only want to find a proof in a possibly stronger proof system $Q$. If $P$ is automatable, then it is possible, in particular, to *decide* whether $\phi$ has a $P$-proof of length $\leq p(m)$, or does not have a proof of length $\leq m$ for some polynomial $p$.
One can show that all sufficiently strong proof systems are not weakly automatable iff there exists a disjoint [**NP**]{} pair that is not separable by a set in [**P**]{}. The latter condition follows, for instance, from the hypothesis that [**P**]{}$\neq$[**NP**]{}$\cap$[**coNP**]{}. Therefore we believe that propositional proof systems are not-automatable except for some weak systems. The hardness assumptions about specific functions used in cryptography, such as the discrete logarithm, imply that already depth-$d$ Frege systems, for some small $d$, are not weakly automatable [@bonet-et-al].
Our original motivation for this paper was the recent result of A. Atserias and M. Müller [@atserias-muller] that Resolution is not automatable unless [**P**]{}=[**NP**]{}. The essence of the proof is to define a polynomial reduction from the [**NP**]{}-complete problem SAT to the problem to decide if a CNF formula has a short Resolution refutation. They define the reduction by mapping a CNF formula $\phi$ to a CNF formula $\rho_{\phi,m}$ expressing that $\phi$ has a Resolution refutation of length $m$ where $m=p(|\phi|)$ for a suitable polynomial $p$. Then they prove:
If $\phi$ is satisfiable, then $\rho_{\phi,m}$ has a Resolution refutation of length $q(m)$ where $q$ is some polynomial;
if $\phi$ is not satisfiable, then $\rho_{\phi,m}$ does not have a Resolution refutation of polynomial length. Hence if there were a polynomial time algorithm for constructing Resolution proofs of length at most polynomially longer than the minimal ones, then it could be used to decide the satisfiability of $\phi$.[^2] Our aim here is not to deal with automatizability, but rather address the natural question: for which propositional proof systems the conditions 1. and 2. are satisfied? The first condition is connected with the reflection principle. The reflection principle is the implication: if $\phi$ is provable, then $\phi$ is true. For every propositional proof system $P$, this can be formalized in the propositional calculus by a series of tautologies. It is well-known that some natural proof system, e.g., $EF$, prove these tautologies by polynomial length proofs. It is not difficult to see that if this is the case, then also formulas of the type $\rho_{\phi,n}$ have polynomial length proofs in the same proof system. Concerning condition 2., this is a question whether a proof system $P$ is able to prove superpolynomial lower bounds on $P$-proofs, i.e., its own proofs. The question is widely open, except for Resolution, and we consider it a more important than automatability.
This paper is organized as follows. After two preliminary sections, in Section 3, we will deal with reflection principles. So far only the full reflection principle and the consistency were studied in the context of the propositional calculus. We will introduce an intermediate concept of the *local reflection principle*, which has only been studied in first order theories. Furthermore, this principle splits into two: the local reflection principle for tautologies and the local reflection principle for non-tautologies. With this finer scale in hand, we will revisit some previous results. Using the lower bound on Resolution proofs on unprovability in Resolution from [@atserias-muller] we will show that Resolution proves efficiently the local reflection principle for non-tautologies, but not for tautologies.
We know that many concrete proof systems prove their reflection principle, or at least the local reflection principle for non-tautologies, with polynomial length proofs. So in order to prove hardness of proof search in these systems, it would now suffice to prove lower bounds on proving lower bounds in these proof systems. But this is a nontrivial problem even for fairly weak proof systems and for strong systems one can only hope for some conditional results. We only know that a sufficiently strong proof system is not able to prove superlinear lower bounds on the proofs in a *stronger* proof system.
At the end of Section 3 we will survey results on reflection principles for some concrete proof systems and state open problems.
In Section 4 we will study connections of first order theories and propositional proof system. For a theory $T$ extending some fragments of arithmetic, say $S^1_2$, one can define two associated proof systems. The first one, which may not always exist and which we call the weak proof system of $T$, is the strongest proof system $P$ whose soundness is provable in $T$. The second one, which always exists and we call it the strong proof system of $T$, is the proof systems in which the proofs are the first order proofs in $T$ of the statements that formulas are tautologies. The basic question is which theories have weak proof systems and which proof systems are weak, or strong, proof systems of theories. One can show that $S^1_2+A$, where $A$ is a $\forall\Sigma^b_0$ sentence, always has the weak proof system, and a sufficiently strong proof system is the weak proof system of some theory iff it proves efficiently its reflection principle (more precisely, if $S^1_2$ proves this fact). Using a result about provability of finite consistency statements, Krajíček proved that the strong proof system of $T$ is equivalent to the weak proof system of $T+Con_T$, where $Con_T$ is the statement expressing the consistency of $T$; see [@krajicek-proof]. We will connect the problem whether or not the weak proof system polynomially simulates the strong one with a conjecture about finite consistency statements.
#### Acknowledgment
I would like to thank Emil Jeř' abek, Erfan Khaniki, J' an Pich, and especially Jan Kraj' iček for their remarks on the draft.
Basic concepts
==============
Propositional proof systems
---------------------------
In this paper *”a proof system”* will always mean *a proof system for classical propositional logic*. We will consider propositional proof systems in the sense of Cook and Rackhow [@cook-reckhow]. According to this definition, a proof system is any polynomial time computable function $P$ that maps the set of strings $\{0,1\}^*$ onto the set of propositional tautologies [*Taut*]{}. The set [*Taut*]{}, of course, depends on the chosen basis of connectives and often we will only consider DNF tautologies. We will also study what is provable about proof systems in some first-order theories. Then it is important how the proof system is represented. Proof systems will always be represented by polynomial time algorithms, which we can formalize, say, by Turing machines, but we will also need that the theory in question can “recognize” some properties of the proof system. In order to prove nontrivial facts about proof systems, we will assume some conditions that guarantee that the proof systems are well-behaved. Thus we will also need that these conditions be provable in the theories in which we formalize our arguments about proof systems.
We will prove some general statements about strong proof systems, at least as strong as Extended Frege systems, and survey some results about weak systems, specifically Resolution, Cutting Planes, and bounded depth Frege systems. We assume that the reader knows these basic systems. Unless stated otherwise, we will treat Resolution as a proof system for proving DNF tautologies, rather than refutation system for refuting unsatisfiable CNFs.
Concerning the strong systems, it is more convenient to use Circuit Frege systems instead of the more familiar Extended Frege systems. Circuit Ferege systems are polynomislly equivalent to Extended Frege systems, Substitution Frege systems, Extended Resolution, and the sequent calculus with the extension rule [$ePK$]{}. The advantage of Circuit Frege systems is that one can use formalizations of polynomial time relations by circuits. If we have to formalize relations by formulas, we need additional *extension variables* and *extension axioms*. This is only a minor technicality, but it complicates notation, therefore we will use Circuit Frege proof system in this article.
*Circuit Frege systems,* abbreviated $CF$, as defined by Jeřábek [@jerabek], use circuits instead of formulas and the same rules as a Frege system with and additional rule that enables one to derive from a circuit $C$ a circuit $C'$ that unfolds in the same formula. Thus circuits are viewed as compressed forms of formulas that enable us to represent them more succinctly. Alternatively, we can compress circuits to a canonical incompressible form and identify circuits that have the same canonical form.
It is not very important which representation one uses for representing formally circuits; we will assume that they are represented by straight-line programs and coded by 0–1 sequences. We will use the *length of the sequence coding the circuit* as the measure of the complexity of circuits.
We say that *a proof system $P$ extends a Circuit Frege system,* if $P$ is a a Circuit Frege system augmented with a polynomial time decidable set of sound axiom schemas $A$. Elements of $A$ are formulas or circuits that are tautologies and being schemas means that a proof may use *any substitution instance* of a proposition $\alpha$ from $A$. When we formalize statements about a proof system $P$ that extend a Circuit Frege system in a theory $T$, we will always assume that $P$ is represented in such a way that $T$ proves that $P$ extends $CF$, i.e., $T$ proves that $P$-proofs are sequences of circuits etc.
It is well-known that $CF$ efficiently simulates the substitution rule [@jerabek] and this also holds for extensions of $CF$ by schemas. We will state a special case of this fact for further reference.
Let $P$ be a proof system that extends $CF$. Suppose that $D$ is a $P$-proof of $\alpha(x_1\dts x_n)$ and let $\gamma_1\dts\gamma_n$ be circuits. Then one can construct a $P$-proof of $\alpha(\gamma_1\dts \gamma_n)$ in time polynomial in $|D|+\sum_i|\gamma_i|$. Furthermore, this is provable in $S^1_2$ provided that $P$ is formalized so that $S^1_2$ recognizes $P$ as an extension of $CF$. Given a $P$-proof of $\alpha(x_1\dts x_n)$, we substitute $\gamma_1\dts\gamma_n$ for $x_1\dts x_n$ in all circuits in $P$.
In the context of proof systems extending $CF$, the word *‘proposition’* will mean a Boolean circuit (a special case of which is a formula). We will use it in particular when the circuit is supposed to express some truth.
Efficiently provable tautologies
--------------------------------
In propositional proof complexity theory we are mainly interested in asymptotical bounds. The typical question is whether or not a given sequence of tautologies has polynomial size proofs in a given proof system. We are often also interested whether the proofs can be efficeintly constructed and verified in a weak theory. Thus given a proof system $P$ and a sequence of tautologies $\{\phi_n\}$ we may consider three basic properties of increasing strength:
$\{\phi_n\}$ have $P$-proofs of polynomial length,
$P$-proofs of $\{\phi_n\}$ can be constructed in polynomial time,
$S^1_2$ proves that $\{\phi_n\}$ are provable in $P$. The third condition implies the first two if $\{\phi_n\}$ is constructible in polynomial time, which we will usually assume. In the propositions and definitions in this paper we will use only one of these properties, mostly the third one, which does not mean that they do not have versions with the other two properties. We leave to the reader to prove the other versions when it is possible. We will abbreviate the last two properties by
$P$ $p$-proves $\{\phi_n\}$, and
$P$ provably $p$-proves $\{\phi_n\}$. If $\{\phi_n\}$ is constructible in polynomial time, we will also say that
$P$ $p$-proves $\{\phi_n\}\to\{\psi_n\}$, if $P$-proofs of $\{\psi_n\}$ can be constructed in polynomial time from $\{\phi_n\}$.
Theories
--------
Although some results could be generalized to a large class of theories, we will only consider finite extensions of $S^1_2$, the fragment of Bounded Arithmetic introduced by Buss [@buss]. In $S^1_2$ polynomial time computations have natural formalization. Sometimes it convenient to have terms, rahter than formulas, for polynomial time algorithms. Then we will assume that the theories contain $S^1_2(PV)$ as defined in [@buss]. In this theory every polynomial time function can be represented by a term. We will assume that first-oder proofs are formalized by the standard Hilbert-style calculus.
Bounded formulas are classified into classes $\Sigma^b_i$ and $\Pi^b_i$ according to the number of alternations of bounded existential and universal quantifier, ignoring sharply bounded quantifiers. A sharply bounded quantifier is a quantifier where the variable is bounded by a term of the form $|t|$, with $t$ an arbitrary arithmetical term and $|\dots|$ denoting the length, which is approximately the binary logarithm. Thus sharp bounds restrict the range of quantification to a polynomial size domain. In particular, a $\Sigma^b_1$ is a formula in prenex form with quantifiers of the form $\exists x\leq t$ and $\forall y\leq|s|$, where $t,s$ are terms (or a formula equivalent to it). We will denote by $\forall\Sigma^b_0$ formulas in the prenex form starting with one, or several universal quantifiers, folowed by sharply bounded quantifiers. A $\Pi_1$ formula is a prenex formula starting with unbounded universal quantifiers followe by bounded quantifiers.
We will use two basic results about arithmetical theories.
Let $T=S^1_2+A$ where $A$ is a set of $\Pi_1$ sentences in the language of $S^1_2$, let $\phi(x,y)$ be a bounded formula with two variables, and suppose that $T\vdash\ \forall x\exists y.\phi(x,y)$. Then there exists a term $t(x)$ such that $T\vdash\ \forall x\exists y(y\leq t(x)\wedge\phi(x,y))$.[^3]
Let $T=S^1_2+A$ where $A$ is a set of $\forall\Sigma^b_0$ sentences, let $\phi(x,y)$ be a $\Sigma^b_1$ formula with two variables, and suppose that $T\vdash\ \forall x\exists y.\phi(x,y)$. Then there exists a polynomial time computable function $f(x)$ such that $\N\models\ \forall x.\phi(x,f(x))$. Furthermore, if we extend $T$ with $PV$ to $T(PV)$ (i.e., $S^1_2(PV)+A$), then $T(PV)$ proves $\forall x.\phi(x,t(x))$ for some $PV$ term $t(x)$.
For a true $\forall\Sigma^b_0$ sentence $A$ of the form $\forall x.\alpha(x)$ with $\alpha\in\Sigma^b_0$, one can construct a sequence of polynomial length tautologies $[\![A]\!]_n$ that express that $\alpha(x)$ is satisfied for all $x$ of length $\leq n$. If we express $[\![A]\!]_n$ as a circuit, then it has $n$ variables; if it is a formula, then the number of variables bounded by a polynomial. The following is a fundamental result about theories and propositional proof system.
\[t-cook\] If $S^1_2$ proves a $\forall\Sigma^b_0$ sentence $A$, then $CF$ provably $p$-proves $[\![A]\!]_n$. This theorem has been extended to a number of theories and proof systems (see [@cook-nguyen]), for example to fragments of Bounded Arithmetic and fragments of the sequent calculus for quantified propositions [@krajicek-pudlak-quantified]. We will say more about it in Section 4.
Propositional proof systems vs. theories
----------------------------------------
The last theorem shows that we can view provability of $\forall\Sigma^b_0$ sentences as a uniform way of proving tautologies. This is not the only possible way to represent uniform provability; e.g., the concept of $p$-provability can be viewed as being intermediate between the nonuniform provability and uniform provability. The concept of provably $p$-provability is very close to the uniform provability as defined by provability of $\forall\Sigma^b_0$ sentences in first-order theories. In [@cook], Cook proved the original version of Theorem \[t-cook\] for the equational theory $PV$, which is another intermediate step between propositional proof systems and first-order theories. Gaisi Takeuti introduced restricted types of first-order proofs that were close to propositional proofs. His aim was to use Gödel’s second incompleteness theorem to prove separations between fragments of Bounded Arithmetic. One of such concepts appeared in a joint work with Jan Krajíček [@krajicek-takeuti].
Reflection principles and soundness
===================================
Definitions and general facts
-----------------------------
We are interested in reflection principles in propositional calculus, but we will start with reflection principles for first order theories as a paradigm. (For a survey of reflection pricniples in arithmetical theories, see Smorynski [@smorynski-HB].) The *local reflection principle* for a theory $T$ is the statement $$\mbox{\it for all $x$, if $x$ is a $T$-proof of $\phi$, then }\phi.$$ The principle is called *local* because it is stated for a single proposition. Usually we study a *schema*, i.e., a set of such sentences for all propositions from some class of sentences $\cal C$. Given a class of sentences, we can also state the *uniform reflection principle* for a theory $T$ and for class $\cal C$: $$\mbox{\it for all $x$ and $y$, if $y\in{\cal C}$ and $x$ is a $T$-proof of $\phi$, then $y$ is true.}$$ In order for this principle to be stated in the language of $T$, it must be possible to define the truth of sentences in $\cal C$. By Tarski’s Theorem (which is easily provable using the fix-point lemma) it is not possible to define truth for all formulas. In particular, in an arithmetical theory we can define truth for classes $\Pi_n$, but not for all arithmetical formulas. We will see that this is different in the propositional calculus. What is also different is that in the propositional calculus we can only speak about proofs up to some length.
Now we want to state reflection principles for a propositional proof system $P$ in the propositional calculus. A natural way to do it is first to state it as an arithmetical formula of the form $\forall\Sigma^b_0$ and translate the formula into a sequence of propositions that express the principle for proofs up to length $n$.
For a propositional proof system $P$ and a proposition $\phi$, we will denote by $\mbox{\it LRfn}_{P,\phi}$ a siutable natural formalization of the local reflection principle in $S^1_2$. In principle we can translate $\mbox{\it LRfn}_{P,\phi}$ to the propositional calculus as a sequence of propositions for every length $m$ of a proof, but for typical calculi the proofs are at most exponentially long in the length of the formula, so it does not make sense to state it for larger lengths. We will denote by $\mbox{\it lrfn}_{P,\phi,m}(\vec x)$ a suitable circuit expressing the principle for proofs $\vec x$ of length at most $m$, where $\vec x$ is a string of propositional variables $x_1\dts x_n$ representing a proof. The length of the circuit $\mbox{\it lrfn}_{P,\phi,m}(\vec x)$ is bounded by a polynomial in $|\phi|$ and $m$. We denote by $|\phi|$ the length of the bit representation of $\phi$. If we cannot use circuits in a proof system in which we want to formalize the local reflexion principle, $\mbox{\it lrfn}_{P,\phi,m}$ will be a formula with additional extension variables.
Informally, the *global reflection principle* for a proof system $P$ is the statement $$\mbox{\it for all $x$ and $y$, if $x$ is a $P$-proof of $y$, then $y$ is true}$$ where *‘$y$ is true’* means that $y$ is a tautology. Again we can first state the principle in an arithmetical theory, namely $S^1_2$, and then translate it to the propositional calculus. The relation *‘$x$ is a $P$-proof of $y$’* is formalized by a $\Sigma^b_1$ formula ${\mbox{\it Prf}}_P(x,y)$. The property that a number (or a bit string) encodes a tautology can be easily formalized by a strict $\Pi^b_1$ formula, which we will denote by $Taut(y)$.[^4] Thus the global reflection principle for a proof system $P$ is formalized by the following $\forall\Sigma^b_0$ formula: $${\mbox{\it Rfn}}_P:= \forall x,y(\mbox{\it Prf}_P(x,y)\to Taut(y)).$$ In the rest of the paper we will omit the specification *‘global’* unless we need to stress the difference between global and local principles. It should be noted that the reflection principle for $P$ expresses the *soundness* of $P$; in this paper we will only refer to it by the term *reflection principle.*
In the propositional calculus we can define circuits $\mbox{\it prf}_{m,n}(\vec x,\vec y)$ that express the relation *‘$x$ is a $P$-proof of $y$’* where $|x|=m$ and $|y|=n$, and $sat_n(\vec y,\vec z)$ that express that the circuit encoded by $\vec y$ is satisfied by assignment $\vec z$ where $|y|=|z|=n$ ($z$ may have more bits than the number of variables of the proposition encoded by $x$). The propositional formalization of the reflection principle is the set of circuits $$\mbox{\it rfn}_{P,m,n}(\vec x,\vec y,\vec z):=\mbox{\it prf}_{P,m,n}(\vec x,\vec y)\to sat_n(\vec y,\vec z)$$ for all $m,n\in\N$. Since the reflection principle for $P$ is stated for all circuits, it is the same thing as the *soundness* of $P$.
Using the formalization of provability relation in $P$ we can state the local reflection principle $\mbox{\it lrfn}_{P,\phi,m}(\vec x)$ more explicitly by $$\mbox{\it prf}_{P,m,n}(\vec x,\lceil\phi\rceil)\to\phi,$$ where $n$ is the length of $\phi$ and $\lceil\phi\rceil$ is the bit string representing $\phi$.
If a proof $P$ system satisfies some basic properties, then the global reflection principle implies the local principle using short $P$-proofs.
\[f1\] Suppose that $P$ allows substitution of truth constants and $P$-proofs of propositions $\phi(\vec z)\equiv sat_n(\lceil\phi\rceil,\vec z)$ can be constructed in polynomial time. Then $P$-proofs of $\mbox{\it lrfn}_{P,\phi,m}(\vec x,\vec u)$ can be constructed in polynomial time from proofs of $\mbox{\it rfn}_{P,m,n}(\vec x,\vec y,\vec z)$ where $m=|\phi|$ (polynomial time is in $m$ and $n$).
The condition about $\phi(\vec z)\equiv sat_n(\lceil\phi\rceil,\vec z)$ is not very restrictive provided that we formalize $sat_n$ in natural way. We can prove the equivalence by induction on the complexity of $\phi$. Then looking at this proof we can easily see that it actually gives us a polynomial time algorith to construct a $P$-proof of the equivalence.
The opposite implication in Fact \[f1\] is probably not true in general, but we do not have any example of a proof system with short proofs of all the instances of the local reflection principle and only long proofs of the global reflections principle. On the other hand, in sufficiently strong proof systems, in particular in proof systems that extend $CF$, already one particular instance of the local reflection principle implies the global reflection principle by short proofs. Namely, if we apply the local reflection principle with the formula $\bot$ (the constant representing contradiction), then it expresses the consistency of the proof system: $$\mbox{\it prf}_{P,m,n_0}(\vec x,\lceil\bot\rceil)\to\bot,$$ which is equivalent to $\neg\mbox{\it prf}_{P,m,n_0}(\vec x,\lceil\bot\rceil)$, where $n_0$ is the length of the representation of $\bot$. We will denote this proposition by $con_{P,m}$.
Suppose that $S^1_2$ proves that $P$ is an extension of $CF$. Then $S^1_2$ proves: if $P$ is consistent, then it is sound. Suppose $D$ is a proof a formula $\alpha(\vec x)$. If $\alpha(\vec x)$ is not a tautology, then for some assignment $\vec a$, $\alpha(\vec a)$ is false. Since $P$ is an extension of $CF$, $P$ proves $\alpha(\vec a)$ (with the proof obtained from $D$ by substituting $\vec a)$ and it also proves $\neg\alpha(\vec a)$, which is a contradiction. So if $P$ is consistent, then $\alpha(\vec x)$ must be a tautology.
The above argument is elementary and the construction of the contradiction can be done in polynomial time. Hence the argument can be formalized in $S^1_2$.
\[c3.2\] Let $P$ be an extension of $CF$. Then $P$-proofs of the propositions $\mbox{\it rfn}_{P,m,n}(\vec x,\vec y,\vec z)$ expressing the reflection principle for $P$ can be constructed in polynomial time from $P$-proofs of propositions $con_{P,m}$ expressing the consistency of $P$.
Using our abbreviations, the conclusion can be stated as: $$\mbox{$P$ $p$-proves $\{con_{P,m}\}\to\{\mbox{\it rfn}_{P,m,n}(\vec x,\vec y,\vec z)\}$.}$$ According to the previous proposition, $S^1_2$ proves $$\forall z\neg {\mbox{\it Prf}}_P(z,\lceil\bot\rceil)\to\forall x,y({\mbox{\it Prf}}_P(x,y)\to \forall u.Sat(y,u)).$$ By Buss’s Theorem, there exists a $PV$ term $t$ such that $$S^1_2(PV)\ \vdash\ \forall x,y,u(\neg {\mbox{\it Prf}}_P(t(x,y,u),\lceil\bot\rceil)\to({\mbox{\it Prf}}_P(x,y)\to Sat(y,u)).$$ Hence $CF$-proofs of propositional translations \_[P,m\_1,n\_0]{}(,)(\_[P,m\_2,n]{}(x,y)sat\_n(y,u)) can be constructed in polynomial time, where $\vec\gamma$ is a string of circuits representing $t$. Since $t$ represents a polynomial time computable function, $|t(x,y,u)|$ is polynomial in $|x|,|y|,|u|$, in fact, polynomial in $|x|$ and $|y|$, because $|u|\leq|x|$. This means that $m_1$ is polynomial in $m_2$ and $n$. Hence we get $P$-proofs of (\[e-rf\]) in polynomial time where the polynomial bound is in the length of ${\mbox{\it rfn}}_{P,m_2,n}$. Lemma [\[l-m\]]{} guaranties that we can get $P$-proofs of $\neg{\mbox{\it prf}}_{P,m_1,n_0}(\vec\gamma,\lceil\bot\rceil)$ from instances of $con_{P,n}$ in polynomial time. Finally, we get proofs of the instances the reflection principle for $P$ using modus ponens.
We can furthermore distinguish two special cases of the local reflection principle: the *local reflection principle for tautologies* and the *local reflection principle for non-tautologies*. Consider the local reflection principle for a proposition $\phi$ stated as a disjunction $$\phi\vee\neg\mbox{\it prf}_{P,m,n}(\vec x,\lceil\phi\rceil).$$ Suppose it has a short $P$-proof. If $\phi$ is a non-tautology, we can substitute a falsifying assignment into $\phi$ and thus get a short proof of the fact that $\phi$ does not have proof of length $\leq m$. Now suppose that $\phi$ is a tautology and, moreover, $P$ has the *feasible disjunction property*. The latter property means that given a $P$-proof of a disjuntion of propositions with disjoint sets of variables, one can construct in polynomial time a $P$-proof of one of the disjuncts. So if the reflection principle has a short $P$-proof, then either $\phi$ has a short $P$-proof, or the fact that it does not have proof of length $\leq m$ has a short proof.
We now state these facts formally.
Suppose a proof system $P$ has the property that from a $P$-proof of a disjucntion $\phi(\vec x)\vee\psi(\vec y)$, where $\phi(\vec x)$ is a nontautology, one can derive by a polynomial length $P$-proof of $\psi(\vec y)$, then TFAE
the instances of the local reflection principle for non-tautologies have polynomial length proofs, and
the propositions of the form $\neg\mbox{\it prf}_{P,m,n}(\vec x,\lceil\phi\rceil)$ for non-tautologies have polynomial length $P$-proofs.
Suppose a proof system $P$ has the feasible disjunction property and proves the instances of its local reflection principle for tautologies by proofs of polynomial lengths. Let $\{\phi_n\}$ be a sequence of tautologies such that $|\phi_n|\leq n$, and the shortest $P$-proof of $\phi_n$ has length $\geq m(n)$ where $m(n)\geq n$ is some function. Then $P$ can prove lower bounds on $P$ proofs of $\{\phi_n\}$ of the form $\geq m(n)$ using proofs whose length is polynomial in $m(n)$. This fact was mentioned in [@disjointNP] for the reflection principle. Here we use a weaker assumption, the local reflection principle for tautologies.
Non-automatability of proof search and provable lower bounds
------------------------------------------------------------
In order to use the approach of Atserias and Müller for proving that proof search is hard for $P$,[^5] it suffices to prove for some polynomials $p(x)$ and $q(y)$ for every $n$, that
$\neg\mbox{\it prf}_{P,p(n),n}(\vec x,\lceil\phi\rceil)$ has $P$-proofs of length $q(n)$ when $\phi$ is *not* a tautology and
$\neg\mbox{\it prf}_{P,m,n}(\vec x,\lceil\phi\rceil)$ does not have polynomial length proofs if $\phi$ is a tautology.
The first condition is satisfied when $P$ has polynomial length proofs of the instances of its local reflection principle for non-tautologies. In the next section we will show that there are many natural proof systems that prove their global reflection principle, hence this condition is satisfied for them. We will also mention lower bounds, i.e., condition 2, that have been proved for some weak systems. Whether or not the second condition is satisfied by strong proof systems is not clear. We are only able to prove a weaker statement, which is not sufficient for non-automatibility. First we need a lemma.
Suppose a proof system $Q$ extending $CF$ is not sound. Then it proves every proposition using a linear length proof. Let $Q$ prove $\alpha(\vec x)$ and suppose there is an assignment $\vec a$ that falsifies $\alpha(\vec x)$. Since $CF$ is complete, it proves $\neg\alpha(\vec a)$. Given a $Q$-proof of $\alpha(\vec x)$, substitute $\vec a$ for $\vec x$. Thus we obtain a $Q$-proof of $\alpha(\vec a)$. Hence $Q$ proves a contradiction, from which any proposition can be derived by a $CF$-proof of linear length.
Let $P$ and $Q$ be extensions of $CF$ such that $P$ provably $p$-proves its reflection principle. Let $\{\alpha_n\}$ be a polynomial time computable sequence of tautologies with $|\alpha_n|=n$. Then there exists a constant $c$ such that if $P$ provably $p$-proves a lower bound $>cn$ on $Q$-proofs of $\alpha_n$ for some $n_0$, then $P$ polynomially simulates $Q$. Proving lower bound $cn$ on $Q$-proofs of $\alpha_n$ means provably $p$-proving $\neg{\mbox{\it prf}}_{Q,cn,n}(x,\lceil\alpha_n\rceil)$. The idea of the proof is simple: if $P$ provably $p$-proves a lower bound, then it also must prove the consistency of $Q$, hence also the reflection principle for $Q$. We postpone the proof of this proposition until we elaborate connections between propositional proof systems and first order theories.
There are two weak points in Proposition \[p-unprov\]. First, it only refers to *provable* $p$-provability, while we would like to have mere $p$-provability. Second, it only shows unprovablity for proof systems strictly stronger than $P$. Can any of the two weaknesses be removed?
The impossibility of proving superlinear lower bound resembles the situation with lower bounds on circuit complexity of concrete Boolean functions where we are only able to prove a lower bound $cn$ for some constant slightly larger than $3$. Is it only a coincidence?
Reflection principles in concrete proof systems
-----------------------------------------------
### Resolution and Cutting Planes
We will denote the Resolution proof system by $Res$. Recall that in this article we view Resolution as proof system for proving DNF tautologies. We will use the following important result.
\[l39\] There exist constants $c>1$ and $\delta>0$ such that for every $n$, $m\geq n^c$, and proposition $\phi$, $|\phi|=n$, there is no Resolution proof $$\neg\mbox{\it prf}_{Res,m,n}(\vec x,\lceil\phi\rceil)$$ of length $n^\delta$. This is a generic statement of two different lemmas, because the encoding of Resolution proofs in the two cited papers are different.
There is no subexponential upper bound on the proofs of the $Res$-local reflection principle for tautologies in $Res$. Specifically, $Res$-proofs of the local reflection principle for $\alpha_n$, $|\alpha_n|=n$, have exponential length for [any sequence of tautologies $\{\alpha_n\}$ for which we have an exponential lower bound.]{}
Suppose $$\neg\mbox{\it prf}_{Res,m,n}(\vec x,\lceil\alpha_n\rceil)\vee \alpha_n,$$ where $m\geq n^c$, has a proof of length $\leq S$. By the feasible disjunction property of $Res$, either $\neg\mbox{\it prf}_{Res,m,n}(\vec x,\lceil \alpha_n\rceil)$ has a proof of length $S^{O(1)}$, or $\alpha_n$ has a proof of length $S^{O(1)}$. Since both propositions have only exponentially large proofs, $S$ must be exponential.
Since we do have sequences of tautologies that require exponential size proofs in Resolution, we also have such a lower bound on the *local reflection principle for tautologies*. This implies an exponential lower bound for the *global reflection principle* in resolution. On the other hand, one can construct polynomial length $Res$-proofs of $\neg\mbox{\it prf}_{Res,m,n}(\vec x,\lceil\phi\rceil)$ for non-tautologies $\phi$. This was proved in [@disjointNP] and reproved in [@atserias-muller]. Hence the instances of the local reflection principle for non-tautologies have polynomial length proofs in $Res$. We state it for further reference. \[p39\] There exists a polynomial $p$ such that for every nontautology $\phi$, and $m\geq n$, $\neg\mbox{\it prf}_{Res,m,n}(\vec x,\lceil\phi\rceil)$ has a Resolution proof of length at most $p(m)$. In this proof we will treat Resolution as a refutation system. If $\phi$ is a non-tautology, then $\neg\phi$ is satisfiable. Suppose we have an assignment that satisfies all initial clauses in a refutation of $\phi$. We will gradually prove that every clause in the proof is satisfied, which produces a contradiction, because the last clause is empty. The whole point is that one can express the fact [*“a clause is satisfied by a given assignment”*]{} by a clause; therefore this argument can be performed in Resolution.
In more detail, let the $j$th clause $C_j$ be represented by variables $y_{e,i,j}$ where the intended meaning is that $x_i$ ($\neg x_i$) is present in the clause if $e=1$ (respectively, $e=0$). Then the fact that the clause is satisfied by an assignment $(e_1\dts e_n)$ is expressed by $$y_{e_1,1,j}\vee\dots\vee y_{e_n,n,j}.$$
The lower bound on the reflection principle was proved already in 2004 by Atserias and Bonet [@atserias-bonet]. It is based on an idea different from the one presented at the beginning of this section. We will explain that proof using a new concept.
A disjunction $\phi(\vec p,\vec x)\vee\psi(\vec p,\vec y)$ with $\vec x$ and $\vec y$ disjoint strings of variables is *friendly for a proof system $P$* if
for every assignment $\vec p\vec x:=\vec a\vec b$ that falsifies $\phi$, $\psi(\vec a,\vec y)$ has a polynomial size $P$-proof,
for every assignment $\vec p\vec y:=\vec a\vec c$ that falsifies $\psi$, $\phi(\vec a,\vec x)$ has a polynomial size $P$-proof. More precisely, friendliness should be defined for a sequence of disjunctions $\{\phi_n(\vec p,\vec x)\vee\psi_n(\vec p,\vec y)\}$ and using $p$-provability in $P$ (we leave it to the reader).
We say that $\phi(\vec p,\vec x)\vee\psi(\vec p,\vec y)$ is *semi-friendly* for a proof system $P$ if condition 1. is satisfied, while condition 2. may fail. For proving a lower bound on the reflection principle, it would suffice to use semifriendly disjunctions, but we will use the stronger, more natural concept.
Here are some examples of friendly disjunctions.
Every $p$-provable disjunction in any proof system.
$Clique_{k+1}$-$Coloring_k$ tautology is freindly in Cutting Planes.
Reflection principle is friendly in Resolution.
1\. is trivial.
2\. The tautology has the form $$\neg Clique_{k+1}(K,G)\vee \neg Coloring_k(\chi,G).$$ where $K$ stands for a $k+1$-clique, $G$ for a graph, and $\chi$ for a $k$-coloring. One can show:
For a fixed graph $G$ and clique $K$, $\neg Coloring_k(\chi,G)$ follows from $PHP^{k+1}_k$.
For a fixed graph $G$ and coloring $\chi$, $\neg Clique_{k+1}(K,G)$ follows from $PHP^{k+1}_k$.
3\. The tautology has the form $$\neg\mbox{\it prf}_{P,m,n}(\vec x,\vec y)\vee sat_n(\vec y,\vec z)$$ One can show:
Given a $Res$-proof $\Pi$ of $\phi$, we also have a $Res$-proof of the equivalent formula $sat_n(\lceil\phi\rceil,\vec z)$, because $\phi\equiv sat_n(\lceil\phi\rceil,\vec z)$ is $p$-provable in $Res$.
Given $\phi$ and a falsifying assignment $\vec a$, we get a polynomial size proof of $\neg\mbox{\it prf}_{P,m,n}(\lceil\phi\rceil,\vec y)$ by Proposition \[p39\].
We will first sketch the lower bound on the reflection principle in the Cutting Plane proof system.
The idea is, roughly speaking, to construct a monotone polynomial reduction of the [*Clique-Coloring*]{} disjoint [**NP**]{}-pair to the canonical pair of $CP$, which is, essentially, the pair defined by the reflection principle of $CP$. The reduction is defined by the mapping $$G\ \mapsto\ \neg Clique_{k+1}(\vec x,G).$$ We observe that if $G$ has a $k$-coloring, then we can construct a proof of $\neg Clique_{k+1}(\vec x,G)$ using $PHP^{k+1}_k$. Here we use the fact that the Clique-Coloring disjunction is (semi)friendly in Cutting Planes. Let $p$ be a polynomial bound on such proofs. Thus we map the disjoint [**NP**]{} pair $$\left(\{G\ |\ Clique_{k+1}(\vec x,G) \}, \{G\ |\ Color_k(\vec y,G)\}\right)$$ to the disjoint [**NP**]{} pair $$\left(\{G\ |\ Clique_{k+1}(\vec x,G) \}, \{G\ |\ \neg Clique_{k+1}(\vec x,G) \mbox{ has a $CP$-proof of size }p(|G|)\}\right).$$ From the fact that the first pair cannot be separated by subexponential monotone circuits, we get that the second one also cannot. Using the monotone interpolation for $CP$ we get an exponential lower bound on $CP$ proofs of the disjunctions Clique\_[k+1]{}(x,y)\_[CP,p(n),n]{}(z,Clique\_[k+1]{}(x,y)(y)) that express the disjointness of the latter pair. The expression $\lceil\neg Clique_{k+1}(\vec x,\vec y)\rceil(\vec y)$ represents a circuit that given an assignment $\vec y:=\vec a$ of $0$s and $1$s, produces the string that is a code of the proposition $Clique_{k+1}(\vec x,\vec a)$.[^6] Hence the propositional variables of\
$\neg{\mbox{\it prf}}_{CP,p(n),n}(\vec z,\lceil\neg Clique_{k+1}(\vec x,\vec y)\rceil(\vec y))$ are only $\vec z$ and $\vec y$. Surely, in $CP$ we cannot represent nontrivial circuits, unless we use extension variables and extension axioms, but if we choose a suitable encoding, the circuit will be trivial—just constants and variables. Thus $\lceil\neg Clique_{k+1}(\vec x,\vec y)\rceil(\vec y)$ is simply a string of truth constants and variables $\vec y$.
Hence the disjunction (\[e-negcl\]) is, essentially, the reflection principle restricted to formulas of the form $\neg Clique_{k+1}(\vec x,\vec y)$. To get the special case of the reflection principle formally correct, we only need to replace $\neg Clique_{k+1}(\vec x,\vec y)$ with the polynomially equivalent formula $Sat_n(\lceil\neg Clique_{k+1}(\vec x,\vec y)\rceil,\vec u)$.
$Res$-proofs of the reflection principle of $Res$ have size $\geq 2^{n^\epsilon}$ for some $\epsilon>0$. The idea is to modify the [*Clique-Coloring*]{} tautology so that it becomes *quasipolynomially* friendly in $Res$. This is done as follows:
consider $2k$-cliques vs. $k$ colorings;
add extension axioms for conjunctions up to $\log n$ that do not mix clique variables with coloring variables; this produces an [**NP**]{}-disjoint pair equivalent to the canonical pair;
translate the quasipolynomial proofs of $PHP^{2k}_k$ in $Res(\log)$ into $Res$ proofs with the extension axioms. It does not matter that we only get quasipolynomially friendly disjuction, because the lower bound on the Clique-Coloring tautology is exponential.
The following concept is just a curiosity.
A disjunction $\phi(\vec p,\vec x)\vee\psi(\vec p,\vec y)$ is *strongly friendly* for a proof system $P$ if
it is friendly and
for every $\vec a\in\{0,1\}^n$ such that both $\phi(\vec a,\vec x)$ and $\psi(\vec a,\vec y)$ are tautologies, both formulas are hard for $P$. A *trivial strongly friendly disjunction* is one that has a polynomial size proof and one term is equivalent to the negation of the other.
There exists a nontrivial strongly friendly disjunction for Resolution.
We will use the fact that there are two polynomials $p_1$ and $p_2$ such that
if $\psi$ is a nontautlogy of length $n$ and $m\geq n$, then $\neg{\mbox{\it prf}}_{Res,m,n}(\vec z,\lceil\psi\rceil)$ has a $Res$-proof of length $p_1(m)$, (polynomial upper bound on the local reflection principle for nontautologies, Proposition \[p39\]);
if $\psi$ is a tautology and $m\geq p_2(n)$, then $\neg{\mbox{\it prf}}_{Res,m,n}(\vec z,\lceil\psi\rceil)$ has only exponentially long proofs (Lemma \[l39\]). Let $p$ be the maximum of $p_1$ and $p_2$. In the rest of the proof we will omit the subscript $Res$.
Our strongly friendly disjunction is: $$\neg {\mbox{\it prf}}_{m',n'}(\vec x,\lceil\neg {\mbox{\it prf}}_{m,n}(\vec z,\vec y)\rceil(\vec y))\vee
sat_{n'}(\lceil\neg {\mbox{\it prf}}_{m,n}(\vec z,\vec y)\rceil(\vec y),\vec u),$$ where $m=p(n)$, $n'=|\lceil\neg {\mbox{\it prf}}_{m,n}(\vec z,\vec y)\rceil(\vec y)|$ and $m'=p(m)$. The expression $\lceil\neg {\mbox{\it prf}}_{m,n}(\vec z,\vec y)\rceil(\vec y)$ has a similar meaning as $\lceil\neg Clique_{k+1}(\vec x,\vec y)\rceil(\vec y)$ in the proof of the lower bound on the reflection principle for $CP$. Namely, it is a string of constants and variables $\vec y$ such that if we substitute a code of a formula $\psi$ for $\vec y$, we get a code of the formula $\neg {\mbox{\it prf}}_{m,n}(\vec z,\lceil\psi\rceil)$. The disjunction is a restriction of the reflection formula to a certain type of formulas, hence it is friendly. It remains to prove that if we instantiate $\vec y$ by plugging in some formula $\psi$ and obtaining $$\neg {\mbox{\it prf}}_{m',n'}(\vec x,\lceil\neg {\mbox{\it prf}}_{m,n}(\vec z,\lceil\psi\rceil)\rceil)\vee
sat_{n'}(\lceil\neg {\mbox{\it prf}}_{m,n}(\vec z,\lceil\psi\rceil)\rceil,\vec u),$$ then either one of the two disjuncts is a nontautology, or both are hard for Resolution. We will consider three cases. Since $\psi$ is fixed, we can simplify the second term and get $$\neg {\mbox{\it prf}}_{m',n'}(\vec x,\lceil\neg {\mbox{\it prf}}_{m,n}(\vec z,\lceil\psi\rceil)\rceil)\vee
\neg {\mbox{\it prf}}_{m,n}(\vec z,\lceil\psi\rceil).$$ [*Case (i), $\psi$ is a nontautology.*]{} Then $\neg {\mbox{\it prf}}_{m,n}(\vec z,\lceil\psi\rceil)$ has a proof of length $m'$, hence the first disjunct is a nontautology.
[*Case (ii), $\psi$ is a tautology and has a proof of length $\leq m$.*]{} Then $\neg {\mbox{\it prf}}_{m,n}(\vec z,\lceil\psi\rceil)$ is a nontautology.
[*Case (iii), $\psi$ is a tautology and has no proof of length $\leq m$.*]{} Then
$\neg {\mbox{\it prf}}_{m,n}(\vec z,\lceil\psi\rceil)$ does not have a subexponential proof according to fact 2. above;
since $\neg {\mbox{\it prf}}_{m,n}(\vec z,\lceil\psi\rceil)$ is a tautology, $\neg {\mbox{\it prf}}_{m',n'}(\vec x,\lceil\neg {\mbox{\it prf}}_{m,n}(\vec z,\lceil\psi\rceil)\rceil)$ also does not have a subexponential proof according to fact 2.
It would be interesting to find a nontrivial strongly friendly disjunction defined combinatorially. It could be one of the well known tautologies, such as the Clique-Coloring tautology, or the Broken Mosquito Screen tautology of [@cook-haken]. If one disjunct defines an [**NP**]{}-complete set, then such a pair would give us an alternative, syntax-free proof of non-automatibility of Resolution based on the condition [**P$\neq$NP**]{}.
### Bounded depth Frege systems
We will denote by $F_d$ the bounded depth Frege system formalized by the propositional sequent calculus with formulas restricted to depth $d$. Thus Resolution is $F_0$. The following two statements are probably true:
*$F_{d+1}$ provably $p$-proves the reflection principle for $F_d$,*
*$F_d$ provably $p$-proves the local reflection principle for $F_d$ for non-tautologies.* The first appeared in Beckman et al. [@BPT] with a proof idea. As for the second one, it seems that the proof that [$Res$ provably $p$-proves the reflection principle for non-tautologies for $Res$]{} can be generalized to all systems $F_d$. The reason why believe that these are true fact is that the formulas expressing that a clause is satisfied by an assignment (general in the case of the reflection principle and specific in the case of the local reflection principle) have the appropriate depth, cf. the proof of Proposition \[p39\]. Another result in this vein is:
For every $k\geq 2$, $Res(k+1)$ provably $p$-proves the reflection principle for $Res(k)$, Atserias and Bonet [@atserias-bonet].
It is also possible that other results about Resolution can be generalized to bounded depth Frege systems, but this will certainly require nontrivial work. We state two such generalizations as open problems.
\[p1\] Does $F_d$ $p$-prove its reflection principle for $d\geq 1$? The same question for the local reflection principle for tautologies.
\[p2\] Is it true that for $d\geq 1$, every tautology $\phi$, and $m$ sufficiently large w.r.t. $n$, $\neg\mbox{\it prf}_{F_d,m,n}(\vec x,\lceil\phi\rceil)$ doesn’t have polynomial length $F_d$-proofs?
The negative answer to Problem \[p1\] would show $F_d$ is weaker than $F_{d+1}$ on $DNF$ tautologies. A small superpolynomial separation has been proved by Impagliazzo and Krajíček [@impagliazzo-krajicek],[^7] but a superquasipolynomial separation is still an open problem. A positive answer to Problem \[p2\] would imply that $F_d$ is non-automatable if [**P**]{}$\neq$[**NP**]{} (but it would not refute 1).
To prove a superquasipolynomial separation of $F_d$ and $F_{d+1}$, it would suffice to prove a a superquasipolynomial lower bound on $F_d$-proofs of $\mbox{\it lrfn}_{P,\phi,m}$ for one sequence of tautologies $\phi_n$ that do not have quasipolynomial size $F_d$ proofs, e.g., $PHP_n$: $$\neg\mbox{\it prf}_{F_d,m,(n+1)n}(\vec x,\lceil PHP_n\rceil)\vee PHP_n,$$ but we do not know how to prove lower bounds on the proof even if one only takes the first term of the disjunction. This is an instance of the fundamental problem: *how difficult is it to prove a lower bound on the lengths of proofs?* In particular, we do not know the answer to the following:
Is there a $d\geq 1$ and a sequence of $DNF$ tautologies $\{\alpha_n\}$ such that a superpolynomial lower bound on the Resolution proofs of $\{\alpha_n\}$ can be proved in $F_d$ using polynomial size proofs? More precisely: is there a function $f$ growing more than polynomially such that tautologies $\neg{\mbox{\it prf}}_{Res,f(n),n}(\vec x,\lceil\phi\rceil)$ have polynomial size $F_d$-proofs?
### Frege and Circuit Frege systams
We have already mentioned, Corollary \[c3.2\], that all extensions of $CF$ provably $p$-prove their reflection principles. This result is just an easy generalization of Cook’s proof of this fact for Extended Resolution [@cook]. Buss constructed explicitly polynomial size proofs of the reflection principle for a Frege system in the Frege system [@buss-reflection]. One can also prove this fact using the theory $\mbox{\it V}NC^1$ that is associated with Frege systems, see [@cook-nguyen].
The picture that emerges from what we know about Resolution, bounded depth Frege Systems on one side, and Frege, Circuit Frege systems and their extensions on the other is that these two kinds of proof systems have essentially different properties. However, so far we only know that the strong systems $p$-prove their reflection principles, while the weak ones probably do not. It would be interesting to find more differences. The most interesting question is whether the strong proof systems can prove lower bounds on their proofs. To prove that they cannot, of course, is hard, becasue we cannot prove any lower bounds on them, but it is conceivable that one can prove at least that such lower bounds are not provable in the theories associated with them, i.e., strengthen Proposition \[p-t-unprov\] below.
Theories and proof systems
==========================
We know that the fragments $G_i$ of the quantified propositional calculus prove their reflection principle using polynomial length proofs (see [@krajicek-pudlak-quantified]). We would like to argue that a proof system proves efficiently its reflection principle is rather a rule than an exception. To this end we will study theories associated with proof systems and proof systems associated with theories. Given an arithmetical theory $T$, we can associate two kinds of proof systems with $T$. The first one, which we will call weak, may not always exist. The second one, which we will call strong, is always defined. Our terminology *weak/strong proof system* is new. In [@cook-nguyen] the weak proof system of theory $T$ is called *a proof system associated with $T$*; in [@krajicek-proof] it is called *a proof system corresponding to $T$*. The strong proof system has been defined before ([@krajicek-pudlak-propositional] is, maybe, the first reference), but no name was given to it.
The weak proof systems of theories
----------------------------------
\[d-1\] Let $T$ be a theory and $P$ a proof system. We will say that $P$ *provably $p$-simulates* $T$ if for every $\forall\Sigma^b_0$ sentence $A$, $P$ provably $p$-proves propositions $[\![A]\!]_n$.
We say that $P$ is a *weak proof* system of an arithmetical theory $T$ if
$T$ proves the soundness of $P$, and
$P$ provably $p$-simulates $T$. Given a theory we can always define a proof $P$ sytem that provably $p$-simulates $T$ and vice versa, given a proof system, we can define a theory that proves the reflection principle (soundness) of $P$, but these constructions in general do not ensure that both properties hold simultaneously. The most important fact concerning weak systems of theories is the following theorem.
\[t4.1\] Let $P$ be a weak proof system of a theory $T$. Then $P$ is the strongest proof system whose soundness is provable in $T$, i.e., every proof systems $P'$ whose soundness is provable in $T$ can be polynomially simulated by $P$. Moreover, the latter fact is provable in $S^1_2$. The following argument can be formalized in $S^1_2$. Let $P$ be a weak proof system of a theory $T$. If $T$ proves the soundness of $P'$, which is the sentence ${\mbox{\it Rfn}}_{P'}$, then $P$ proves the propositional instances of ${\mbox{\it Rfn}}_{P'}$, which are propositional formulation of the reflection principle ${\mbox{\it rfn}}_{P',m,n}$. Since we are arguing in $S^1_2$, the $P$-proofs of ${\mbox{\it rfn}}_{P',m,n}$ can be constructed in polynomial time in $n$ and $m$. If we want to prove a proposition $\phi$ in $P$ and we are given a $P'$-proof $D$ of $\phi$, we only need to substitute $\phi$ and $D$ into ${\mbox{\it rfn}}_{P',m,n}$ where $m=|\phi|,n=|D|$.
If $P$ and $P'$ are weak proof systems of a theory $T$, then they are polynomially equivalent.
According to this corollary, a weak proof system for a theory $T$ is determined up to polynomial simulation. We will denote by $P_T$ one of these weak proof systems (when they exist). Note also that Theorem \[t4.1\] gives us an equivalent definition of the weak propositional proof system of a theory.
Let $A\in\forall\Sigma^b_0$. We will denote by $CF+\{[\![A]\!]_n\}$ the Circuit Frege proof system extended with axiom schemas $\{[\![A]\!]_n\}$, which means that the system can use any proposition of the form $[\![A]\!]_n(x_1/\beta_1\dts x_m/\beta_m)$ as an axiom, where $x_1\dts x_m$ are the propositional variables of $[\![A]\!]_n$ and $x_i/\beta_i$ denotes the substitution of a proposition $\beta_i$ for variable $x_i$.
\[t43\]
For every true $\forall\Sigma^b_0$ sentence $A$ sentence, $CF+\{[\![A]\!]_n\}$ is a weak proof system of $S^1_2+A$.
A proof system $P$ that extends $CF$ is a weak proof system of $S^1_2+\mbox{\it Rfn}_P$ iff $P$ provably $p$-proves its reflection principles. 1. First we show that $S^1_2+A$ proves the soundness of $CF+\{[\![A]\!]_n\}$. It proves the soundness of the (substitution instances of) axioms $\{[\![A]\!]_n\}$ using $A$. Then the argument is the same as for $CF$ alone. Since $S^1_2+A$ proves that the rules preserve soundness, $PIND$-$\Pi^b_1$ implies that the proof system is sound.
Now we prove that $CF+\{[\![A]\!]_n\}$ simulates $S^1_2+A$. Let $A:=\forall y.\phi(y)$, where $\phi\in\Sigma^b_0$. Suppose $S^1_2+A$ proves $\forall x.\psi(x)$ for some $\psi\in\Sigma^b_0$. By Buss’s theorem, $$S^1_2\vdash\ \forall x(\phi(t(x))\to \psi(x)),$$ where $t(x)$ represents a polynomial time computable function. Since $CF$ simulates $S^1_2$, we know that $S^1_2$ proves that $[\![\phi(f(x)\to\psi(x)]\!]_n$ are provable in $CF$. Since the substitution instances of $[\![A]\!]_n$ are axioms of $CF+\{[\![A]\!]_n\}$, $S^1_2$ proves that $[\![\phi(f(x))]\!]_n$ are provable in $CF+\{[\![A]\!]_n\}$. Hence it also proves that $\{[\![\psi(x)]\!]_n\}$ are provable in $CF+\{[\![A]\!]_n\}$.
2\. ($\Rightarrow$) If $P$ is a weak proof system of $S^1_2+\mbox{\it Rfn}_P$, then in particular $S^1_2$ proves that $P$ proves $[\![\mbox{\it Rfn}_P]\!]_n$ for all $n$. These propositions are equivalent to the propositions ${\mbox{\it rfn}}_{P,k,l}(\vec x,\vec y,\vec z)$ expressing the reflection principle in the propositional calculus.
($\Leftarrow$) Suppose that $S^1_2$ proves that $P$ proves its reflection principle. Trivially, $S^1_2+\mbox{\it Rfn}_P$ proves the soundness of $P$. It remains to show that $P$ simulates $S^1_2+\mbox{\it Rfn}_P$. The argument is very similar to the one in the proof in part 1 of the simulation of $S^1_2+A$ by $CF+\{[\![A]\!]_n\}$. The difference is that we now use $\mbox{\it Rfn}_P$ instead of $A$, and we use the fact that $P$ proves the propositional instances of the reflection principle instead of $\{[\![A]\!]_n\}$.
\[c4.4\]
For every $\forall\Sigma^b_0$ sentence $A$, $S^1_2+A$ has a weak proof system.
For every proof system $P$ that extends $CF$, $P+\{{\mbox{\it rfn}}_{P,m,n}\}_{m,n}$ is a weak proof system of some theory. Furthermore, $P+\{{\mbox{\it rfn}}_{P,m,n}\}_{m,n}$ provably $p$-proves propositions expressing its reflection principle.
1\. Follows immediately from Theorem \[t43\], 1.
2\. By Theorem \[t43\], 1, $CF+\{{\mbox{\it rfn}}_{P,m,n}\}_{m,n}$ is a weak proof system of $S^1_2+{\mbox{\it Rfn}}_P$. Further, $CF+\{{\mbox{\it rfn}}_{P,m,n}\}_{m,n}$ polynomially simulates $P+\{{\mbox{\it rfn}}_{P,m,n}\}_{m,n}$, because $CF+\{{\mbox{\it rfn}}_{P,m,n}\}_{m,n}$ polynomially simulates $P$; the reverse simulation is trivial. Thus $P+\{{\mbox{\it rfn}}_{P,m,n}\}_{m,n}$ is a weak system of $S^1_2+{\mbox{\it Rfn}}_P$. By Theorem \[t43\], 2, $P+\{{\mbox{\it rfn}}_{P,m,n}\}_{m,n}$ provably $p$-proves its reflection principle.
Now we are almost ready to prove that nonlinear lower bounds on $P$-proofs as stated in Proposition \[p-unprov\] are not provably $p$-provable in $P$. It only remains to prove the unprovability in the theory associated with $P$.
Let $P$ be a weak proof system of a theory $T\supseteq S^1_2$, let $Q$ be a proof system which extends $CF$, and let $\{\alpha_n\}$ be a polynomial time computable sequence of tautologies with $|\alpha_n|=n$. Then there exists a constant $c$ such that for every $n_0$, if $T$ proves a lower bound $>cn$ for $n\geq n_0$ on $Q$-proofs of $\alpha_n$, then $T$ proves the soundness of $Q$, hence $P$ polynomially simulates $Q$.[^8] Lemma \[l5.1\] can be formalized in $S^1_2$. Thus there exists a constant $c_1$ depending only on $Q$ such that $T$ proves:
[*if $Q$ is not sound, then there exists an $r$ such that all propositions have $Q$-proofs of length $\leq c_1n+r$.* ]{} Let $c>c_1$ and suppose $T$ proves a lower bound $>cn$ for $n\geq n_0$ on $Q$-proofs of $\alpha_n$. Then $T$ proves that for every $r$ there exists an $n$ such that $cn\geq c_1n+r$. Hence according to (\*), $T$ proves that $Q$ is sound.
Let proof systems $P$, $Q$ and tautologies $\{\alpha_n\}$ be given, and suppose $P$ provably $p$-proves its reflection principle and a lower bound $>cn$ for $n\geq n_0$ on $Q$-proofs of $\alpha_n$. Let $T$ be $S^1_2+{\mbox{\it Rfn}}_P$. By Theorem \[t43\], $P$ is a weak proof system of $T$. Our assumption is that $S^1_2$ proves that $P$ proves a lower bound $>cn$ for $n\geq n_0$ on $Q$-proofs of $\{\alpha_n\}$. Since $T$ proves the soundness of $P$, it also proves a lower bound $>cn$ for $n\geq n_0$ on $Q$-proofs of $\{\alpha_n\}$. Hence, by the previous proposition, $P$ polynomially simulates $Q$.
One can prove unprovability of superpolynomial lower bounds on $EF$ in $PV_1$, which is a theory slightly weaker than $S^1_2$ and for which $EF$ is the weak proof system.
For every function $f:\N\to\N$ such that for every $c\in\N$, $PV_1$ proves that $f(n)$ eventually dominates $n^c$, $PV_1$ does not prove the sentence $$\forall x\exists y (|x|<|y|\wedge Taut(y)\wedge\forall z(|z|\leq f(|y|)\to\neg{\mbox{\it Prf}}_{EF}(z,y))).$$ This can certainly be generalized to stronger theories and proof systems, but we do not see a way to derive from it a statement about unprovability in propositional proof systems.
The strong proof systems of theories
------------------------------------
Let $T$ be a consistent theory extending Robinson’s Arithmetic with the set of axioms decidable in polynomial time. The *strong proof system of* $T$, denoted by $Q_T$, is the propositional proof system where $d$ is a proof of a proposition $\phi$, if $d$ is a $T$-proof of the sentence $Taut(\lceil\phi\rceil)$. Since Robinson’s Arithmetic proves all true bounded sentences and $T$ is consistent, $Q_T$ is complete and sound. There is an *alternative definition* that works for all theories that have infinite models. This is based on a more direct translation of propositional formulas into first-order sentences. Let $\phi(p_1\dts p_n)$ be a propositional formula where $p_1\dts p_n$ are the propositional variables of $\phi$. We define its first-order translation $tr_\phi$ by taking $n+1$ distinct first-order variables $x_1\dts x_n,y$ and putting $$tr_\phi:= \forall x_1\dots x_ny.\phi(x_1=y\dts x_n=y).$$
[*Exercise.*]{} Construct polynomial size proofs of the sentences $tr_\phi\equiv Taut(\lceil\phi\rceil)$ in Robinson’s Arithmetic.
Let $T$ be theory such that $P_T$ is defined. Then $Q_T$ polynomially simulates $P_T$.
As $T$ proves ${\mbox{\it Rfn}}_P$, given a proposition $\phi$ and its $P$-proof, we can construct in polynomial time a $T$-proof of $Taut(\lceil\phi\rceil)$.
One reason for calling $Q_T$ strong is the following fact.
\[f3\] $T$ does not prove the soundness of $Q_T$.
By Gödel’s 2nd incompleteness theorem. This fact, however, does not exclude the possibility that $Q_T$ is not strong, e.g., that it is polynomially equivalent to $P_T$, although we consider this possibility unlikely. On the contrary, it seems reasonable to conjecture that for every theory $T$ that possesses a weak proof system, $Q_T$ is strictly stronger than $P_T$. The following corollary was suggested by J. Pich.
$T$ does not prove superlinear lower bounds on $Q_T$-proofs of any sequence of tautologies $\{\alpha_n\}$.
From Fact \[f3\] and Proposition \[p-t-unprov\].
Consistency statements
----------------------
For a theory $T$ whose set of axioms is polynomial time decidable, we denote by $Con_T(x)$ an arithmetical sentence that formalizes the statement that there is no proof of contradiction in $T$ whose length is $\leq x$. In this formula, $x$ is a variable ranging over natural numbers. Thus the consistency of $T$ can be expressed by $\forall x.Con_T(x)$. For a fixed $n\in\N$, we denote by $\bar n$ a closed arithmetical term of length $O(\log n)$ whose value is $n$. Hence $Con_T(\bar n)$ is a sentence of length $O(\log n)$ formalizing the statement that there is no proof of contradiction of length $\leq n$.
\[t4.5\] Let $T$ be a finitely axiomatized sequential theory. Then there exists a polynomial time computable sequence of $T$-proofs $\{D_n\}$ such that $D_n$ is a proof of $Con_T(\bar n)$ for $n=1,2,\dots$. Moreover, it is provable in $S^1_2$ that for all $n$, $D_n$ is a $T$-proof of $Con_T(\bar n)$. In [@finitistic; @improved] we constructed a sequence of $T$-proofs $\{D_n\}$ of $Con_T(\bar n)$ of polynomial length in $n$. One can easily check that the construction can be done in polynomial time and this is also provable in $S^1_2$. In contrast to Theorem \[t4.5\], if $\{E_n\}$ is a sequence of $T$-proofs of $Con_{T+Con_T}(\bar n)$, then this fact is not provable even in $T+Con_T$.
Let $T$ be a consistent computably axiomatized theory containing $S^1_2$. Then for no sequence $\{E_n\}$, $T+Con_T$ proves that for all $n$, $E_n$ is a $T$-proof of $Con_{T+Con_T}(\bar n)$. By way of contradiction, suppose $\{E_n\}$ is such a sequence. Since $T+Con_T$ proves the consistency of $T$, it also proves the uniform $\Pi_1$ reflection principle for $T$. Since it proves that $\{E_n\}$ is a sequence of $T$-proofs of $Con_{T+Con_T}(\bar n)$, it also proves, using the reflection principle, $\forall x.Con_{T+Con_T}(x)$, which is equivalent to $Con_{T+Con_T}$. This is in contradiction with the Second Incompleteness Theorem.
If $T$ contains $S^1_2$ and $T+Con_T$ is consistent, then there exists a computable sequence $\{E_n\}$ of $T$-proofs of $Con_{T+Con_t}(\bar n)$. It is a formalization of the brute-force search for contradiction and the proofs have exponential lengths in $n$. We have conjectured that there are no such proofs of polynomial length, see [@finitistic]. A slightly weaker conjecture says that such proofs cannot be constructed in polynomial time. In the next section we will show a link with the conjecture that the strong system of a theory $T$ is strictly stronger than the weak one.
Strong vs. weak proof systems
-----------------------------
\[t3.1\] Let $T=S^1_2+A$, where $A$ is a true $\Pi_1$ sentence. Then $Q_T$ is a weak proof system of $T+Con_T$, i.e., $$Q_T\equiv P_{T+Con_T}.$$ Clearly, $T+Con_T$ proves the consistency of $T$, hence also the soundness of $Q_T$. It remains to prove that provably in $S^1_2$, $Q_T$ simulates $T+Con_T$. Let a $\Sigma^b_0$ formula $\theta(x)$ be given and suppose that $T+Con_T$ proves $\forall x.\theta(x)$. Then $$T\vdash(\forall y.Con_T(|y|))\to\forall x.\theta(x).$$ By Parikh’s theorem, there exists a polynomial $p$ such that $$T\vdash \forall x\exists y(|y|\leq p(|x|)\wedge (Con_T(|y|)\to\forall x.\theta(x))).$$ This implies $$T\vdash \forall x(Con_T(p(|x|))\to\theta(x)).$$ Since $T\vdash\ u\leq v\to (Con_T(v)\to Con_T(u))$, we have $$T\vdash \forall x(Con_T(p(|x|))\to\forall z(|z|\leq|x|\to\theta(z)).$$ Let $D$ be a $T$-proof of this sentence. By Theorem \[t4.5\], provably in $S^1_2$, there exists an algorithm that in time polynomial in $n$ constructs a $T$-proof $D_n$ of $Con_T(n)$. Thus $S^1_2$ proves that for all $n$, $D+D_{p(n)}$ is a $T$-proof of $\forall z(|z|\leq \bar n\to\theta(z))$. We also have $$T\vdash\forall z(|z|\leq \bar n\to\theta(z))\equiv Taut([\![\theta]\!]_n)$$ with a polynomial size proof and provably in $S^1_2$. Thus $Q_T$ provably $p$-simulates $T+Con_T$.
Let $T=S^1_2+A$, where $A$ is a $\forall\Sigma^b_0$ true sentence, and let $S=T+Con_T$. Then TFAE:
$T$-proofs of $Con_S(\bar n)$ can be constructed in time polynomial in $n$.
The weak system of $S$ polynomially simulates the strong system of $S$.
We will use the observation that $Con_S(\bar n)$ is equivalent to the statement that $con_{Q_S,n}(\vec x)$ is a tautology, i.e., S\^1\_2 Con\_S(|n) Taut(con\_[Q\_S,n]{}).
1$\Rightarrow$2. Suppose $T$-proofs of $Con_S(\bar n)$ can be constructed in polynomial time. This is equivalent to constructing $Q_T$-proofs of $con_{Q_S,n}$ in polynomial time. By Corollary \[c3.2\] we get $Q_T$-proofs of the reflection principle for $Q_S$ in polynomial time. Hence, given a $Q_S$-proof of some $\phi$, we get a $Q_T$ proof of $\phi$ in polynomial time. Thus $Q_T$ polynomially simulates $Q_S$. By Theorem \[t3.1\], this implies that $P_S$ polynomially simulates $Q_S$.
2$\Rightarrow$1. By Theorem \[t4.5\], there exists an algorithm that in time polynomial in $n$ constructs $S$-proofs $Con_S(n)$. Hence by (\[e-con\]) above, one can construct in polynomial time $Q_S$-proofs of $Taut(con_{Q_S,n})$. Assuming $P_S$ polynomially simulates $Q_S$, we get $P_S$-proofs of $Taut(con_{Q_S,n})$. By Theorem \[t3.1\], this implies that we get $Q_T$-proofs of $Taut(con_{Q_S,n})$ in polynomial time. Using (\[e-con\]) once again, we get $T$-proofs of $Con_S(\bar n)$.
One can easily check that a version of this corollary with polynomial lengths of proofs instead of polynomial time algorithms is also true. It seems that in general $Q_T$ is *much* stronger than $P_T$. Let’s consider an example.
Let $T$ be $S^1_2$. Then, by [@cook; @buss], $P_T$ is the Extended Frege proof system. It is well-known that $S^1_2$ interprets the entire bounded arithmetic $T_2$ on an initial segment of the natural numbers. $T_2$ proves the soundness of all fragments $G_i$ of the quantified propositional sequent calculus [@krajicek-pudlak-quantified]. From this, one can easily deduce that the strong system of $S^1_2$ polynomially simulates all $G_i$s. But in fact, the strong system of $S^1_2$ simulates (apparently) much stronger proof systems. Krajíček defined a construction that, from a proof system $P$, produces the *implicit* $P$, denoted by $iP$, and this can be iterated (see [@krajicek-implicit]). The implicitation construction seems to always produce a stronger system. E.g., $iEF$ polynomially simulates $G$, the quantified propositional sequent calculus, while we believe that $EF$ does not. The strong proof system of $S^1_2$ polynomially simulates all iterated implicitations $i_kEF$ of $EF$. This is because
$I\Delta_0+Exp$ proves the soundness of all $i_kEF$, and
if $I\Delta_0+Exp$ proves a $\forall\Sigma^b_0$ sentence $\phi$, then there exists a formula $\alpha(x)$ with one free variable that defines an initial segment of numbers closed under the successor, $+$, and $\times$ in $S^1_2$ such that $S^1_2$ proves the relativization of $\phi$ to $\alpha$; see [@hajek-pudlak].
Characterize the strong proof system of $S^1_2$.
We conjecture that $i_\infty EF$, a suitably formalized union of all $i_kEF$s, is polynomially equivalent to the strong proof system of $S^1_2$. We do have some ideas how to prove this conjecture, but for the time being it is an open problem.
[11]{}
A. Atserias and M. L. Bonet: *On the automatizability of Resolution and related propositional proof systems.* Information and Computation 182(2) (2004), 182-201.
A. Atserias, M. Müller: *Automating Resolution is NP-Hard.* Proc. 60th Annual IEEE Symp. on Foundations of Computer Science, (2019), 498-509.
A. Beckmann, P. Pudlák, N. Thapen: *Parity games and propositional proofs.* ACM Transaction on Computational Logic 15(2), article 17, 2014.
, [ *Non-automatizability of bounded-depth [F]{}rege proofs*]{}. Comput. Complexity, 13 (2004), 47–68.
S. R. Buss: *Bounded Arithmetic.* Bibliopolis, Naples, Italy, 1986.
S. R. Buss: *Propositional consistency proofs.* Annals of Pure and Applied Logic 52 (1991), 3-29.
*Feasibly constructive proofs and the propositional calculus.* In Proc. 7th Annual ACM Symp. on Theory of Computing (STOC) (1975), 83–97.
S. A. Cook and A. Haken: *An exponential lower bound for the size of monotone real circuits.* J. Computer and System Science 58(2) (1999), 326-335.
S. A. Cook and P. Nguyen: *Logical Foundations of Proof Complexity.* Cambridge Univ. Press, 2009.
S.A. Cook, R. A. Reckhow: *The Relative Efficiency of Propositional Proof Systems.* J. Symbolic Logic 44(1) (1979), 36–50.
M. Garl' ik: *Resolution Lower Bounds for Refutation Statements.* 44th Symp. on Mathematical Foundations of Computer Science (MFCS 2019), Leibniz International Proceedings in Informatics, Vol. 138 (2019), pp. 37:1-37:13.
M. Garl' ik: *Failure of Feasible Disjunction Property for k-DNF Resolution and NP-hardness of Automating It*, preprint March, 2020.
M. Göös, J. Nordström, T. Pitassi, R. Robere, D. Sokolov, and S. F. de Rezende: *Automating Algebraic Proof Systems is NP-Hard.* Electronic Colloquium on Computational Complexity (ECCC) 27: 64 (2020).
P. Hájek and P. Pudlák: *Metamathematics of first order arithmetic.* Springer-Verlag, 1993.
A. Haken: *The Intractability of Resolution.* Theor. Comput. Sci. 39: 297-308 (1985)
R. Impagliazzo and J. Krajíček: *A note on conservativity relations among bounded arithmetic theories.* Math. Logic Quarterly 48(3) (2002), 375-377.
E. Jeřábek: *Dual weak pigeonhole principle, Boolean complexity, and derandomization,* Annals of Pure and Applied Logic 129 (2004), 1-37.
J. Krajíček: *Implicit proofs.* J. of Symbolic Logic, 69(2), (2004), 387-397.
J. Krajíček: *Proof Complexity.* Cambridge Univ. Press, 2019.
J. Krajíček, P. Pudlák: *Propositional proof systems, the consistency of first order theories and the complexity of computations,* J. of Symbolic Logic Vol.54, No.3, 1989, pp.1063-1079.
J. Krajíček, P. Pudlák: *Quantified propositional calculi and fragments of bounded arithmetic,* Zeitschrift für Math. Logik 36 (1990), 29-46.
J. Krajíček, P. Pudlák: *Propositional provability and models of weak arithmetic.* In: Proc. Computer Science Logic’89, Eds. Borger, Kleine-Buning, Richter, Springer-Verlag LNCS 440, 1990, 193-210.
J. Krajíček and G. Takeuti: On induction-free provability. Annals of Mathematics and Artificial Intelligence 6 (1992), 107-126.
R. Parikh: *Existence and feasibility in arithmetic.* J. Symbolic Logic 36 (1971), 494-508.
P. Pudlák: *On the length of proofs of finitistic consistency statements in first order theories.* In: Logic Colloquium 84, North Holland P.C., 1986 pp.165-196.
P. Pudlák: *Improved bounds to the length of proofs of finitistic consistency statements.* In: Contemporary mathematics Vol.65, 1987 pp.309-331.
P. Pudlák: *On reducibility and symmetry of disjoint NP-pairs.* Theor. Comput. Science 295 (2003), 323-339.
C. Smorynski: *The incompleteness theorem.* In Handbook of Mathematical Logic, ed. J. Barwise, North-Holland 1989, 821-866.
[^1]: The author is supported by the project EPAC, funded by the Grant Agency of the Czech Republic under the grant agreement no. 19-27871X, and the institute grant RVO: 67985840.
[^2]: Atserias and Müller prove in fact a stronger lower bound in 2., namely there is no subexponential proof of $\rho_{\phi,n}$.
[^3]: This is not the original form of Parik’s Theorem and it is certainly not the most general form.
[^4]: A strict $\Pi^b_1$ formula is a formula that starts with one, or several, bounded universal quantifiers followed by a $\Sigma^b_0$ formula, i.e. formula with only sharply bounded quantifiers.
[^5]: This is certainly not the only possible way of proving hardness of proof search. The formulas used in the spin-off papers that proved hardness of proof search for proof systems stronger than resolution still used $\neg\mbox{\it prf}_{Res,p(n),n}(\vec x,\lceil\phi\rceil)$, it only was suitably lifted.
[^6]: In logic various different notations are used for the Gödel numbers as functions of some variables. E.g., Smorynski [@smorynski-HB] would write $\lceil\phi(x,\dot{y})\rceil$ for what we would denote by $\lceil\phi(\vec x,\vec y)\rceil(\vec y)$.
[^7]: In [@impagliazzo-krajicek] the result is only stated in terms of bounded arithmetical theories; cf. [@krajicek-proof] Section 14.5 for the proof of the separation of proof systems.
[^8]: Note that this does not exclude the possibility that $T$ proves a superlinear lower bound on $Q$ that is stronger than $P$ *assuming in $T$ that $Q$ is sound.*
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'In this paper we study some aspects of search for an immobile target by a swarm of $N$ non-communicating, randomly moving searchers (numbered by the index $k$, $k = 1, 2, \ldots, N$), which all start their random motion simultaneously at the same point in space. For each realization of the search process, we record the unordered set of time moments $\{\tau_k\}$, where $\tau_k$ is the time of the first passage of the $k$-th searcher to the location of the target. Clearly, $\tau_k$’s are independent, identically distributed random variables with the same distribution function $\Psi(\tau)$. We evaluate then the distribution $P(\omega)$ of the random variable $\omega \sim \tau_1/\overline{\tau}$, where $\overline{\tau} = N^{-1} \sum_{k=1}^N \tau_k$ is the ensemble-averaged realization-dependent first passage time. We show that $P(\omega)$ exhibits quite a non-trivial and sometimes a counterintuitive behaviour. We demonstrate that in some well-studied cases ([[*e.g.*]{}]{}, Brownian motion in finite $d$-dimensional domains) the *mean* first passage time is not a robust measure of the search efficiency, despite the fact that $\Psi(\tau)$ has moments of arbitrary order. This implies, in particular, that even in this simplest case (not saying about complex systems and/or anomalous diffusion) first passage data extracted from a single particle tracking should be regarded with an appropriate caution because of the significant sample-to-sample fluctuations.'
address:
- '$^1$ Laboratory of Physical Properties, Technical University of Madrid, Av. Complutense s/n, 28040 Madrid, Spain'
- '$^2$Laboratoire de Physique Théorique de la Matière Condensée (UMR CNRS 7600), Université Pierre et Marie Curie/CNRS, 4 place Jussieu, 75252 Paris Cedex 5 France'
- '$^3$ Laboratoire de Physique Théorique (UMR CNRS 8627), Université de Paris-Sud/CNRS, France'
author:
- 'Carlos Mejía-Monasterio$^1$, Gleb Oshanin$^2$ and Grégory Schehr$^3$'
title: First passages for a search by a swarm of independent random searchers
---
[*Keywords*]{}: Random motion, First passage times, Random search
Introduction {#introduction .unnumbered}
============
Search processes are ubiquitous in Nature: In order to survive, predators have to hunt the prey and the prey have to forage [@stephens; @bell; @klafter]. In order to convert into required reaction products, the reactants involved in chemical or biochemical reactions have first to find each other [@loverdo]. In many biophysical processes ligands search for binding sites [@net; @net1], proteins seek the target sequences on DNA’s [@berg], etc. Human beings look for a better job, partners, shelter, files in databases. Even an attempt to unlock a pin-protected device can be considered as a search in the space of all possible passwords [@pass].
Search for a desired target depends generally on a variety of different conditions and may take place in different environments: targets may be sparse, hidden, difficult to detect even when found. The targets may be immobile or mobile, try to avoid searchers or to evade from the searched area [@pnas; @kam]. They may have no or may have their own life-time and vanish before they are detected.
Searchers may be immobile, as it happens, [[*e.g.*]{}]{}, in visual search [@visual], may move freely or interact with the environment. Their motion may be hindered under conditions of molecular crowding [@crowding], [[*e.g.*]{}]{}, in cell’s cytoplasm or in dynamical backgrounds formed by other randomly moving particles, or facilitated due to interactions with the molecular motors. The searchers may search “blindly” detecting the target only upon an encounter with it, or “smell” (or “see”) the target somehow at long distances correcting their motion [@pnas; @smell]. They may have no memory of previously visited area or adapt their strategy “on-line” repelling themselves from their footprints on the searched substrate. Finally, the searchers may act individually or in swarms [@massimo; @gelenbe; @satya].
In general, for each specific situation different search strategies may be realized, and the question of efficient ones has motivated a great deal of work within the last years. While earlier works have considered deterministic search algorithms (see, e.g., Refs.[@stephens; @bell; @stone] and references therein) specific to such human activities as, say, search for natural resources or rescue operations, more recent studies focused on random search strategies. It was realized that the strategies based on Lévy flights or walks [@klafter; @viswanathan; @boyer], in which a searcher performs excursions whose lengths are random variables with heavy-tailed distributions, in some aspects are more advantageous than a search based on a conventional Brownian motion, or on random walks which step on nearest-neighbors only. Naturally, in this case the large-scale dynamics of searchers is superdiffusive.
Following the observation of trajectories of foraging animals in which active local search phases randomly alternate with relocation phases (see, e.g., Refs. [@kramer]) another type of random search - an intermittent search - has been proposed. In this algorithm the search process is characterized by two distinct types of motion - ballistic relocation stage when the searcher is non-receptive to the target and a relatively slow phase with a random Brownian-type motion when the target may be detected [@inter1; @inter2; @inter3; @inter4]. Much effort has been invested recently in understanding different optimization schemes for such a random search. In particular, one looked for the conditions allowing to minimize the mean first passage time for the process which is unlimited in time [@inter1], or seeked to enhance the chances of successful detection by minimizing the non-detection probability for the search process constrained to happen within a finite time interval [@inter2; @inter3; @inter4]. Note that for such a search the large-scale dynamics is diffusive, albeit intermittent.
Finally, a combination of a Lévy-based and intermittent search has been proposed in Ref. [@combined], in which the length of the relocation stage was taken as a random variable with a heavy-tailed distribution. It was shown that such a combined strategy is advantageous in the critical case of rare targets.
In this paper we discuss some aspects of a blind search by a swarm of $N$ independent, *non-communicating* searchers. We consider a situation, as depicted in Fig. \[target\], in which $N$ searchers (numbered by the index $k$, $k = 1, 2, \ldots, N$) occupy initially the same position in space, at some distance $x_0$ apart of an immobile target, start their random motion simultaneously and arrive for the first time to the location of the target at times $\tau_k$, respectively. Note that $\tau_k$s are not ordered. Clearly, in such a situation the first passage times $\tau_k$’s are independent, identically distributed random variables with the same distribution $\Psi(\tau)$.
We focus here on the random variable $$\omega = \frac{1}{N} \, \frac{\tau_1}{\overline{\tau}}, \label{def}$$ where $\overline{\tau}$ is the averaged, over the ensemble of $N$ searchers, realization-dependent first passage time, $$\overline{\tau} = \frac{1}{N} \, \sum_{k = 1}^N \tau_k.$$ Hence, the random variable $\omega$ probes the first passage time of a given searcher relative to the ensemble-averaged first passage time for $N$ independent searchers. The scaling factor $1/N$ in Eq. (\[def\]) is introduced here for convenience, so that regardless of the value of $N$, the random variable $\omega$ has a support on $[0,1]$. We note parenthetically that random variables such as in Eq. (\[def\]) were previously studied in Refs. [@iddo1; @sid; @iddo2; @iddo3; @greg1; @greg2] within a different context.
![A sketch of trajectories of four searchers starting at the same point (filled black circle) at a distance $x_0$ from the target and reaching the target for the first time at different time moments.[]{data-label="target"}](FIG1.eps){width="45.00000%"}
Our goal is to calculate the distribution function $$P(\omega) = \Big \langle \delta\left(\omega - \frac{1}{N} \, \frac{\tau_1}{\overline{\tau}}\right)\Big \rangle \;,$$ where $\delta(\cdot)$ is the delta-function and the angle brackets denote averaging over different realizations of independent, identically distributed random variables $\tau_k$.
For arbitrary normalized $\Psi(\tau)$, the distribution $P(\omega)$ is normalized, possesses all moments and the first moment $\langle \omega
\rangle \equiv \int^1_0 \omega \, d\omega \, P(\omega) = 1/N$. Note that $P(\omega)$ can be seen as a measure of the robustness of a given search algorithm. Clearly, if $P(\omega)$ appears to be sharply peaked at $\omega = 1/N$, this would signify that the underlying search algorithm is quite robust. Otherwise, if the distribution appears to be broad, or even to have a multi-modal shape, this would imply that the performance of such an algorithm is rather poor and sample-to-sample fluctuations matter. From yet another conceptual perspective, one can say that $P(\omega)$ probes the validity of the *mean* first passage time as the proper measure of a search process efficiency. We proceed to show that in many situations $P(\omega)$ has a rather complicated structure of which the mean behavior is not representative. We will show that in different situations $P(\omega)$ may have completely different shapes (modality) and also change the shape (say, from a unimodal bell-shaped form to a bimodal $U$- or an $M$-shaped one) when some of the parameters are slightly modified. A similar phenomenon of a shape reversal has been previously observed for different mathematical objects in Refs. [@sid; @greg1; @greg2].
We turn next to the key feature: the distribution $\Psi(\tau)$ of $\tau_k$’s, which encodes all the information on the specific properties of the searchers’ random motion and on their initial location relative to the target. Without a significant lack of generality, we suppose that for a search in infinitely large systems, in which all searchers perform a blind search and detect the target with probability $1$ upon a first encounter with it, $\Psi(\tau)$ can be defined as $$\label{distribution}
\Psi(\tau) = \frac{a^{\mu}}{\Gamma(\mu)} \, \exp\left(- \frac{a}{\tau}\right) \, \frac{1}{\tau^{1 + \mu}} ,$$ where $\Gamma(\cdot)$ is the gamma function, $a$ is a characteristic parameter which sets the cutoff of the distribution at small values of $\tau$, and $\mu \geq 0$ is the so-called the persistence exponent [@satya_review_persistence]. Note that depending on the dimensionality of space and type of random motion, one can encounter completely different values of $\mu$; $\mu$ can be $0 \leq \mu < 1$, $\mu = 1$ or $\mu > 1$. The distribution in Eq. (\[distribution\]) is normalized but does not possess already a first moment, (i.e., the mean first passage time), for $0 \leq \mu < 1$.
The distribution in Eq. (\[distribution\]) is exact for a Brownian motion (BM) in semi-infinite one-dimensional (1D) systems, in which case $\mu = 1/2$ and $a = x_0^2/4 D$, $D$ being the diffusion coefficient [@redner]. In fact, the case $\mu = 1/2$ appears to be very representative. According to the theorem due to Sparre Andersen [@sparre], in 1D systems for any discrete-time random walk with each step length chosen from a continuous, symmetric but otherwise arbitrary distribution the first passage time distribution $\Psi(n)$ decays with a number of steps $n$ as $n^{-3/2}$. For continuous-time $\tau$ Markov processes, an analogous result is $\Psi(\tau) \sim \tau^{-3/2}$. This universality, of course, is broken when subordination effects (i.e., long tailed waiting time distributions in a CTRW sense) are present.
Although not exact, the distribution in Eq. (\[distribution\]) is physically quite plausible for other types of random motion in infinite 1D systems provided that an appropriate choice of the exponent $\mu$ is made [@redner]: In particular, Eq. (\[distribution\]) with $\mu = 1 - 1/\alpha$ is a reasonable approximation for the first-arrival probability density for Lévy flights with Lévy index $\alpha$, $1 < \alpha < 2$ [@chechkin] and with $\mu = 1 - H$ [@molchan] - for the first-passage-time distribution for fractional BM with Hurst index $H$, $0 < H < 1$. One may also claim that $\mu = 1 - d_f/d_w$ for $d_f < d_w$ (compact exploration [@pgg]), where $d_f$ is a non-integer spatial dimension and $d_w$ is the fractal dimension of random motion trajectories, since the first passage distribution is defined as the time derivative of the survival probability of an immobile target or of the normalized current through the surface of the target [@redner]. Hence, its long-time tail is the same as the long-time tail of the time derivative of the Smoluchowski constant [@osh0; @osh1].
The case $\mu = 0$ can be encountered in border-line situations of compact exploration, when the dimension of space $d_f$ equals the fractal dimension $d_w$ of the random motion trajectories. In particular, this situation is realized for standard BM ($d_w = 2$) in two-dimensional (2D) space. Here, one finds that the *long-time* tail of the distribution $\Psi(\tau)$ follows (see, [[*e.g.*]{}]{}, Ref.[@redner]) $$\label{distribution_log}
\Psi(\tau) \sim \frac{1}{\tau \ln^2(\tau)}.$$ This situation will be considered in more detail in what follows.
Finally, one finds $\mu = d/2 - 1$, (or, more generally, $\mu = d_f/d_w
- 1$), for Brownian (anomalous) motion in $d > 2$ ($d_f > d_w$) dimensional systems. Note, however, that here the expression in Eq. (\[distribution\]) is not normalized and thus can not be considered as a distribution since only a finite fraction of trajectories will visit the target within an infinite time so that the searchers will have a finite probability to escape to infinity (non compact exploration [@pgg]). One can, however, normalize $\Psi(\tau)$ in Eq. (\[distribution\]) by hand, treating it as the conditional probability distribution of the first passage times for such trajectories which visit the target within an infinite time.
Next, we will consider a normalized, exponentially-truncated version of the distribution in Eq. (\[distribution\]): $$\label{truncated_dist}
\Psi(\tau) = \frac{\left(a b\right)^{\mu/2}}{2 K_{\mu}(2 \sqrt{a/b})} \, \exp\left(- \frac{a}{\tau}\right) \, \frac{1}{\tau^{1 + \mu}} \, \exp\left(- \frac{\tau}{b}\right),$$ where $K_{\mu}(\cdot)$ is the modified Bessel function. In contrast to $\Psi(\tau)$ in Eq. (\[distribution\]), the distribution in Eq. (\[truncated\_dist\]) possesses moments of arbitrary order. Note that the latter point is crucial and, according to a common belief, the first moment of the distribution, [[*i.e.*]{}]{}, the *mean* first passage time can be regarded as a robust measure of the search process efficiency. We set out to show that in many situations this is not the case due to the significant sample-to-sample fluctuations.
The distribution in Eq. (\[truncated\_dist\]) is exact for a BM in semi-infinite 1D systems in presence of a constant bias pointing towards the target (see, [[*e.g.*]{}]{}, Ref.[@redner]). In this case one has $a = x_0^2/4 D$, where $x_0$ is the starting point and $b = 4 D/v^2$, $v$ being the drift velocity.
One may argue, as well, that the expression in Eq. (\[truncated\_dist\]) is an appropriate approximation for the first-passage time distribution for search in finite systems or search assisted by smell. In general, of course, for random motion in finite systems the distribution $\Psi(\tau)$ will be represented as a series of exponentials; the form in Eq. (\[truncated\_dist\]) is thus tantamount to a heuristic approximation of this series in which one takes a behavior specific to an infinite system and truncates it by an exponential function with the characteristic decay time equal to the largest relaxation time. Depending on a particular situation, the cut-off parameter $b$ will be either proportional to the volume of the system (for BM in $d > 2$ systems), to $L^2$ where $L$ is the length of the interval for BM in 1D, to $S \ln(S) $ for BM in finite 2D systems of area $S$, or inversely proportional to the strength of the bias for biased diffusion. Some of these situations will be discussed in detail in the sequel.
We finally remark that the distribution in Eq. (\[truncated\_dist\]) appears in many other physical problems. To name but a few we mention the distribution of times between action potentials (or the ISI distribution) in the integrate-and-fire model of neuron dynamics [@man], the distribution of the stopping distances for the sliding motion of a solid block on an inclined heterogeneous plane [@lima], the avalanche life-time distribution in the mean-field version of the Bak-Sneppen model [@flyv], the distribution of the probability current in finite disordered one-dimensional samples [@osh] or the distribution of the number of times that a particle diffusing in a sphere hits its boundary during some time interval [@net1]. Thus our subsequent analysis applies to these systems as well.
The outline of our paper is as follows: We start in section 1 with a general remark on the tails of the first passage time distribution in case with $N$ independent randomly moving searchers. In section 2 we derive an exact result for the distribution $P(\omega)$ for $\Psi(\tau)$ in Eq. (\[distribution\]) with arbitrary $\mu$ and for arbitrary $N$. In section 3 we discuss several exactly solvable particular cases. Section 4 is devoted to the asymptotic analysis of $P(\omega)$ for large $N$ and arbitrary $\mu$. In section 5 we evaluate $P(\omega)$ for the exponentially-truncated distribution in Eq. (\[truncated\_dist\]) in case of two searchers. We also furnish here exact calculations of $P(\omega)$ for two BMs in finite 1D, 2D and 3D spherical domains. Next, in section 6, we discuss the form of $P(\omega)$ for $N$ searchers whose first passage time distribution is given by Eq. (\[truncated\_dist\]). Finally, in section 7 we conclude with a summary of our results and some generalizations.
First passage to the target for $N$ independent searchers
=========================================================
We begin with a somewhat evident but conceptually very important remark on the tails of the first passage time distribution in case of $N$ independent, non-communicating randomly moving searchers. For the original results and a discussion we address the reader to Ref. [@katja]. Some other interesting aspects of this model were discussed in Refs. [@gelenbe; @paul0; @paul].
Let $P_N(\tau)$ denote the probability that up to time moment $\tau$ neither of $N$ searchers has visited the target, [[*i.e.*]{}]{}, the target remained non-detected up to time $\tau$. For the situation under consideration, clearly, $$P_N(\tau) = P_1(\tau)^N,$$ where $P_1(\tau)$ is an analogous probability for a single searcher. Supposing that the target is found as soon as any of the searchers arrives to its location for the first time, we have that for $N$ searchers the first passage density $\Psi_N(\tau)$ is given by $$\label{qq}
\Psi_N(\tau) = - \frac{d P_N(\tau)}{d \tau}.$$ This yields immediately that $\Psi_N(\tau)$ follows (for the parental distribution in Eq. (\[distribution\])), as $\tau \to
\infty$, $$\label{dist1}
\Psi_N(\tau) \sim \frac{1}{\tau^{1 + \mu N}}.$$ A remarkable feature of this simple result is that depending on the number of searchers $N$, the distribution in Eq. (\[dist1\]) may have finite moments even if $\mu \leq 1$, contrary to the parent distribution in Eq. (\[distribution\]) which does not have any moment for such values of $\mu$. For instance, for $0 < \mu < 1$ and $\mu N > 1$, Eq. (\[dist1\]) has a finite first moment, [[*i.e.*]{}]{}, a finite mean first passage time, for $\mu N > 2$ it has a finite second moment, and generally, for $\mu N > k$ it has $k$ first finite moments. This is a crucially important advantage of search processes involving $N$ searchers, which makes the search process more efficient even for non-communicating searchers.
Next, it might be instructive to consider some exactly solvable case, [[*i.e.*]{}]{}, BM in semi-infinite 1D systems, and to calculate the $N$-dependence of the mean first passage time. In this case, the probability that a single searcher has not visited the location of the target (the origin) up to time moment $\tau$, starting at distance $x_0$ from the target, obeys $P_1(\tau) = {\rm erf}\left(x_0/\sqrt{4 D
\tau}\right)$, where ${\rm erf}(.)$ is the error function. Hence, the first moment of the distribution in Eq. (\[qq\]), [[*i.e.*]{}]{}, the mean first passage time $\big<\tau_N\big>$ for an ensemble of $N$ independent searchers is defined as $$\big< \tau_N \big> = - \int^{\infty}_0 d\tau \, \tau \, \frac{d P_N(\tau)}{d \tau} = \int^{\infty}_0 d\tau \, {\rm erf}^{N}\left(\frac{x_0}{\sqrt{4 D \tau}}\right).$$ The integral in the latter equation is convergent, as we have already remarked, for $N \geq 3$. One verifies that $\big< t_N \big>$ is a slowly *decreasing* function of the number of searchers $N$, and, for $N \gg 1$, we find that $$\label{t}
\big< \tau_N \big> \sim \frac{x_0^2}{4 D \ln\left(N\right)}.$$ Therefore, the more searchers one has, the less the mean first passage time is. Curiously enough, this essentially 1D result, $\big< \tau_N \big>$ in Eq. (\[t\]), coincides exactly with the mean residence time which $N$ BMs spend simultaneously together in a circular disc of radius $x_0$ on a 2D plane within an infinite time interval [@occupation], or with the mean first exit time of one of $N$ BMs from such a disc [@yuste].
Note finally that the mean first passage time may acquire a much stronger dependence on $N$ and decrease much faster for communicating searchers which share the information on the location of the target [@massimo]. Note, as well, that increasing the number of “searchers” allows to decrease substantially the time necessary to reach the target site on a DNA for proteins which strongly bind to other nonspecific sites acting as deep temporal traps [@raphael]. Some properties of the first passage time distribution for the event in which $N$ random walks appear for the first time simultaneously at the same lattice site have been discussed in Ref. [@pnas] within the context of a survival of an evasive prey.
General form of $P(\omega)$ for heavy-tailed first passage time distributions
=============================================================================
Let $\Big< \exp\left( - \lambda \omega\right)\Big>$, $\lambda \geq 0$, denote the moment generating function of the random variable $\omega$, Eq. (\[def\]). For arbitrary $\Psi(\tau)$, it can be formally represented as an $N$-fold integral: $$\left< e^{- \lambda \omega}\right> = \int^{\infty}_0 \ldots \int^{\infty}_0 \left( \prod_{n = 1}^N d \tau_n \, \Psi(\tau_n)\right) \,
\exp\left(-\lambda \frac{\tau_1}{\tau_1 + \tau_2 + \ldots + \tau_N} \right).$$ Integrating over $d \tau_1$, we change the integration variable $\tau _1 \to \omega$, to get $$\begin{aligned}
\left< e^{- \lambda \omega}\right> &=& \int^1_0 \frac{d \omega}{\left(1 - \omega\right)^2} \, e^{- \lambda \omega}
\int^{\infty}_0 \ldots \int^{\infty}_0 \left( \prod_{n = 2}^N d \tau_n \, \Psi(\tau_n)\right) \nonumber\\
&\times& \left(\tau_2 + \ldots + \tau_N \right) \Psi\left(\frac{\omega}{1 - \omega} \left(\tau_2 + \ldots + \tau_N \right) \right).\end{aligned}$$ Using next the following integral representation $$\tau \Psi(\tau) = \int^{\infty}_0 dp \, Q(p) \; e^{ - p \tau},$$ where the kernel $Q(p)$ is some unknown function defined via the inverse Laplace transform of the distribution $\Psi(\tau)$, we obtain the following general result for the probability density $P(\omega)$ in case of $N$ identic $\tau$-variables: $$\begin{aligned}
\label{general}
P(\omega) = \frac{1}{\omega^2} \int^{\infty}_0 d\lambda \, Q\left(\frac{1 - \omega}{\omega} \lambda\right) \; \Phi^{N - 1}\left(\lambda\right),\end{aligned}$$ with $\Phi(\lambda) = \left< \exp(- \lambda \tau)\right>$ being the characteristic function of the distribution $\Psi(\tau)$.
Now, for the distribution in Eq.(\[distribution\]) we have $$\label{1}
Q(p) = \frac{a^{(\mu + 1)/2}}{\Gamma(\mu)} p^{(\mu - 1)/2} J_{\mu - 1}\left(2 \sqrt{a p}\right),$$ and $$\label{2}
\Phi(\lambda) = \frac{2 a^{\mu/2}}{\Gamma(\mu)} \lambda^{\mu/2} K_{\mu}\left(2 \sqrt{a \lambda}\right),$$ where $J_{\nu}(\cdot)$ is the Bessel function.
Substituting the expressions in Eqs. (\[1\]) and (\[2\]) into Eq. (\[general\]), we find that the probability density $P(\omega)$ is given by $$\begin{aligned}
\label{m} P(\omega) = \frac{2^{-
\mu}}{\Gamma(\mu)} \frac{\left(1 - \omega\right)^{(\mu - 1)/2}}{\omega^{(3 + \mu)/2}} \,
\int^{\infty}_0 du \, u^{\mu } \, J_{\mu - 1}\left(\sqrt{\frac{1 - \omega}{\omega}} \, u\right) \, \Xi^{N-1}(u),\end{aligned}$$ independently of $a$, with $$\label{K} \Xi(u) = \frac{2^{1 - \mu} }{\Gamma(\mu)} \, u^{\mu} \,
{K}_{\mu}\left( u\right).$$ The result in Eq. (\[m\]) defines an exact distribution $P(\omega)$ for arbitrary $\mu$ and $N$. In several particular cases, the integral in Eq. (\[m\]) can be performed in closed form: when $N =
2$ or $N = 3$ and arbitrary $\mu > 0$, or when $\mu$ is equal to a half of an odd integer, while $N$ is arbitrary. We discuss below some of these cases, as well as present an asymptotic analysis of $P(\omega)$ in the limit $N \gg 1$.
Exactly solvable cases for the heavy-tailed distributions
=========================================================
Two non-communicating random searchers
--------------------------------------
Consider first the case of just two random searchers, $N = 2$. For this simple situation, one readily finds from Eq. (\[m\]) that $P(\omega)$ is given explicitly by: $$\label{lim}
P(\omega) = \frac{\Gamma(2 \mu)}{\Gamma^2(\mu)} \, \omega^{\mu - 1} \left(1 - \omega\right)^{\mu - 1},$$ [[*i.e.*]{}]{}, in this case $P(\omega)$ is a beta-distribution. This result has been also obtained within a different context in Ref. [@iddo1].
Notice now that, despite its simplicity, the result in Eq. (\[lim\]) contains a surprise: it has a completely different shape (modality) depending on whether $0 < \mu < 1$, $\mu = 1$ or $\mu > 1$ (see Fig. \[fig2\]).
![The distribution $P(\omega)$ in Eq. (\[lim\]) for $\mu = 1/2$, $\mu = 1$ and $\mu = 3/2$.[]{data-label="fig2"}](FIG2.eps){width="65.00000%"}
When $0 < \mu < 1$, the distribution $P(\omega)$ has a characteristic $U$-shape, so that the most probable values of $\omega$ are $0$ and $1$. Strikingly, the mean $\big< \omega\big> = 1/2$ corresponds in this case to the *least probable* value of the distribution. This signifies that in this case there is no symmetry between two identical searchers and both arrive to the target for the first time at distinctly different times. Note that for $\mu = 1/2$ (two BMs in 1D), $P(\omega) = 1/\pi \sqrt{\omega (1 - \omega)}$ and hence, the probability $P(\omega \leq \Omega)$ that $\omega$ attains any value from the interval $[0, \Omega]$ obeys $$P(\omega \leq \Omega) = \frac{2}{\pi} {\rm arcsin}\left(\sqrt{\Omega}\right),$$ which is the continuous arcsine distribution.
When $\mu = 1$, the distribution $P(\omega)$ in Eq. (\[lim\]) is uniform, which means that for either of the searchers its first passage time, relative to the average over the ensemble of two searchers, may take any value with equal probability.
Finally, for $\mu > 1$ the distribution $P(\omega)$ is unimodal and centered at $\omega = 1/2$. This signifies that in this case two searchers will most likely arrive to the target simultaneously.
Therefore, in situations when $0 < \mu \leq 1$ ([[*i.e.*]{}]{}, when $\Psi(\tau)$ does not have already the first moment) the first passage times of two searchers will be most probably distinctly different. On contrary, for $\mu > 1$ the first passage times will be most probably the same and hence, the algorithm resulting in such values of $\mu$ will be robust.
We end up this subsection by noticing that a similar transition was found in Ref. [@sanjib_sinai] in the related Sinai model with a linear drift of strength $\mu$. There the occupation time distribution on the positive axis exhibits a transition at $\mu=1$ and is also given, in certain limiting cases, by a beta-distribution, as in Eq. (\[lim\]).
Three non-communicating random searchers
----------------------------------------
For three non-communicating searchers, whose first passage times obey a non-truncated distribution in Eq. (\[distribution\]), we find $$\begin{aligned}
\label{N3}
P(\omega) &=& \frac{\sqrt{\pi}}{2^{4 \mu - 1}} \; \frac{\Gamma(2 \mu) \Gamma(3 \mu)}{\Gamma^3(\mu) \Gamma(2 \mu + 1/2)} \omega^{-1 - \mu} \left(1 - \omega\right)^{\mu - 1} \nonumber\\
&\times& \;_2F_1\left(2 \mu,
3 \mu; 2 \mu +\frac{1}{2}; - \frac{1 - \omega}{4 \omega}\right),\end{aligned}$$ where $_2F_1$ is a hypergeometric series. The distribution in Eq. (\[N3\]) for three different values of $\mu$ is depicted in Fig. \[fig3\].
![The distribution $P(\omega)$ in Eq. (\[N3\]) for $\mu = 1/2$, $\mu = 1$ and $\mu = 3/2$.[]{data-label="fig3"}](FIG3.eps){width="65.00000%"}
Consider first behavior of the distribution in Eq. (\[N3\]) in the vicinity of $\omega = 0$ and $\omega = 1$. Recollecting the definition of the hypergeometric series, one immediately observes from Eq. (\[N3\]) that $P(\omega) \sim C_2 (1-\omega)^{\mu - 1}$ when $\omega \to 1$. A little bit more involved analysis shows that $P(\omega) \sim C_1 \omega^{\mu - 1}$ when $\omega \to 0$. This means that similarly to the $N = 2$ case, the distribution $P(\omega)$ diverges at both edges when $\mu < 1$, and most probable values of $\omega$ are $0$ and $1$. Note, however, that $C_1 > C_2$ and hence, the distribution is skewed to the left favoring small values of $\omega$. Therefore, for $\mu < 1$ the most probable situation is that one of three searchers arrives to the location of the target much earlier than two others. Clearly, the mean $\langle \omega \rangle =
1/3$ does not have any significance \[apart, of course, of the fact that this is just the first moment of the distribution in Eq. (\[N3\])\].
Further on, the distribution in Eq. (\[N3\]) exhibits a qualitative change of behavior for $\mu \geq 1$. Here $P(\omega)$ is always a bell-shaped function of the variable $\omega$ centered at the most probable value $\omega = \omega_m$. The only difference between the $\mu = 1$ and $\mu > 1$ cases is that for the former $P(\omega)$ attains a non-zero values at the edges, $P(\omega = 1) = 1/3$ and $P(\omega=0) = 2$, while in the latter $P(\omega = 0) = P(\omega = 1) \equiv 0$. It is important to observe, however, that $\omega_m$ is always appreciably *less* than $\big< \omega \big> = 1/3$. For example, for $\mu
=3$, one has $\omega_m \approx 0.2719$, for $\mu = 10$ one has $\omega_m \approx 0.3102$ and etc. In fact, $\omega_m \to 1/3$ only when $\mu \to \infty$.
$N$ non-communicating random searchers for $\mu = 1/2$ and $\mu = 3/2$
----------------------------------------------------------------------
For $\mu = 1/2$ and arbitrary $N$, ([[*i.e.*]{}]{}, for $N$ BMs starting from the same point on a semi-infinite line), we find the following simple law: $$\label{mu12}
P(\omega) = \frac{N - 1}{\pi} \frac{1}{\sqrt{\omega \left(1 - \omega\right)}} \frac{1}{1 - \omega + (N - 1)^2 \omega},$$ while for $\mu = 3/2$ and arbitrary $N$ we get $$\begin{aligned}
\label{mu32}
\fl P(\omega) = \frac{\Gamma(N)}{\pi} \frac{\left(1 - \omega\right)^{1/2}}{\omega \left(1 - \omega + (N - 1)^2 \omega\right)^{3/2}} \sum_{p=0}^{N-1} \frac{p + 1}{\Gamma(N - p)} \nonumber\\
\times \left(\frac{\omega}{1 - \omega + (N-1)^2 \omega}\right)^{p/2} {\rm U}_{p + 1}\left( \left(\frac{\omega (N - 1)^2}{1 - \omega + (N-1)^2 \omega}\right)^{1/2}\right),\end{aligned}$$ where ${\rm U}_k(\cdot)$ are the Chebyshev’s polynomials of the second kind. These distributions, for a particular case $N = 10$, are depicted in Fig. (\[fig4\]).
![The distribution $P(\omega)$ in Eqs. (\[mu12\]) (red line) and (\[mu32\]) (blue line) for $N = 10$.[]{data-label="fig4"}](FIG4.eps){width="65.00000%"}
Note that for $\mu = 1/2$ and arbitrary $N$, the distribution $P(\omega)$ has a characteristic $U$-shape and diverges at the edges $\omega = 0$ and $\omega = 1$. The distribution function is strongly skewed towards the left edge; for $\omega \to 0$ the distribution behaves as $P(\omega) \sim (N - 1) \, \omega^{-1/2}$, while for $\omega \to 1$ we have $P(\omega) \sim (N - 1)^{-1} \, (1 -\omega)^{-
1/2}$, i.e., the amplitudes are $(N - 1)^2$ times different. This means that for $N \gg 1$ an event in which one out of a swarm of $N$ independent searchers arrives to the target much earlier than the rest of searchers is much more probable than an event in which one of the searchers arrives to the target location much later than others.
For $\mu = 3/2$ the distribution $P(\omega)$ is unimodal and peaked at some value $\omega = \omega_m$, which defines the most probable first passage time of a given searcher relative to the ensemble average value. Note, however, that $\omega_m$ is always less than the mean $\big< \omega \big> = 1/N$. Actually, as one may observe from Fig. \[fig4\], $\omega_m$ is more than two times less than $1/10$. Note that in this case the mean first passage time for an individual searcher exists, but apparently is not a proper measure of the search process since it is different from the most probable value.
Asymptotic large-$N$ behavior of $P(\omega)$ for heavy-tailed first passage time distributions
==============================================================================================
We consider next the large-$N$ asymptotic behavior of the probability density $P(\omega)$ in Eq. (\[m\]). To do this, it might be helpful first to look at the Eq. (\[m\]) from a different perspective and to realize that $P(\omega)$ can be expressed as the probability density function $P_{N-1}(|{\bf r}|)$ for the position of a random walker in some Rayleigh’s-type random flight model with a variable flight length in which $N$ - (the number of independent searchers) - will play the role of “time”.
Consider a Rayleigh’s random flight process (see, e.g., Ref. [@hughes]) in a $d = 2 \mu$-dimensional continuum. If $d$ is thought to be an integer, this would restrict the analysis to some particular values of $\mu$. However, as noted in Ref. [@kingman], formal considerations can be used to define an isotropic random flight in a space of arbitrary positive dimension $d$, with $d$ not necessarily restricted to integral values. This relaxes any constraint on the value of $\mu$.
Suppose that flights lengths are independent, identically distributed random variables and let the probability density function $p({\bf
r},{\bf r'})$ for a transition from ${\bf r}$ to ${\bf r'}$ be a function of $\rho = |{\bf r} - {\bf r'}|$ only, which means that the process is translationally invariant and $p({\bf r},{\bf r'}) =
p(\rho)$. Choose next $$\label{step} p(\rho) = \frac{\Gamma(2 \mu)}{\pi^{\mu}
\Gamma(\mu)} \frac{1}{\left(1 + \rho^2\right)^{2 \mu}},$$ [[*i.e.*]{}]{}, a heavy-tailed (but normalizable for any $\mu > 0$) distribution.
Using Fourier transform technique (see, [[*e.g.*]{}]{}, Ref. [@hughes]), one readily finds that the probability density function of a random walker, starting at the origin, to be at a distance $\rho$ from the origin after $N - 1$ such flights, is given by $$\label{rw}
\label{ray} P_{N-1}(\rho) = \frac{1}{2^{\mu} \pi^{\mu} \rho^{\mu - 1}} \; \int^{\infty}_0 du \, u^{\mu } \, J_{\mu -
1}\left(\rho \, u\right) \, \Xi^{N-1}(u),$$ with $\Xi(u)$ defined by Eq. (\[K\]).
Consequently, we find the following relation between the probability density $P(\omega)$ in Eq. (\[m\]) and $P_{N-1}(\rho)$ in Eq. (\[ray\]): $$\label{rel_rw}
\omega \left(1 - \omega\right) P(\omega) =
\left. \frac{\pi^{\mu}}{\Gamma(\mu)} \rho^{2 \mu} P_{N-1}(\rho)\right|_{\rho = \sqrt{(1 - \omega)/\omega}} \, .$$ Therefore, the probability density of the random variable $\omega$, which describes the realization-dependent ratio of the first passage time of a given searcher in an ensemble of $N$ ones, and of the ensemble averaged first passage time, is proportional to the probability density of finding a random walker performing Rayleigh’s random flights with a broad distribution of flight length, Eq. (\[step\]), at distance $\sqrt{(1 - \omega)/\omega}$ away from the origin after $N - 1$ flights.
When $N$ is large, the integral in Eq. (\[m\]) is dominated by the behavior of $\Xi(u)$ in the vicinity of $u = 0$. In turn, the latter depends on the value $\mu$: the cases $\mu > 1$, $\mu =1$ and $\mu <
1$ need to be considered separately.
The case $\mu > 1$
------------------
In this case, the leading small-$u$ behavior of the characteristic function follows: $$\label{u}
\ln\left(\Xi(u)\right) \sim -
u^2/4 (\mu - 1) \;.$$ Plugging the latter asymptotic form into Eq. (\[m\]), and performing integration over $du$, we find that the asymptotic large-$N$ behavior of the distribution $P(\omega)$ is determined by $$\label{form}
P(\omega) \sim \frac{1}{\omega (1 - \omega)} \left(\frac{1- \omega}{\omega N}\right)^{\mu} \exp\left(- (\mu - 1) \frac{(1 - \omega)}{\omega N}\right).$$ Therefore, in the limit $N \gg 1$ for $\mu > 1$ the distribution $P(\omega)$ is always a bell-shaped function of $\omega$, which approaches $0$ exponentially fast when $\omega \to 0$ and as a power-law when $\omega \to 1$. The maximum of $P(\omega)$ is located at $$\begin{aligned}
\omega_m &=& \frac{1}{4 N} \left(\mu - 1 + (\mu + 1) N - \sqrt{\left(\mu - 1 + (\mu + 1) N\right)^2 - 8 (\mu - 1) N } \right) \nonumber\\ &\sim& \frac{\mu - 1}{\mu + 1} \frac{1}{N}.\end{aligned}$$ This substantiates our previous claims that $\omega_m$ is always less than $\big< \omega \big> = 1/N$ and converges to $1/N$ only when $\mu
\to \infty$. Note that in this case the first moment of $\Psi(\tau)$, [[*i.e.*]{}]{}, the mean first passage time, exists but it is not representative of the most probable behaviour.
It is instructive now to reproduce the result in Eq. (\[form\]) using a different type of argument. Notice that $P_{N}(\rho)$ in Eq. (\[rw\]) in the limit $N \to \infty$ becomes a Gaussian distribution in $d = 2 \mu$ dimensional space of one rescaled variable $\rho/N^{1/2}$: $$\begin{aligned}
P_{N}(\rho) \sim \frac{1}{N^{\mu}} \exp{\left(-(\mu-1) \frac{\rho^2}{N}\right)}.\end{aligned}$$ Inserting the latter expression into the relation in Eq. (\[rel\_rw\]), we recover the result in Eq. (\[form\]).
The case $\mu = 1$
------------------
In this borderline case one finds $$\label{Ximu1}
\ln\left(\Xi(u)\right) \sim - \left(1 - 2 \gamma + 2 \ln 2 - 2 \ln(u) \right) u^2/4 \;,$$ where $\gamma$ is the Euler constant.
Note that here one has an additional logarithmic factor $\ln(u)$, as compared to the leading small-$u$ behavior in Eq. (\[u\]). Since logarithm is a slowly varying function, we can repeat essentially the same argument: $P_{N}(\rho)$ in Eq. (\[rw\]) converges, as $N \to
\infty$, to a Gaussian distribution of the scaling variable $\rho/({N
\log N})^{1/2}$: $$P_{N}(\rho) \sim \frac{1}{N \ln(N)} \exp\left(- \frac{\rho^2}{N \ln(N)}\right).$$ Hence, in virtue of the relation in Eq. (\[rel\_rw\]), we find $$\begin{aligned}
P(\omega) \sim \frac{1}{\left(N \ln(N) \omega\right)^2} \exp\left(- \frac{(1 - \omega)}{N \ln(N) \omega}\right)\;.\end{aligned}$$ This is again a bell-shaped function of $\omega$ with the most probable value $\omega_m~\sim~1/(2 N \log N)$.
The case $\mu < 1$
------------------
For $\mu <1$ the leading small-$u$ behavior of the characteristic function reads: $$\ln\left(\Xi(u)\right) \sim - \frac{\Gamma(1 - \mu)}{4^{\mu} \Gamma(1 + \mu)} u^{2 \mu } \;,$$ so that for $N \to \infty$, $P_{N}(\rho)$ in Eq. (\[rw\]) becomes $$P_{N}(\rho) = \frac{1}{N} S\Big(x = \frac{\rho}{N^{1/2\mu}}, \alpha = 2 \mu, \sigma = \frac{\Gamma(1 - \mu)}{\Gamma(1+\mu)}\Big),$$ where $S(\cdot)$ is the one-sided, $2 \mu$-dimensional stable law with index $\alpha = 2 \mu$ and scale $\sigma = \Gamma(1 -
\mu)/\Gamma(1+\mu)$ [@kol]. In consequence, the distribution $P(\omega)$ is given, for $N \to \infty$, by $$\label{onesided}
P(\omega) \sim \frac{(1 - \omega)^{\mu - 1}}{N \omega^{\mu + 1}} \, S\Big(x = \frac{1}{N^{1/2\mu}} \sqrt{\frac{1 - \omega}{\omega}}, \alpha = 2 \mu, \sigma = \frac{\Gamma(1 - \mu)}{\Gamma(1+\mu)}\Big).$$ Asymptotic behavior of $S(\cdot)$ has been discussed in detail in Ref. [@kol]. When $x \ll 1$, $S(\cdot) \to const$, and hence, $$P(\omega) \sim \frac{(1 - \omega)^{\mu - 1}}{N},$$ [[*i.e.*]{}]{}, $P(\omega)$ diverges as $\omega \to 1$. On the other hand, when $x \gg 1$, [[*i.e.*]{}]{}, when $\omega$ is sufficiently close to $0$, $S(\cdot) \sim
1/x^{4 \mu}$ [@kol], which yields $$P(\omega) \sim N \omega^{\mu - 1}.$$ This means that $P(\omega)$ diverges when $\omega \to 0$ and, generally, in this domain $0 < \mu < 1$ the distribution has a characteristic $U$-shaped form strongly skewed towards small values of $\omega$ since the amplitudes differ by a factor $N^2$. Note that for $\mu = 1/2$ the distribution in Eq. (\[onesided\]) becomes the Cauchy distribution so that the simple form in Eq. (\[mu12\]) follows immediately.
The distribution $P(\omega)$ for exponentially-truncated first passage time distributions. Two non-communicating searchers
==========================================================================================================================
We turn next to the analysis of the distribution $P(\omega)$ in case when the parent first passage time distribution $\Psi(\tau)$, Eq. (\[truncated\_dist\]), possesses moments of arbitrary order.
For two non-communicating searchers and the exponentially-truncated distribution in Eq. (\[truncated\_dist\]), we get the following result for the distribution of the random variable $\omega$: $$\label{2trunc}
P(\omega) = \frac{1}{2 K^2_{\mu}(2 \sqrt{a/b})} \, \frac{1}{\omega (1 - \omega)} \, K_{2 \mu}\left(2 \sqrt{\frac{a}{b \, \omega ( 1- \omega)}}\right).$$ Two remarks are in order. First, one readily notices that $P(\omega)$ vanishes exponentially fast when $\omega \to 0$ or $\omega \to 1$ so that $P(\omega = 0) = P(\omega = 1) = 0$. Second, $P(\omega)$ is clearly symmetric under the replacement $\omega \to 1 - \omega$. Since here we deal with a truncated distribution $\Psi(\tau)$ which possesses the moments of arbitrary order, our first guess would be that $P(\omega)$ is always a bell-shaped function with a maximum at $\omega = 1/2$. To check this guess, we expand $P(\omega)$ in the Taylor series around $\omega = 1/2$: $$\begin{aligned}
\fl P(\omega) = \frac{2 K_{2 \mu}(4 \sqrt{a/b})}{K^2_{\mu}(2 \sqrt{a/b})} \, \Big[ 1 + \nonumber\\
+
4 \left(1 - \mu - 2 \sqrt{\frac{a}{b}} \frac{K_{2 \mu - 1}(4 \sqrt{a/b})}{K_{2 \mu}(4 \sqrt{a/b})}\right) \left(\omega - \frac{1}{2}\right)^2 + \mathcal{O}\left(\omega - \frac{1}{2}\right)^4\Big].\end{aligned}$$ Inspecting the sign of the coefficient before the quadratic term, [[*i.e.*]{}]{}, $$\label{g}
g = 1 - \mu - 2 \sqrt{\frac{a}{b}} \frac{K_{2 \mu - 1}(4 \sqrt{a/b})}{K_{2 \mu}(4 \sqrt{a/b})},$$ we notice that
- For $\mu > 1$, $g$ is always negative for any value of $b/a$ so that here the distribution $P(\omega)$ is a bell-shaped function with a maximum at $\omega = 1/2$.
- For $\mu = 1$, $g$ is negative and approaches $0$ from below when $b/a \to \infty$. It means that $P(\omega)$ is generally a bell-shaped function with a maximum at $\omega = 1/2$, but it is becoming progressively flatter when $b/a$ is increased, so that ultimately $P(\omega) \approx 1$ apart of very narrow regions at the edges for $b/a \gg 1$.
- For $0 \leq \mu < 1$ there always exists a critical value $y_c(\mu)$ of the parameter $y = b/a$ which is defined implicitly as the solution of the equation $g = 0$, Eq. (\[g\]). For $b/a <
y_c(\mu)$, the distribution $P(\omega)$ is unimodal with a maximum at $\omega = 1/2$. For $b/a = y_c(\mu)$, the distribution is nearly uniform except for narrow regions in the vicinity of the edges. Finally, which is quite surprising in view of the fact that in this case $\Psi(\tau)$ possesses all moments, for $b/a >
y_c(\mu)$ the distribution $P(\omega)$ is bimodal with a characteristic $M$-shaped form, two maxima close to $0$ and $1$ and $\omega = 1/2$ being the least probable value.
We depict in Fig. \[fig5\] three characteristic forms of $P(\omega)$ for $\mu =
1/2$ and three different values of $b/a$.
![The distribution $P(\omega)$ in Eq. (\[2trunc\]) for $\mu = 1/2$ and different values of $b$ ($a$ is set equal to $1$).[]{data-label="fig5"}](FIG5.eps){width="65.00000%"}
Therefore, for $\mu \geq 1$ two non-communicating searchers will most probably reach the target simultaneously. For $0 \leq \mu < 1$ two distinctly different situations are possible: if $b/a$ is less than some well-defined critical value $y_c(\mu)$, then most likely both searchers will arrive to the target for the first time together. If, on the contrary, $b/a$ exceeds this critical value, the event in which both searchers arrive to the location of the target simultaneously will be the least probable one.
Two independent BMs in a semi-infinite 1D system with a bias
------------------------------------------------------------
Recall now that the distribution in Eq. (\[truncated\_dist\]) with $\mu = 1/2$ is exact for a BM in semi-infinite 1D systems in presence of a constant bias pointing towards the target. In this case one has $a = x_0^2/4 D$, where $x_0$ is the starting point and $D$ - the diffusion coefficient, and $b = 4 D/v^2$, $v$ being the drift velocity. Hence, $2 \sqrt{a/b} = {\it Pe} = x_0 |v|/2 D$ is the Peclet number (see, [[*e.g.*]{}]{}, Ref. [@redner]). Consequently, we can make a following statement:
Consider two independent, absolutely identical BMs on a semi-infinite line, starting at the same point $x_0$, having the same diffusion coefficient $D$ and experiencing the same bias $F$ which points towards the origin so that the drift velocity of both BMs is $v < 0$. Then, an event in which both BMs arrive for the first time to the origin simultaneously is
- the *least* probable if $Pe < Pe_c$,
- the *most* probable if $Pe > Pe_c$,
where $Pe_c$ is the solution of the transcendental equation $$1 = 2 Pe_c \frac{K_0(2 Pe_c)}{K_1(2 Pe_c)}.$$ An approximate solution of the latter equation gives $Pe_c \approx
0.666...$.
Therefore, the mean first passage time to the target might be an appropriate measure of the search efficiency for sufficiently large Peclet numbers, but definitely is not the one in case of small $Pe$. In the latter case the sample-to-sample fluctuations are significant and the mean value is not representative of the actual behaviour.
Two independent unbiased BMs on a finite interval
-------------------------------------------------
We have already remarked that the expression in Eq. (\[truncated\_dist\]) is a reasonable approximation for the first passage time distribution for random motion in finite systems. For a one-dimensional bounded interval of length $L$ and for $\mu = 1/2$, the parameter $b \sim L^2/D$, while $a \sim x_0^2/D$, where $x_0$ is the starting point. Consequently, $\sqrt{a/b}$ should be $ \sim
x_0/L$ and independent of the particles’ diffusion coefficient $D$. This suggests a somewhat strange result that the very shape (or the modality) of the distribution $P(\omega)$ will crucially depend on the starting point $x_0$.
![Two BMs on a finite interval $[0,L]$ with a reflecting boundary at $x = L$. The distribution $P(\omega)$ in Eq. (\[pw1d\]) for different values of $x_0/L$, $x_0$ being the starting point of both BMs.[]{data-label="1dd"}](FIG6.eps){width="65.00000%"}
To verify such a prediction, we focus now on a BM on a finite 1D interval and re-examine this question using an exact form of the normalized first passage time distribution $\Psi(\tau)$. For the BM with a starting point at $x_0$ on an interval of length $L$ with an adsorbing boundary at $x=0$ and a reflecting boundary at $x = L$, $\Psi(\tau)$ is given by $$\label{fpt1}
\Psi(\tau) = \frac{2 \pi D}{L^2} \sum_{n = 0}^{\infty} A^{(d=1)}_n\left(\frac{x_0}{L}\right) \exp\left(- \frac{\pi^2 (n + 1/2)^2 D \tau}{L^2}\right),$$ where $$A^{(d=1)}_n\left(\frac{x_0}{L}\right) = \left(n + \frac{1}{2}\right)\sin\left(\frac{\pi (n + 1/2) x_0}{L}\right) \ .$$ Consequently, the normalized distribution $P(\omega)$ in case of two independent, identical BMs has the following form $$\begin{aligned}
\label{pw1ddd}
P(\omega) &=& \frac{4}{\pi^2} \sum_{m,n=0}^{\infty} \frac{A^{(d=1)}_n\left(x_0/L\right) \, A^{(d=1)}_m\left(x_0/L\right)}{\left(\omega (n + 1/2)^2 + (1 - \omega) (m+ 1/2)^2\right)^2}.\end{aligned}$$ Using next an equality $$\label{equality}
\frac{1}{\lambda_m^2} \frac{d}{d\omega} \frac{1}{\lambda_n^2 + \frac{1 - \omega}{\omega} \lambda_m^2} = \frac{1}{\left(\omega \lambda_n^2 + (1 - \omega) \lambda_m^2\right)^2},$$ and the following representation of the characteristic function $\Phi(\lambda)$ of the distribution $\Psi(\tau)$ in Eq. (\[fpt1\]), $$\begin{aligned}
\Phi(\lambda) &=& \int_0^{\infty} d\tau \, \Psi(\tau) \, \exp\left( - \lambda \, \tau\right) \nonumber\\
&=& \frac{2}{\pi} \sum_{n = 0}^{\infty} \frac{ A^{(d=1)}_n\left(x_0/L\right)}{(n + 1/2)^2 + L^2 \lambda/\pi^2 D} \nonumber\\
&=& \frac{\cosh\left((L - x_0) \sqrt{\lambda/D}\right)}{\cosh\left(L \sqrt{\lambda/D}\right)},\end{aligned}$$ we can conveniently rewrite Eq. (\[pw1ddd\]) as $$\begin{aligned}
\label{pw1d}
P(\omega) = \frac{2}{\pi} \frac{d}{d\omega} \sum_{m = 0}^{\infty} \frac{\sin\left(\pi (m+1/2) \frac{x_0}{L}\right)}{m + 1/2} \, \Phi\left(\lambda = \frac{\pi^2 D}{L^2} \frac{1 - \omega}{\omega} \left(m + 1/2\right)^2\right) \nonumber\\
= \frac{2}{\pi} \frac{d}{d\omega} \sum_{m = 0}^{\infty} \frac{\sin\left(\pi (m+1/2) \frac{x_0}{L}\right)}{m + 1/2} \, \frac{\cosh\left(\pi (m+1/2) \sqrt{\frac{1 - \omega}{\omega}} \left(1 - \frac{x_0}{L}\right)\right)}{\cosh\left(\pi (m+1/2) \sqrt{\frac{1 - \omega}{\omega}}\right)} \,.\end{aligned}$$
The distribution $P(\omega)$ tends to zero exponentially, $P(\omega) \sim \exp(- \pi x_0/2 L \sqrt{\omega})/\omega^{3/2}$, when $\omega \to 0$, precisely in the same way as $P(\omega)$ in Eq. (\[2trunc\]) obtained for the exponentially truncated first passage time distribution. By symmetry, we expect the same behavior when $\omega \to 1$. In principle, $P(\omega)$ in Eq. (\[pw1d\]) can be represented in closed form as a complicated combination of elliptic function. However, we prefer to proceed with a numerical analysis of the rapidly convergent series in Eq. (\[pw1d\]), in order to understand whether $P(\omega)$ is always a bell-shaped function of $\omega$, or undergoes a transition to an $M$-shaped form at a certain value of $x_0/L$.
In Fig. \[1dd\] we depict the distribution $P(\omega)$ in Eq. (\[pw1d\]) for different values of the ratio $x_0/L$. One notices that $P(\omega)$ has a different shape depending whether $x_0/L$ is less or greater than the critical value $\approx 0.605...$. This allows us to make the following statement:
Consider two independent, identical, unbiased BMs starting at the same point $x_0$ on a finite interval $[0,L]$ with a reflecting boundary at $x = L$. Then, an event in which both BMs arrive simultaneously to the origin is
- the *least* probable if $x_0/L \lesssim 0.605...$,
- the *most* probable if $x_0/L \gtrsim 0.605...$.
From a common sense point of view such a behavior seems a bit counterintuitive - indeed, why should two BMs arrive to the target at progressively distinct times the closer they are to its location, and should most probably arrive together when they are far from it? On the other hand, such a behavior is quite a natural one: Indeed, $x_0$ ($a^{1/2}$) and $L$ ($b^{1/2}$) define the effective size of the window in which the decay of the first passage time distribution is governed by the intermediate power-law tail. The larger is this window, the closer we are to the situation described in Section III. Therefore, the origin of such a disproportionate behavior of two identical, independent BMs is precisely the same as the one behind the famous arcsine law for the distribution of the fraction of time spent by a random walker on a positive half-axis [@arcsine]: Once one of the BMs goes away from the target, it finds it more difficult to return than to keep on going away.
![A BM on a finite interval $[0,L]$ (with a reflecting boundary at $x = L$) starting at $x = x_0$. The first passage time distribution $\Psi(\tau)$ in Eq. (\[fpt1\]) for $L=
100$, $D = 1/2$ and different values of $x_0/L$.[]{data-label="fpt1dd"}](FIG7.eps){width="70.00000%"}
To substantiate this claim, we plot in Fig. \[fpt1dd\] the distribution in Eq. (\[fpt1\]) for three different values of $x_0$ and fixed $L$. Note that all three curves show an exponential behavior for both small and large values of $\tau$ (which, in fact, is an argument in favor of our choice of the exponentially truncated distribution in Eq. (\[truncated\_dist\])). For large $\tau$ all three curves merge which signifies that at such values of $\tau$ the characteristic decay time is dependent only on $L$. The lower cut-off is clearly dependent only on the starting point $x_0$. Further on, notice that the closer (for a fixed $L$) the starting point $x_0$ to the reflecting boundary is, the narrower is the distribution $\Psi(\tau)$. On contrary, the smaller $x_0$ is (the closer to the target), the more pronounced the intermediate power-law behavior $\sim
t^{-3/2}$ becomes (see the dashed line in Fig. \[fpt1dd\]). Actually, the fact that the farther away the starting point from the reflecting wall is, the broader is the first passage time distribution (the BM simply does not “know” that it is in a finite system up to times of order $\sim L^2/D$) has been already discussed in detail in Ref. [@redner].
![The coefficient of variation $\gamma_v$, the skewness $\gamma_a$ and the kurtosis $\gamma_e$ of the distribution $\Psi(\tau)$ in Eq. (\[fpt1\]) versus $x_0/L$. Dotted vertical line defines the critical value $x_0/L \approx 0.605...$ at which the distribution $P(\omega)$ changes the modality.[]{data-label="kurtosis"}](FIG8.eps){width="65.00000%"}
Consider next some standard characteristics of the first-passage-time distribution in Eq. (\[fpt1\]), such as the mean first passage time, the skewness, the coefficient of variation and the kurtosis, to see if any of them reflects the transition observed for $P(\omega)$ at $x_0/L \approx 0.605...$. Note that the moments of arbitrary order of the distribution $\Psi(\tau)$ in Eq. (\[fpt1\]) can be straightforwardly calculated in closed form: $$\big<\tau^m\big> \equiv \int^{\infty}_0 \tau^m \, d\tau \, \Psi(\tau) = \frac{(-1)^m \sqrt{\pi}}{\Gamma(m + 1/2)} \left(\frac{ L^2}{D}\right)^m E_{2 m}\left(\frac{x_0}{2 L}\right) \,,$$ where $E_{2 k}(.)$ are the Euler polynomials. Consequently, the mean and the variance are given by $$\big<\tau\big> = - 2 \frac{ L^2}{D} \, E_2\left(\frac{x_0}{2 L}\right),$$ and $$k_2 = \big<\tau^2\big> - \big<\tau\big>^2 = \frac{4}{3} \frac{L^4}{D^2} \left(E_4\left(\frac{x_0}{2 L}\right) - 3 E_2^2\left(\frac{x_0}{2 L}\right)\right) \,$$ respectively. One can readily check that both characteristic properties are monotonically increasing functions of $x_0/L$ and do not show any sign of a particular behavior at $x_0/L \approx 0.605...$. Further on, we define the coefficient of variation $$\gamma_v = \sqrt{\frac{k_2^2}{\big<\tau\big>^2}} = \sqrt{\frac{E_4\left(x_0/2 L\right) - 3 E_2^2\left(x_0/2 L\right)}{3 E_2^2\left(x_0/2 L\right)}},$$ the skewness $$\gamma_a = \frac{k_3}{k_2^{3/2}},$$ where $k_3$ is the third cumulant: $$\begin{aligned}
k_3 = \big<\tau^3\big> - 3 \big<\tau\big> \big<\tau^2\big> + 2 \big<\tau\big>^3,\end{aligned}$$ and the kurtosis (coefficient of excess) $$\gamma_e = \frac{k_4}{k_2^2},$$ with $k_4$ being the fourth cumulant of the distribution in Eq. (\[fpt1\]): $$k_4 = \big<\tau^4\big> - 3 \big<\tau^2\big>^2 - 4 \big<\tau\big> \big<\tau^3\big> + 12 \big<\tau\big>^2 \big<\tau^2\big> - 6\big<\tau\big>^4 \,.$$ In Fig. \[kurtosis\] we plot the coefficient of variation, the skewness and the kurtosis of the distribution in Eq. (\[fpt1\]). One observes a strong variation of these properties reflecting an influence of extreme events for sufficiently small values of $x_0/L$. For $x_0/L > 0.6$ the variation of $\gamma_a$ and $\gamma_e$ becomes rather small. However, neither of these properties shows a clear demarkation line between different regimes exhibited by $P(\omega)$.
We finally remark that in many practically interesting physical problems the starting point $x_0$ is not fixed, but the searcher rather starts from some random location which is uniformly distributed on the interval. The distribution $P_{av}(\omega)$ appropriate to such a situation is obtained by merely averaging $P(\omega)$ in Eq. (\[pw1d\]) over $x_0$, [[*i.e.*]{}]{}, $$\label{av}
P_{av}(\omega) = \frac{1}{L} \int^L_{0} dx_0 \, P(\omega).$$ A quick inspection of $P(\omega)$ in Eq. (\[pw1d\]) shows that, due to the orthogonality of $A^{(d=1)}_n(x_0/L)$, the averaged distribution $P_{av}(\omega)$ does not depend on $\omega$, so that $$P_{av}(\omega) \equiv 1.$$ This seems to be a general property of $P_{av}(\omega)$, associated with the probability conservation and thus, as will be checked for further examples, is independent of the dimension of space.
Two independent BMs in a disc with a reflecting boundary
--------------------------------------------------------
Recall that in an infinite 2D system the first passage time distribution has an algebraic (with a logarithmic correction) tail with $\mu = 0$, Eq. (\[distribution\_log\]). Consequently, we may expect essentially the same behavior as we observed in two previous subsections.
![Two BMs in a (2D or 3D) bounded spherical domain of radius $L$ with a reflecting (thick black line) boundary. The target is a circle (in 3D, a sphere) of radius $r$ centered at the origin. Two BMs start from the same point (a filled circle) at a distance $x_0$ from the center and arrive to the target boundary for the first time at time moments $\tau_1$ and $\tau_2$, respectively.[]{data-label="target2"}](FIG9.eps){width="45.00000%"}
Consider an immobile target of radius $r$ fixed at the origin of a disc of radius $L$ with a reflecting boundary. Suppose next that a BM starts at some point at distance $x_0$ from the origin and hits the target for the first time at time moment $\tau$. Then, the first passage time distribution $\Psi(\tau)$ in such a situation is given explicitly by
![A BM in a two-dimensional disc of radius $L$ with a reflecting boundary. First passage time distribution $\Psi(\tau)$ in Eq. (\[fpt2\]) for $D = 1/2$, $L= 100$, $r = 5$ and different values of $x_0/L$.[]{data-label="fpt2d"}](FIG10.eps){width="70.00000%"}
$$\label{fpt2}
\Psi(\tau) = \frac{r D}{Z} \sum_{n = 0}^{\infty} A_n^{(d = 2)}(r,x_0,L) \exp\left(- \lambda_n^2 D \tau\right),$$
where $Z$ is the normalization, $$A_n^{(d = 2)}(r,x_0,L) = \frac{U_0(\lambda_n x_0) U_0'(\lambda_n r)}{\lambda_n^2 L^2 U_0^2(\lambda_n L) - r^2 \left(U_0'(\lambda_n r)\right)^2},$$ $$U_0(x) = Y_0(\lambda_n r) J_0(x) - J_0(\lambda_n r) Y_0(x),$$ $$U_0'(\lambda_n r) = \frac{d U_0(\lambda_n x)}{dx}\bigg|_{x=r},$$ while $\lambda_n$ are the roots of the function $$Y_0(\lambda_n r) J_1(\lambda_n L) - J_0(\lambda_n r) Y_1(\lambda_n L),$$ arranged in an ascending order, and $Y_n(\cdot)$ are Bessel functions of the second kind. Note that $\lambda_n$ depends on $L$ and $r$.
In Fig. \[fpt2d\] we depict the distribution in Eq. (\[fpt2\]) for fixed $L$, $D$ and $r$, and several values of $x_0$. Note that, apart of the intermediate behavior, which follows a power-law with a logarithmic correction $\sim 1/\tau \log^2(\tau)$ (see the dashed line in Fig. \[fpt2d\]), we have here essentially the same trend as the one we observed for a BM on a 1D finite interval. Namely, the closer is the starting point to the location of the target, the broader is the first passage time distribution so that the intermediate power-law behavior gets more pronounced.
Suppose next that we have two BMs starting at the same point some distance $x_0$ apart of the origin. Then, from Eq. (\[fpt2\]), we get the following result for the distribution $P(\omega)$: $$\label{pw222}
P(\omega) = \frac{r^2}{Z^2} \sum_{n,m=0}^{\infty} \frac{A_n^{(d=2)}(r,x_0,L) \, A_m^{(d = 2)}(r,x_0,L)}{\left(\omega \lambda_n^2 + (1 - \omega) \lambda_m^2\right)^2} \,,$$ which can be conveniently rewritten, using the equality in Eq. (\[equality\]), as $$\label{pw2}
P(\omega) = \frac{r}{Z} \frac{d}{d\omega} \sum_{m=0}^{\infty} \frac{A_m^{(d = 2)}(r,x_0,L)}{\lambda_m^2} \, \Phi\left(\lambda = \frac{1 - \omega}{\omega} D \lambda_m^2\right) \,,$$ where $\Phi(\lambda)$ is the characteristic function of the first passage time distribution in Eq. (\[fpt2\]) defined by [@redner]: $$\Phi(\lambda) = \frac{I_0\left(\sqrt{\frac{\lambda}{D}} x_0\right) K_1\left(\sqrt{\frac{\lambda}{D}} L\right) + K_0\left(\sqrt{\frac{\lambda}{D}} x_0\right) I_1\left(\sqrt{\frac{\lambda}{D}} L\right)}{I_0\left(\sqrt{\frac{\lambda}{D}} r\right) K_1\left(\sqrt{\frac{\lambda}{D}} L\right) + K_0\left(\sqrt{\frac{\lambda}{D}} r\right) I_1\left(\sqrt{\frac{\lambda}{D}} L\right)} \,.$$
We depict the result in Eq. (\[pw2\]) for fixed $L$ and $r$, and several values of $x_0/L$.
![Two Brownian motions in a two-dimensional disc. The distribution $P(\omega)$ in Eq. (\[pw2\]) for $L= 100$, $r = 5$ and different values of $x_0/L$.[]{data-label="pw2d"}](FIG11.eps){width="65.00000%"}
One notices that the shape of the distribution $P(\omega)$ is again different, depending on the value of the ratio $x_0/L$. For $x_0/L
\lesssim 0.643$, the distribution has an M-shaped form with a minimum at $\omega = 1/2$, for $x_0/L \approx 0.643$ the distribution is nearly uniform except for narrow regions near the edges, and lastly, for $x_0/L \gtrsim 0.643$ it becomes unimodal, but still it is very broad and has a pronounced flat region around the maximum. This signifies that even in this regime the sample-to-sample fluctuations are significant.
Consider finally the situation when the starting point of two BM is uniformly distributed within the space between two concentric circles of radius $r$ and radius $L$. One readily finds, by averaging the result in Eq. (\[pw2\]), that here as in 1D one has $P_{av}(\omega) \equiv 1$.
Two independent BMs in a sphere with a reflecting boundary
----------------------------------------------------------
We now turn our attention to a Brownian motion in a 3D spherical domain with a reflecting boundary. The target is supposed to be a sphere of radius $r$ which is fixed at the origin.
The distribution of the first passage time $\tau$ of a BM starting at a distance $x_0$ from the origin to the surface of the target is given explicitly by $$\label{fpt3}
\Psi(\tau) = \frac{D}{Z} \sum_{n=0}^\infty A^{(d=3)}_n(r,x_0,L)
\exp\left(-\lambda_n^2D\tau\right) \ ,$$ where $$A^{(d=3)}_n(r,x_0,L) = \frac{2 u_0(\lambda_n x_0) u_0^\prime(\lambda_n r)}
{G(r,L,\lambda_n)} \ ,$$ $$u_0(x) = y_0(\lambda_n r) j_0(x) - j_0(\lambda_n r) y_0(x) \ ,$$ the derivative $u_0^\prime(\lambda_n r) = du_0(\lambda_n x)/dx|_{x=r}$, and $$\begin{aligned}
G(r,L,\lambda_n) &=& Lj_0(\lambda_nL)
u_0(\lambda_nL)(\lambda_nLj_0(\lambda_nr) - y_0(\lambda_nr))
- \nonumber\\
&-& \frac{j_0(\lambda_nr)y_0(\lambda_nr)}{\lambda_n}
+ \frac{L-r}{r^2\lambda_n^4} \ .\end{aligned}$$ The set $\{\lambda_n\}$ are the roots of $ y_0(\lambda_n r)
j_1(\lambda_n L) - j_0(\lambda_n r) y_1(\lambda_n L) $, arranged in an ascending order, while $j_n(\cdot)$ and $y_n(\cdot)$ are the spherical Bessel functions of the first and of the second kind, respectively.
![A BM in a sphere with a reflecting boundary. First passage time distribution $\Psi(\tau)$ in Eq. (\[fpt3\]) for $L= 100$, $r=5$, $D = 1/2$ and different values of $x_0/L$. []{data-label="fpt3d"}](FIG12.eps){width="70.00000%"}
The distribution of the first passage time in Eq. (\[fpt2\]) is plotted in Fig. \[fpt3d\] for three different values of $x_0/L$. An intermediate power-law $\sim \tau^{-3/2}$ is apparent for $x_0/L=0.055$, persists for about a decade for $x_0/L=0.3$ and is entirely absent for $x_0/L=0.9$, [[*i.e.*]{}]{}, when the BM starts close to the reflecting boundary.
From Eq. (\[fpt3\]), we obtain the normalized distribution $P(\omega)$ for two independent BMs starting at a distance $x_0$ from the origin: $$\begin{aligned}
\label{pw3}
P(\omega) &=& \frac{1}{Z^2} \sum_{n,m=0}^\infty
\frac{A^{(d=3)}_n(r,x_0,L) A^{(d=3)}_m(r,x_0,L)}{\left(\omega \lambda_n^2 + (1 - \omega) \lambda_m^2\right)^2} = \nonumber\\
&=& \frac{1}{Z} \sum_{m=0}^\infty \frac{A^{(d=3)}_m(r,x_0,L)}{\lambda_m^2} \, \Phi\left(\lambda = \frac{1 - \omega}{\omega} D \lambda_m^2\right) \,,\end{aligned}$$ where $\Phi(\lambda)$ is the characteristic function of $\Psi(\tau)$ in Eq. (\[fpt3\]), defined by [@redner] $$\Phi(\lambda) = \frac{r}{x_0} \frac{\sinh\left(\sqrt{\frac{\lambda}{D}} \left(L - x_0\right)\right) - \sqrt{\frac{\lambda}{D}} \, L \, \cosh\left(\sqrt{\frac{\lambda}{D}} \left(L - x_0\right)\right)}{ \sinh\left(\sqrt{\frac{\lambda}{D}} \left(L - r\right)\right) - \sqrt{\frac{\lambda}{D}} \, L \, \cosh\left(\sqrt{\frac{\lambda}{D}} \left(L - r\right)\right)} \,.$$
This distribution is plotted in Fig. (\[pw3d\]) for fixed $L$ and $r$, and several values of $x_0/L$.
![Two BMs in a sphere with a reflecting boundary. The distribution $P(\omega)$ in Eq. (\[pw3\]) for $L= 100$, $r=5$ and different values of $x_0/L$. []{data-label="pw3d"}](FIG13.eps){width="65.00000%"}
Therefore, also in 3D we find that $P(\omega)$ has a different modality depending on the value of the ratio $x_0/L$. For $x_0/L \lesssim 0.68..$, the distribution has an M-shaped form with $\omega = 1/2$ being the least probable value. For $x_0/L > 0.68..$, the distribution has, in principle, a maximum at $\omega = 1/2$ but this maximum is almost invisible so that visually the distribution looks more like a uniform one, as compared to the 1D case in which the maximum is more apparent.
In case when the starting point $x_0$ is uniformly distributed between two concentric spheres of radii $r$ and $L$, we again find $P_{av}(\omega) \equiv 1$.
Two independent BMs in a sphere with an adsorbing boundary
----------------------------------------------------------
Consider finally a geometrically different situation in which two BMs start from the same point within a three-dimensional sphere at a fixed distance $x_0$ from the origin but now the target is the surface of the sphere. In this case, the normalized distribution of the time of the first passage of a BM with diffusion coefficient $D$ to any point on the surface of the sphere of radius $L$ from a point at distance $x_0$ from the origin is given by the series (see, [[*e.g.*]{}]{}, [@net1; @gros]): $$\label{ads1}
\Psi(\tau) = \frac{2 \pi D}{x_0 L} \sum_{n=1}^{\infty} (-1)^{n + 1} n \, \sin\left(\pi n \frac{x_0}{L}\right) \, \exp\left(- \frac{\pi^2 n^2 D \tau}{L^2}\right),$$ whose moments of arbitrary order $m$ are defined as $$\big<\tau^m\big> = \frac{(-1)^m \sqrt{\pi}}{\Gamma(m + 3/2)} \, \frac{L}{x_0} \left(\frac{L^2}{D}\right)^m \, B_{2 m + 1}\left(\frac{x_0 + L}{2 L}\right),$$ where $B_m(.)$ are the Bernoulli polynomials.
![A BM starting at a distance $x_0$ from the origin of a sphere with an adsorbing boundary. The first passage time distribution $\Psi(\tau)$ in Eq. (\[ads1\]) for $L=
100$, $D = 1/2$ and different values of $x_0/L$.[]{data-label="f1"}](FIG14.eps){width="70.00000%"}
The distribution in Eq. (\[ads1\]) is depicted in Fig. \[f1\] for fixed $L$ and $D$, and different values of the ratio $x_0/L$. Note that here the situation is inverse to the one in which the target is situated in the origin - the most pronounced intermediate time power-law behavior is observed for $x_0/L \sim 1$ and is absent for small values of $x_0/L$.
![Two BMs starting at a distance $x_0$ from the origin of a sphere with an adsorbing boundary. The distribution $P(\omega)$ in Eq. (\[ads\]) for $L=
100$ and different values of $x_0/L$.[]{data-label="f2"}](FIG15.eps){width="70.00000%"}
The normalized distribution $P(\omega)$ in this case has the form: $$\begin{aligned}
\label{ads}
P(\omega) &=& \frac{4 L^2}{\pi^2 x_0^2} \sum_{n,m=1}^{\infty} (-1)^{n + m} \frac{n \, m \, \sin\left(\pi n x_0/L\right) \,
\sin\left(\pi m x_0/L\right)}{\left(\omega n^2 + (1 - \omega) m^2\right)^2} \nonumber\\
&=& - \frac{i L^2}{2 \pi x_0^2} \, \frac{1}{\omega^{3/2} \left(1 - \omega\right)^{1/2}} \left. \frac{d}{d y} \ln\left(\frac{\theta_3\left(\frac{x_0 ( 1 + i y)}{2 L}, e^{- \pi y}\right)}{\theta_3\left(\frac{x_0 (1 - i y)}{2 L}, e^{- \pi y}\right)}\right)\right|_{y = \sqrt{\frac{1 - \omega}{\omega}}} \,\end{aligned}$$ where $\theta_3$ is the Jacobi theta-function: $$\theta_3(v,q) = \sum_{m=-\infty}^{\infty} q^{n^2} \exp\left(2 \pi i m v\right).$$ The distribution in Eq. (\[ads\]) is depicted in Fig. \[f2\] for fixed $L$ and different values of $x_0/L$. We observe here a transition from a bell-shaped form with a maximum at $\omega = 1/2$ and an M-shaped form with a minimum at $\omega = 1/2$ and maxima close to $0$ and $1$. The transition takes place at $x_0/L = 0.643...$. Averaging Eq. (\[ads\]) over the starting point $x_0$, we again find that $P_{av}(\omega) \equiv 1$.
The distribution $P(\omega)$ for exponentially-truncated first passage time distributions. Three and more non-communicating searchers
=====================================================================================================================================
We turn finally to the situation with more than two searchers whose first passage time distribution obeys an exponentially-truncated form in Eq. (\[truncated\_dist\]). For $N = 3$ we find the following general result $$\begin{aligned}
\label{k}
P(\omega) &=& \frac{(b/4 a)^{3 \mu/2}}{2 K^3_{\mu}\left(2 \sqrt{a/b}\right)} \frac{\omega^{\mu - 1}}{(1 - \omega)^{\mu + 1}}
\int^{\infty}_0 x^{\mu} dx J_{\mu - 1}\left(x\right) \nonumber\\ &\times& \left(x^2 + \frac{4 a}{b \omega}\right)^{\mu}
K^2_{\mu}\left(\sqrt{\frac{\omega}{1 - \omega}\left(x^2 + \frac{4 a}{b \omega}\right)} \right).\end{aligned}$$
![Three non-communicating Brownian searchers. $P(\omega)$ in Eq. (\[3exp\]) for different values of the parameter $b$ ($a$ is set equal to $1$).[]{data-label="sketch6"}](FIG16.eps){width="65.00000%"}
A straightforward analysis shows that $P(\omega)$ in Eq. (\[k\]) is always a bell-shaped function for $\mu \geq 1$. The most probable value $\omega_m$ is, however, always substantially less than $1/3$, approaching this value only when $\mu \to \infty$ or $b \to 0$.
The case $0 < \mu < 1$ is different. Focusing on $\mu = 1/2$ ([[*e.g.*]{}]{}, biased BMs on a semi-infinite line), for which Eq. (\[k\]) simplifies, $$\label{3exp}
P(\omega) = \frac{4}{\pi} \sqrt{\frac{a}{b}} \frac{e^{6 \sqrt{a/ b}}}{\omega (1 - \omega) \sqrt{1 + 3 \omega}}
K_1\left(2 \sqrt{\frac{a}{b} \frac{1 + 3 \omega}{\omega (1- \omega)}}\right),$$ we discuss a sequence of different regimes which may be observed when $b/a$ is gradually varied, see Fig. \[sketch6\]. For $b/a \ll 1$, $P(\omega)$ is peaked at $\omega_m \approx 1/3$. For larger $b/a$, $\omega_m$ moves towards the origin and $P(\omega_m)$ decreases (see the inset of Fig. \[sketch7\]). For yet larger $b/a$, $\omega_m$ keeps moving towards the origin but now $P(\omega_m)$ passes through a minimum and then starts to grow. At $b/a \approx 140$ a second extremum emerges at $\omega \approx 0.84$ which then splits into a minimum and a maximum (see Fig. \[sketch7\]) so that $P(\omega)$ becomes bimodal. For still larger $b/a$, the minimum moves towards $\omega =
1/2$, while the second maximum moves to $\omega = 1$.
![Three non-communicating Brownian searchers. The loci of the extrema of $P(\omega)$ in Eq. (\[3exp\]). Blue (green) circles define the position of the first (second) maximum. Red circles define the position of the minimum of $P(\omega)$. The inset shows the maximal value $P(\omega_m)$ vs $b$ ($a$ is set equal to $1$).[]{data-label="sketch7"}](FIG17.eps){width="65.00000%"}
In fact, the result in Eq. (\[3exp\]) can be straightforwardly generalized for arbitrary $N$, yielding $$\begin{aligned}
\label{generalBM}
&&P(\omega) = \frac{2 (N - 1)}{\pi} \sqrt{\frac{a}{b}} \frac{e^{2 N \sqrt{a/ b}}}{\omega (1 - \omega) \sqrt{(N - 1)^2 w + 1 - w}} \nonumber \\
&&\times K_1\left(2 \sqrt{\frac{a}{ b} \left(\frac{1}{w} + \frac{(N - 1)^2}{1 - w}\right)}\right) \;.\end{aligned}$$ The distribution function in Eq. (\[generalBM\]) shows essentially the same behavior as the one in Eq. (\[3exp\]); the only difference is that the critical values of the parameter $b/a$ at which $P_{max}$ attains a minimal value or when the second maximum emerges depend on the number of searchers $N$.
Conclusions
===========
To conclude, in this paper we have studied the distribution $P(\omega)$ of the random variable $\omega \sim
\tau_1/\overline{\tau}$, $\overline{\tau} = N^{-1} \sum_{k=1}^N
\tau_k$, where $\tau_k$’s are the first passage times to an immobile target by $N$ independent searchers, which start their random motion simultaneously from the same point in space. Hence, $\tau_k$’s are independent, identically distributed \[with distribution $\Psi(\tau)$\] random variables. Since $\omega$ equals, by definition, the realization-dependent first passage time of a given searcher relative to the realization-dependent ensemble-average first passage time of $N$ searchers, the distribution $P(\omega)$ can be viewed as a measure of the robustness of a given search algorithm and of the underlying random motion (space exploration), which also probes the validity of the *mean* first passage time of a single searcher as a proper and/or informative measure of the search efficiency.
We have considered two general forms of $\Psi(\tau)$: The one in Eq. (\[distribution\]), which is appropriate for search in unbounded domains and is characterized by a power-law long-time tail $\sim
\tau^{-1-\mu}$, where the exponent $\mu $ encodes the specific details of the searchers’ random motion, and an exponentially tempered form in Eq. (\[truncated\_dist\]), which is a plausible approximation for random motion in finite domains or search assisted by a constant bias (“smell”) towards the target.
We have shown that for a non-truncated distribution $\Psi(\tau)$ in Eq. (\[distribution\]) with $\mu < 1$, the distribution of the random variable $\omega$ has a characteristic $U$-shaped form so that the most probable values of $\omega$ are $0$ and $1$. For $N = 2$ the distribution is symmetric around $\omega = 1/2$ with $\omega = 1/2$ being the minimum of the distribution. This signifies that the “symmetry” between two identical searchers is broken. For $N > 2$ the distribution $P(\omega)$ is skewed (by a factor $N^2$) towards the small values of $\omega$ and can be expressed, in an explicit form, as a one-sided $\alpha$-stable distribution with $\alpha = 2 \mu$. For $\mu \geq 1$, the distribution $P(\omega)$ has a bell-shaped form but the most probable value of $\omega$ is always less than the mean value $\langle \omega\rangle = 1/N$. The most probable and the mean values of $\omega$ coincide only when $\mu \to \infty$.
For the exponentially truncated $\Psi(\tau)$ in Eq. (\[truncated\_dist\]) the distribution $P(\omega)$ always has a bell-shaped form for $\mu \geq 1$. For $\mu < 1$, however, the situation is more complicated and interesting. We realized first that for $N = 2$ there exists some critical value of the parameter $y_c(\mu) = b/a$, so that for $b/a < y_c(\mu)$ the distribution $P(\omega)$ has a bell-shaped form, but for $b/a > y_c(\mu)$ it attains an $M$-shaped form with $\omega = 1/2$ being the least probable value and two maxima close to the edges of the interval. This signifies that, despite the fact that $\Psi(\tau)$ has moments of arbitrary order, two identical searchers will arrive for the first time to the target at distinctly different times. Consequently, in such a situation the *mean* first passage is not a proper measure of the search efficiency. We remind that such a form of $\Psi(\tau)$ is exact for a BM taking place on a semi-infinite one-dimensional line with a constant bias directed towards the target. For this physical situation, our result implies that two identical BMs, starting at some point $x_0$ and having the same drift velocity $v < 0$ and the same diffusion coefficient $D$, will most likely arrive together to the target (the origin) if the Peclet number $Pe = x_0
|v|/2 D$ exceeds some critical value $Pe_c \approx 0.666$. On contrary, if $Pe < Pe_c$, an event that these two BMs arrive simultaneously to the target location is the *least* probable event.
Turning next to an unbiased BM on a finite interval $[0,L]$ we recall that $b$ should be proportional to $L^2/D$ and $a \sim
x_0^2/D$. Hence, $\sqrt{a/b} \sim x_0/L$. This suggests a somewhat strange result that for two unbiased identical BMs on a finite 1D interval the modality of the distribution should depend on how far is the starting point from the reflecting boundary. Since $\Psi(\tau)$ in Eq. (\[truncated\_dist\]), which we have used for the derivation of this result, is an approximate form of the first passage time distribution, we have revisited this problem using exact forms of $\Psi(\tau)$ for a BM in 1D, 2D and 3D spherical domains with a reflecting boundary. We have shown that indeed, the very shape (modality) of the distribution $P(\omega)$ depends on the ratio $x_0/L$.
We have realized that for $x_0/L < \chi_c(d)$, where $\chi_c(d) \approx 0.61, 0.64$ and $0.68$ for 1D, 2D and 3D, respectively, $P(\omega)$ is an $M$-shaped function of $\omega$ with a minimum at $\omega = 1/2$ so that here two unbiased identical BMs will most probably arrive to the target location for the first time at distinctly different times. For $x_0/L > \chi_c(d)$, the distribution has a maximum at $\omega = 1/2$ so that, mathematically, the most probable event is that two BMs arrive for the first time to the location of the target simultaneously. Note, however, that $P(\omega)$ is a “bell-shaped” function only in 1D (although is still rather broad) but in 2D and 3D, $P(\omega)$ is nearly flat in an extended region around the maximum and rather abruptly vanishes in the vicinity of the edges of the interval. This signifies that here sample-to-sample fluctuations are very significant. In case when the starting point $x_0$ of two BMs is uniformly distributed within the domain (outside the target), we found that $P(\omega) \equiv 1$. This allows us to conclude that in neither of these well-studied situations (apart of, with some reservations, 1D case with $x_0/L \sim 1$) the mean first passage time of an individual searcher can be considered as a robust measure of the search process efficiency.
We argue that a similar behavior will take place in finite 1D systems for fractional BM with arbitrary Hurst index $H$ or for $\alpha$-stable Lévy flights with $0 < \alpha < 1$, and, more generally, for *finite* systems of (not necessarily integer) dimension $d_f$ with fractal dimension $d_w$ of random motion trajectories given that a) $d_f \leq d_w$ (compact exploration, $\mu =
1 - d_f/d_w < 1$) or b) $d_f > d_w$ (non-compact exploration, $\mu =
d_f/d_w - 1$) but $d_f < 2 d_w$.
We have evaluated $P(\omega)$ in the case of three and more searchers with the exponentially-truncated first passage time distribution in Eq. (\[truncated\_dist\]). We have shown that for $\mu > 1$ the distribution $P(\omega)$ is a bell-shaped function of $\omega$ for any value of $b/a$. Next, we have demonstrated that for very small values of the ratio $b/a$ the distribution $P(\omega)$ has a bell-shaped form with a maximum close to $1/N$. Further on, we have predicted the following sequence of regimes which can be observed upon gradually increasing $b/a$ (the starting point $x_0$ of $N$ searchers is moving towards the location of the target): the most probable value $\omega_m$ of $\omega$ moves towards the origin. The value of the maximum, $P(\omega_m)$ first decreases, passes through a minimal value and then starts to increase. At a certain threshold value of $b/a$, a second extremum emerges in the vicinity of $\omega = 1$, which then splits into a minimum and a maximum so that the distribution $P(\omega)$ becomes a skewed $M$-shaped one. This signifies, as well, that for such a situation the *mean* first passage time of a given searcher is not a representative characteristic of the search process.
As a final observation, we note that one may encounter a power-law distribution with a more abrupt truncation, compared to the exponential function in Eq. (\[truncated\_dist\]), say, a bounded power-law (see [@greg2]) or a power-law tempered from both sides by a Gaussian function: $$\Psi(\tau) = \frac{\left(a \, b\right)^{\mu/2}}{K_{\mu/2}\left(2 a/b\right)} \exp\left( - \frac{a^2}{\tau^2}\right) \, \frac{1}{\tau^{1 + \mu}} \, \exp\left( - \frac{\tau^2}{b^2}\right).$$
![$P(\omega)$ in Eq. (\[oufff\]) for a faster than exponential truncation of the intermediate power-law behavior. Here the parameter $z = 2 a/b$.[]{data-label="sketch9"}](FIG18.eps){width="65.00000%"}
In this case, the distribution of the random variable $\omega$ is given by $$\label{oufff}
P(\omega) = \frac{1}{K_{-\mu/2}^2\left(2 a/b\right)} \, \frac{1}{\omega (1 - \omega)} \, K_{-\mu} \left(2 \frac{a \left(\omega^2 + (1 - \omega)^2\right)}{b \omega (1 - \omega)}\right).$$ One finds that $P(\omega)$ in Eq. (\[oufff\]) is always a bell-shaped function for $\mu \geq 1$. For $\mu < 1$, depending on the value of $z = 2 a/b$, it may have a unimodal, or a three-modal form.
Acknowledgments {#acknowledgments .unnumbered}
===============
The authors wish to thank O. Bénichou, C. Godrèche, S. N. Majumdar, I. M. Sokolov and M. Vergassola for helpful discussions.
References {#references .unnumbered}
==========
[99]{}
D. W. Stephens and J. R. Krebs, [*Foraging Theory*]{}, (Princeton, NJ, Princeton University Press, 1986). J. W. Bell, [*Searching Behavior, The Behavioural Ecology of Finding Resources*]{}, (Chapman and Hall Animal Behaviour Series, London, Chapman and Hall, 1991) J. Klafter, M. Shlesinger and G. Zumofen, [*Physics Today*]{} [**49**]{} 33 (1996). see, e.g., C. Loverdo, O. Bénichou, M. Moreau and R. Voituriez, Nature Physics [**4**]{}, 134 (2008); O. Bénichou, M. Coppey, M. Moreau and G. Oshanin, J. Chem. Phys. [**123**]{}, 194506 (2005), and references therein. M. J. Ward and J. B. Keller, SIAM J. Appl. Math. [**53**]{}, 770 (1993); I. V. Grigoriev, Y. A. Makhnovskii, A. M. Bereshkovskii and V. Y. Zitserman, J. Chem. Phys. [**116**]{}, 9574 (2002); O. Bénichou and R. Voituriez, Phys. Rev. Lett. [**100**]{}, 168105 (2008); O. Bénichou et al., Phys. Rev. Lett. [**105**]{}, 150606 (2010). G. Oshanin, M. Tamm and O. Vasilyev, J. Chem. Phys. [**132**]{}, 235101 (2010). O. G. Berg, R. B. Winter and P.H. Von Hippel, Biochemistry [**20**]{} 6929 (1981); M. Coppey, O. Bénichou, R. Voituriez and M. Moreau, Biophys. J. [**87**]{} 1640 (2004); M. Slutsky and L. Mirny, Biophys. J. [**87**]{}, 4021 (2004); I. M. Sokolov, R. Metzler, K. Pant and M. C. Williams, Biophys. J [**89**]{}, 895 (2005); L. Hu, A. Yu. Grosberg and R. Bruinsma, Biophys. J. [**95**]{}, 1151 (2008); B. van den Broek et al., Proc. Natl. Acad. Sci. USA 105, 15738 (2008); I. Bonnet et al., Nucleic Acids Res. [**36**]{}, 4118 (2008); M. A. Lomholt et al., PNAS [**106**]{}, 8204 (2009). G. Kabatiansky and G. Oshanin, J. Phys. A [**42**]{}, 434016 (2009). G. Oshanin, O. Vasilyev, P. Krapivsky and J. Klafter, Proc. Natl. Acad. Sci. USA [**106**]{}, 13696 (2009). A. Kamimura and T. Ohira, New J. Phys. [**12**]{}, 053013 (2010); T. Vicsek, Nature [**466**]{}, 43 (2010). M. M. Chun and J. M. Wolfe, Cognitive Physiology [**30**]{}, 39 (1996); T. S. Horowitz and J. M. Wolfe, Nature [**394**]{}, 575 (1998). O. Bénichou et al., Phys. Rev. Lett. [**84**]{}, 511 (2000); O. Bénichou et al., Phys. Rev. B [**63**]{}, 235413 (2001); S. C. Weber, A. J. Spakowitz and J. A. Theriot, Phys. Rev. Lett. [**104**]{}, 238102 (2010); J.-H. Jeon and et al., Phys. Rev. Lett. [**106**]{}, 048103 (2011); G. Oshanin and C. Mejía-Monasterio, Soft Matter [**7**]{}, 993 (2011). G. H. Wadhams and J. P. Armitage, Nat. Rev. Mol. Cell Biol. [**5**]{}, 1024 (2004); P. G. de Gennes, Eur. Biophys. J. [**33**]{}, 691 (2004); M. Vergassola, E. Villermaux and B. I. Shraiman, Nature [**445**]{}, 406 (2007); A. Celani and M. Vergassola, Proc. Natl. Acad. Sci. USA [**107**]{}, 1391 (2010); J. A. Revelli, F. Rojo, C. E. Budde and H. S. Wio, J. Phys. A: Math. Theor. [**43**]{}, 195001 (2010); A. Sengupta, T. Kruppa and H. Löwen, Phys. Rev. E [**83**]{}, 031914 (2011). J. B. Masson, M. Bailly-Bechet and M. Vergassola, J. Phys A [**42**]{}, 434009 (2009). E. Gelenbe, Phys. Rev. E [**82**]{}, 061112 (2010). M. R. Evans and S. N. Majumdar, Phys. Rev. Lett. [**106**]{}, 160601 (2011). L. D. Stone, [*Theory of Optimal Search*]{}, (Arlington, VA, Operations Research of America, 1989) A. M. Edwards et al., Nature [**449**]{}, 1044 (2007). G. Ramos-Fernandez et al., Behav. Ecol. Sociobiol. [**55**]{}, 223 (2003); D. Boyer et al., Proc. Biol. Sci. [**273**]{} 1743 (2006). D. L. Kramer and R. L. MacLaughlin, Am. Zool. [**41**]{}, 137 (2001). O. Bénichou et al., Phys. Rev. Lett. [**94**]{} 198101 (2005); O. Bénichou, M. Coppey, M. Moreau and R. Voituriez, Europhys. Lett. [**75**]{} 349 (2006); M. Moreau, O. Bénichou, C. Loverdo and R. Voituriez, J. Stat. Mech. [**2009**]{} P12006 (2009). G. Oshanin, H. S. Wio, K. Lindenberg and S.F. Burlatsky, J. Phys.: Condens. Matter [**19**]{} 065142 (2007); G. Oshanin, K. Lindenberg, H. S. Wio and S. Burlatsky, J. Phys. A: Math. Theor. [**42**]{}, 434008 (2009); F. Rojo et al., J. Phys. A: Math. Theor. [**43**]{}, 345001 (2010); F. Rojo, P. A. Pury and C. E. Budde, Phys. Rev. E [**83**]{}, 011116 (2011). J. Newby and P. C. Bressloff, J. Stat. Mech. [**2010**]{}, P04014 (2010). I. G. Portillo, D. Campos and V. Méndez, J. Stat. Mech. [**2011**]{}, P02033 (2011). M. A. Lomholt, T. Koren, R. Metzler and J. Klafter, Proc. Natl. Acad. Sci. USA [**105**]{}, 11055 (2008). I. Eliazar, Physica A [**356**]{}, 207 (2005). G. Oshanin and S. Redner, Europhys. Lett. [**85**]{}, 10008 (2009). I. Eliazar and I. M. Sokolov, J. Phys. A [**43**]{}, 055001 (2010). I. M. Sokolov and I. I. Eliazar, Phys. Rev. E [**81**]{}, 026107 (2010). G. Oshanin and G. Schehr, Quantitative Finance, to appear; arXiv:1005.1760v1. G. Oshanin, Y. Holovatch and G. Schehr, Physica A, to appear. S. N. Majumdar, Curr. Sci. [**77**]{}, 370 (1999).
S. Redner, [*A Guide to First-Passage Processes*]{}, (Cambridge, Cambridge University Press, 2001). E. Sparre Andersen, Math. Scand. [**1**]{}, 263 (1953); ibid [**2**]{}, 195 (1954). A. Chechkin et al., J. Phys. A: Math. Gen. [**36**]{}, L537 (2003); T. Koren et al., Phys. Rev. Lett. [**99**]{}, 160602 (2007). A. Hansen, T. Engoy and K. J. Maloy, Fractals [**2**]{}, 527 (1994); S. Maslov, M. Paczuski and P. Bak, Phys. Rev. Lett. [**73**]{}, 2162 (1994); M. Ding and W. Yang, Phys. Rev. E [**52**]{}, 207 (1995); J. Krug et al., Phys. Rev. E [**56**]{}, 2702 (1997); G. M. Molchan, Commun. Math. Phys. [**205**]{}, 97 (1999). P. G. de Gennes, J. Chem. Phys. [**76**]{}, 3316 (1982). G. S. Oshanin, S. F. Burlatsky and A. A. Ovchinnikov, Phys. Lett. A [**139**]{}, 245 (1989). G. Oshanin, M. Moreau and S. F. Burlatsky, Adv. Colloid Interface Sci. [**49**]{}, 1 (1994). G. L. Gershtein and B. B. Mandelbrot, Biophys. J. [**4**]{}, 41 (1964). A. R. Lima, C. F. Moukarzel, I. Grosse and T. J. P. Penna, Phys. Rev. E [**61**]{}, 2267 (2000). H. Flyvbjerg, K. Sneppen and P. Bak, Phys. Rev. Lett. [**71**]{}, 4087 (1993). G. Oshanin, A. Mogutov and M. Moreau, J. Stat. Phys. [**73**]{}, 379 (1993); G. Oshanin, S. F. Burlatsky, M. Moreau and B. Gaveau, Chem. Phys. [**177**]{}, 803 (1993); C. Monthus and A. Comtet, J. Phys. I France [**4**]{}, 635 (1994). K. Lindenberg, V. Seshadri, K. E. Shuler and G. H. Weiss, J. Stat. Phys. [**23**]{}, 11 (1980). S. Redner and P. L. Krapivsky, Am. J. Phys. [**67**]{}, 1277 (1999). P. L. Krapivsky, S. N. Majumdar and A. Rosso, J. Phys. A: Math. Theor. [**43**]{}, 315001 (2010). O. Bénichou et al., J. Phys. A: Math. Gen. [**36**]{} 7225 (2003). S. B. Yuste, L. Acedo and K. Lindenberg, Phys. Rev. [**64**]{}, 052102 (2001). O. Bénichou, Y. Kafri, M. Sheinman and R. Voituriez, Phys. Rev. Lett. [**103**]{}, 138102 (2009). S. N. Majumdar and A. Comtet, Phys. Rev. Lett. [**89**]{}, 060601 (2002); S. Sabhapandit, S. N. Majumdar and A. Comtet, Phys. Rev. E [**73**]{}, 051102 (2006). B. D. Hughes, [*Random walks and random environments*]{}, (Clarendon Press, Oxford, 1995). J. F. C. Kingman, Acta Mathematica [**109**]{}, 11 (1963). V. Kolokoltsov, Proc. London Math. Soc. [**80**]{}, 725 (2000). P. Lévy, Comp. Math. [**7**]{}, 283 (1939). E. N. Govorun et al., Phys. Rev. E [**64**]{}, 040903(R) (2001).
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'We uncover the full null structure of the Maxwell-Dirac system in Lorenz gauge. This structure, which cannot be seen in the individual component equations, but only when considering the system as a whole, is expressed in terms of tri- and quadrilinear integral forms with cancellations measured by the angles between spatial frequencies. In the 3D case, we prove frequency-localized $L^2$ space-time estimates for these integral forms at the scale invariant regularity up to a logarithmic loss, hence we obtain almost optimal local well-posedness of the system by iteration.'
address:
- |
Department of Mathematics\
University of Rome “La Sapienza”\
Piazzale Aldo Moro 2\
I-00185 Rome\
Italy
- |
Department of Mathematics\
University of Ferrara\
Via Macchiavelli 35\
I-44100 Ferrara\
Italy
- |
Department of Mathematical Sciences\
Norwegian University of Science and Technology\
Alfred Getz’ vei 1\
N-7491 Trondheim\
Norway
author:
- 'Piero D’Ancona'
- Damiano Foschi
- Sigmund Selberg
bibliography:
- 'mybibliography.bib'
title: 'Null structure and almost optimal local well-posedness of the Maxwell-Dirac system'
---
Introduction {#A}
============
In this paper we uncover the complete null structure of the Maxwell-Dirac system (M-D) in Lorenz gauge. This structure is expressed in terms of tri- and quadrilinear integral forms with certain cancellations measured by the angles between spatial frequencies. In the 3D case, we prove frequency-localized $L^2$ space-time estimates for these integral forms, at the optimal (i.e., scale invariant) regularity up to a logarithmic loss, and as a consequence we obtain almost optimal local well-posedness of the system by iteration.
The null structure that we have found is not the usual *bilinear null structure* that may be seen in bilinear terms of each individual component equation of a system, but instead depends on the structure of the system as a whole, hence we call it *system null structure*. System null structure has been found first by Machedon and Sterbenz [@Machedon:2004] for the Maxwell-Klein-Gordon system (M-K-G), and later by the present authors for the Dirac-Klein-Gordon system (D-K-G); see [@Selberg:2007d]; in both cases this structure was used to prove the almost optimal local well-posedness of the respective systems.
A key feature, distinguishing our work from earlier results on the regularity of nonlinear gauge field theories, is that we find a null structure relative to the Lorenz gauge condition, whereas up to now the gauge of choice has been the Coulomb gauge, following the seminal works [@Klainerman:1994b; @Klainerman:1995a], where a bilinear null structure was found for M-K-G and also the Yang-Mills equations (Y-M) in Coulomb gauge. The Coulomb gauge has been widely used since, also in [@Machedon:2004]. For Y-M, the temporal gauge has also been used; see [@Tao:2003].
An obvious point in favor of the Lorenz gauge is the fact that it is Lorentz (*sic*) invariant, which entails that the equations take a much more symmetric form (they become nonlinear wave equations) than in Coulomb gauge, where one obtains a mix of hyperbolic and elliptic equations.
In fact, the only advantage of Coulomb gauge seems to be that the 4-potential $A$ of the electromagnetic field is better behaved in this gauge, whereas in Lorenz gauge it appears to have rather poor regularity properties. So if one walks down the usual path, thinking of M-K-G or M-D as systems of PDEs for either a scalar field $\phi$ or a spinor $\psi$, and the potential $A$, then this system will likely not be well-posed in Lorenz gauge near the scaling regularity. However, there is no compelling reason to take this point of view, because the regularity of $A$ in itself is really of no interest. Instead, it is the electric and magnetic fields $\mathbf E$ and $\mathbf B$ (or equivalently the tensor $F$) which matter, and these are perfectly well-behaved in Lorenz gauge.
The systems M-K-G and M-D are related, as can be seen from the fact that by “squaring” the Dirac part of M-D (which will destroy some of its structure) one obtains an equation that looks like the Klein-Gordon part of M-K-G, but with two bilinear terms added; in [@Selberg:2005] it was shown that these additional terms also have a null structure, in Coulomb gauge. Combining this fact with the null structure found in [@Klainerman:1994b] for M-K-G in Coulomb gauge, one can then conclude that all the bilinear terms in the “squared M-D” are null forms. In view of this, it is conceivable that the analysis in [@Machedon:2004], where almost optimal local well-posedness was proved for M-K-G in Coulomb gauge, could be extended to cover also the “squared” M-D in Coulomb gauge, but we do not try to follow this path, which would only add to the already highly complicated analysis in [@Machedon:2004].
In fact, “squaring” M-D is not a good idea, as it clearly destroys most of the spinorial structure of the system. And it is precisely the spinorial structure which allows us to find some very powerful cancellations, which however have a remarkably simple form, being expressed in terms of the six angles between the four spatial frequencies in a certain quadrilinear space-time integral form.
This structure enables us to prove closed estimates for the iterates of M-D at the scale-invariant data regularity, up to a logarithmic loss, but this turns out to be quite difficult, and takes up most of the paper. Since, as remarked, the systems M-D and M-K-G are related, it is natural to make a comparison of the techniques used here with those applied in [@Machedon:2004]. In fact, some superficial similarities aside, our approaches differ significantly.
As does [@Machedon:2004], we rely on the usual dyadic decompositions of frequency space adapted to the null cone, and angular decompositions for the spatial frequencies taking into account the geometry of interacting null cones.
The key difference is that we do everything within the confines of $L^2$ theory. On the one hand, our spaces are very simple: We use only the standard $L^2$-based Sobolev and wave-Sobolev norms. On the other hand, working entirely in $L^2$ comes at a price: The existing $L^2$ theory for bilinear interactions of waves is not sufficient to handle the quadrilinear space-time integral form which is at the core of the M-D regularity problem, and we develop new techniques to deal with the difficulties that arise. Of course, we do use the $L^2$ bilinear generalizations of the $L^4$ estimate of Strichartz for the homogeneous wave equation (see [@Foschi:2000]), but in addition we require a number of modifications of these estimates, proved by the third author in [@Selberg:2008a].
In [@Machedon:2004], by comparison, highly sophisticated spaces were used, built from Besov versions of the $X^{s,b}$ spaces and Tataru’s outer block norms, but only well-known bilinear estimates were applied ($L^2$ and mixed-norm generalizations of the Strichartz estimates for the wave equation).
We now present the M-D system.
On the Minkowski space-time ${\mathbb{R}}^{1+3}$ we use coordinates $t=x^0$ and $x=(x^1,x^2,x^3)$. The corresponding Fourier variable is denoted $X=(\tau,\xi)$, where $\tau \in {\mathbb{R}}$ and $\xi \in {\mathbb{R}}^3$ correspond to $t$ and $x$, respectively. The partial derivative with respect to $x^\mu$ is denoted $\partial_\mu$, and we also write $\partial_t=\partial_0$, $\nabla = (\partial_1,\partial_2,\partial_3)$. Roman indices $j,k,\dots$ run over $1,2,3$, greek indices $\mu,\nu,\dots$ over $0,1,2,3$, and repeated indices are implicitly summed over these ranges. Indices are raised and lowered using the metric $\operatorname{diag}(-1,1,1,1)$. The conventions regarding indices do not apply to mere enumerations, of course, only to indices relating to the coordinates of space-time.
The M-D system describes an electron self-interacting with an electromagnetic field, and is obtained by coupling Maxwell’s equations and the Dirac equation: $$\begin{gathered}
\label{A:2}
\nabla \cdot \mathbf E = \rho,
\qquad
\nabla \cdot \mathbf B = 0,
\qquad
\nabla \times \mathbf E + \partial_t \mathbf B = 0,
\qquad
\nabla \times \mathbf B - \partial_t \mathbf E = \mathbf J,
\\
\label{A:4}
\left(\boldsymbol\alpha^\mu D_\mu + m\boldsymbol\beta\right)\psi = 0.\end{gathered}$$ The unknowns are the fields $\mathbf E = (E^1,E^2,E^3)$ and $\mathbf B = (B^1,B^2,B^3)$, which are ${\mathbb{R}}^3$-valued functions of $(t,x)$, and the Dirac spinor $\psi$, which is a ${\mathbb{C}}^4$-valued function of $(t,x)$; $m \ge 0$ is the rest mass of the electron. We regard elements of ${\mathbb{C}}^4$ as column vectors, hence it makes sense to premultiply them by the $4\times4$ Dirac matrices $$\label{A:5}
\boldsymbol\alpha^0 = \mathbf I_{4\times4},
\qquad
\boldsymbol\alpha^j =
\begin{pmatrix}
0 & \boldsymbol\sigma^j \\
\boldsymbol\sigma^j & 0 \\
\end{pmatrix},
\qquad
\boldsymbol\beta =
\begin{pmatrix}
\mathbf I_{2 \times 2} & 0 \\
0 & - \mathbf I_{2 \times 2} \\
\end{pmatrix},$$ where the $\boldsymbol\sigma^j$ are the Pauli matrices. The matrices in are all hermitian, and satisfy $(\boldsymbol\alpha^\mu)^2 = (\boldsymbol\beta)^2 = \mathbf I_{4 \times 4}$ and $\boldsymbol\alpha^j \boldsymbol\alpha^k + \boldsymbol\alpha^k \boldsymbol\alpha^j = 0$ for $1 \le j < k \le 3$.
Formally, the second and third equations in are equivalent to the existence of a four-potential $A_\mu = A_\mu(t,x) \in {\mathbb{R}}$, $\mu=0,1,2,3$, such that $$\label{A:6}
\mathbf B = \nabla \times \mathbf A,
\qquad
\mathbf E = \nabla A_0 - \partial_t \mathbf A,$$ where we write $\mathbf A = (A_1,A_2,A_3)$. In the absence of an electromagnetic field, the operator $D_\mu$ in would just be $-i\partial_\mu$, but in the presence of a field $(\mathbf E, \mathbf B)$ represented by $A_\mu$, this must be modified by the minimal coupling transformation, so that $D_\mu$ becomes the gauge covariant derivative $$\label{A:8}
D_\mu = D_\mu^{(A)} = \frac{1}{i}\partial_\mu - A_\mu.$$ To complete the coupling we plug into the Dirac four-current density $$\label{A:10}
J^\mu = {\langle \, \boldsymbol\alpha^\mu\psi , \psi \, \rangle} \qquad (\mu=0,1,2,3),$$ where ${\langle \, z , w \, \rangle}$ is the standard inner product on ${\mathbb{C}}^4$; this splits into the charge density $\rho = J^0 = {| \psi |}^2$ and the three-current density $\mathbf J = (J^1,J^2,J^3)$.
Eqs. – constitute the M-D system; to simplify it, we can express also the first and fourth equations in in terms of $A_\mu$. Thus, is replaced by $$\label{A:12}
\square A_\mu - \partial_\mu ( \partial^\nu A_\nu ) = - J_\mu
\qquad \left( \square = \partial_\mu \partial^\mu = -\partial_t^2 + \Delta \right).$$ Here, by the conventions on indices, $J_0 = - J^0$ and $J_j = J^j$ for $j=1,2,3$.
Then M-D consists of and , coupled by and .
This system is invariant under the *gauge transformation* $$\label{A:14}
\psi \longrightarrow \psi' = e^{i\chi} \psi,
\qquad A_\mu \longrightarrow A_\mu' = A_\mu + \partial_\mu \chi,$$ for any $\chi : {\mathbb{R}}^{1+3} \to {\mathbb{R}}$, called the *gauge function*. Indeed, if $(\psi,A_\mu)$ satisfies M-D, then so does $(\psi',A_\mu')$, in view of the identity $$D_\mu^{(A')}\psi' = e^{i\chi}D_\mu^{(A)}\psi.$$ Since the observables $\mathbf E, \mathbf B, \rho, \mathbf J$ are not affected by , two solutions related by a gauge transformation are physically undistinguishable, and must be considered equivalent. In practice, a solution is therefore a representative of its equivalence class, and we can pick a representative whose potential $A_\mu$ is chosen so that it simplifies the analysis as much as possible. This is known as the *gauge freedom*. In this paper we impose the Lorenz gauge condition (due to Ludvig Lorenz, not Hendrik Lorentz), $$\label{A:16}
\partial^\mu A_\mu = 0 \qquad \left( \iff \partial_t A_0 = \nabla \cdot \mathbf A \right),$$ which greatly simplifies . Then the M-D system becomes $$\begin{aligned}
\label{A:20}
\left(-i\boldsymbol\alpha^\mu \partial_\mu + m\boldsymbol\beta\right)\psi &= A_\mu \boldsymbol\alpha^\mu \psi,
\\
\label{A:22}
\square A_\mu &= - {\langle \, \boldsymbol\alpha_\mu\psi , \psi \, \rangle},
\\
\label{A:24}
\partial^\mu A_\mu &= 0. \end{aligned}$$ We consider the initial value problem starting from data $$\label{A:30}
\psi(0,x) = \psi_0(x) \in {\mathbb{C}}^4,
\quad \mathbf E(0,x) = \mathbf E_0(x) \in {\mathbb{R}}^3,
\\
\quad \mathbf B(0,x) = \mathbf B_0(x) \in {\mathbb{R}}^3,$$ which in view of the first two equations in must satisfy the constraints $$\label{A:32}
\nabla \cdot \mathbf E_0 = {| \psi_0 |}^2,
\qquad
\nabla \cdot \mathbf B_0 = 0.$$ The initial data for the four-potential $A_\mu$, which we denote by $$\label{A:40}
A_\mu(0,x) = a_\mu(x) \in {\mathbb{R}},
\qquad
\partial_t A_\mu(0,x) = \dot a_\mu(x) \in {\mathbb{R}}\qquad (\mu=0,1,2,3),$$ must be constructed from the observable data $(\mathbf E_0,\mathbf B_0)$. We write $\mathbf a = (a_1,a_2,a_3)$ and $\dot{\mathbf a} = (\dot a_1,\dot a_2,\dot a_3)$. By and we get the constraints $$\label{A:42}
\dot a_0 = \nabla \cdot \mathbf a.
\qquad
\mathbf B_0 = \nabla \times \mathbf a,
\qquad
\mathbf E_0 = \nabla a_0 - \dot{\mathbf a},$$ which determine $\mathbf a, \dot{\mathbf a}$, given $a_0,\dot a_0$. The simplest choice is $$\label{A:48}
a_0 = \dot a_0 = 0.$$ Then $\mathbf a,\dot{\mathbf a}$ are determined by .
The Lorenz gauge condition is automatically satisfied throughout the time interval of existence, for data satisfying and . It suffices to prove this for smooth solutions, since the solutions that we later obtain are limits of smooth solutions. So assume $(\psi,A_\mu)$ is a smooth solution of and on a time interval $(-T,T)$, with the $A_\mu$’s real-valued, and set $u=\partial^\mu A_\mu$. A calculation using yields $\square u = - 2 \operatorname{Im}{\langle \, -i\boldsymbol\alpha^\mu \partial_\mu \psi , \psi \, \rangle}$, and implies $\square u = 0$; here we use the hermiticity of $\boldsymbol\alpha^\mu$ and $\boldsymbol\beta$. Moreover, $u(0)=\partial_t u(0) = 0$, on account of and , so we conclude that $u(t)=0$ for all $t \in (-T,T)$.
Thus, the Lorenz condition can be removed from the system once we have data satisfying the proper constraints, and we are left with the equations and , but these can be combined into a single nonlinear Dirac equation by splitting the four-potential into its homogeneous and inhomogeneous parts: $$\begin{gathered}
\label{A:60}
A_\mu = A_\mu^{\mathrm{hom.}} + A_\mu^{\mathrm{inh.}},
\\
\label{A:62}
\square A_\mu^{\mathrm{hom.}} = 0,
\qquad
A_\mu^{\mathrm{hom.}}(0,x) = a_\mu(x),
\qquad
\partial_t A_\mu^{\mathrm{hom.}}(0,x) = \dot a_\mu(x),
\\
\label{A:64}
A_\mu^{\mathrm{inh.}} = - \square^{-1} {\left\langle \, \boldsymbol\alpha_\mu\psi , \psi \, \right\rangle}.\end{gathered}$$ Here we use the notation $\square^{-1}F$ for the solution of the inhomogeneous wave equation $\square u = F$ with vanishing data at time $t=0$.
Thus, M-D in Lorenz gauge has been reduced to the nonlinear Dirac equation $$\label{A:70}
\left(-i\boldsymbol\alpha^\mu \partial_\mu + m\boldsymbol\beta\right)\psi
= A_\mu^{\mathrm{hom.}} \boldsymbol\alpha^\mu \psi
- \mathcal N(\psi,\psi,\psi),$$ where $$\label{A:72}
\mathcal N(\psi_1,\psi_2,\psi_3)
=
\left( \square^{-1} {\left\langle \, \boldsymbol\alpha_\mu\psi_1 , \psi_2 \, \right\rangle} \right)
\boldsymbol\alpha^\mu\psi_3.$$
In order to uncover the null structure in , we decompose the spinor as $$\label{B:12}
\psi=\psi_+ + \psi_-,
\qquad
\qquad \psi_\pm \equiv \mathbf\Pi_\pm \psi,$$ where $\mathbf\Pi_\pm \equiv \mathbf\Pi(\pm \nabla/i)$ is the multiplier whose symbol is the Dirac projection $$\label{B:8}
\mathbf\Pi(\xi)
=
\frac12 \left( \mathbf I_{4 \times 4} + \frac{\xi^j \boldsymbol\alpha_j}{{| \xi |}} \right) \qquad (\xi \in {\mathbb{R}}^3).$$ To motivate this, we note that the stationary Dirac operator $-i\boldsymbol\alpha^j\partial_j$ has symbol $\xi^j \boldsymbol\alpha_j$ (recall that $\xi \in {\mathbb{R}}^3$ denotes the Fourier variable corresponding to $x$), whose eigenvalues are $\pm{| \xi |}$, with associated eigenspace projections $\mathbf\Pi(\pm\xi)$, as can be seen using the algebraic properties of the Dirac matrices. Note the identities $$\begin{gathered}
\label{B:14}
\mathbf I_{4\times4} = \mathbf\Pi(\xi)+\mathbf\Pi(-\xi),
\qquad
R^j(\xi)\boldsymbol\alpha_j = \mathbf\Pi(\xi)-\mathbf\Pi(-\xi),
\\
\label{B:16}
\mathbf\Pi(\xi)^* = \mathbf\Pi(\xi),
\qquad
\mathbf\Pi(\xi)^2 = \mathbf\Pi(\xi),
\qquad
\mathbf\Pi(\xi)\mathbf\Pi(-\xi) = 0.\end{gathered}$$ In view of the last two identities, it is no surprise that $$\label{B:17}
{| \mathbf\Pi(\xi_1)\mathbf\Pi(-\xi_2)z |} \lesssim {| z |} {\theta}(\xi_2,\xi_2)
\qquad \left( \forall \xi_1,\xi_2 \in {\mathbb{R}}^3 \setminus \{0\}, \, z \in {\mathbb{C}}^4 \right),$$ where $\theta(\xi_1,\xi_2)$ denotes the angle between nonzero vectors $\xi_1,\xi_2$. This estimate, proved in [@Selberg:2007d Lemma 2], is a key tool for identifying spinorial null structures.
The right member of can also be restated as $$\label{B:10}
-i\boldsymbol\alpha^j\partial_j = {| \nabla |}\,\mathbf\Pi_+ - {| \nabla |}\,\mathbf\Pi_-,$$ where ${| \nabla |}$ is the multiplier with symbol ${| \xi |}$. Combining this with and the fact that $\mathbf\Pi(\xi)\boldsymbol\beta = \boldsymbol\beta \mathbf\Pi(-\xi)$, we see that splits into two equations:
\[B:18\] $$\begin{aligned}
\label{B:20}
\left(-i\partial_t+{| \nabla |}\right) \psi_+
=
- m \boldsymbol\beta \psi_-
+ \mathbf\Pi_+ \left( A_\mu^{\mathrm{hom.}} \boldsymbol\alpha^\mu \psi - \mathcal N(\psi,\psi,\psi) \right),
\\
\label{B:22}
\left(-i\partial_t-{| \nabla |}\right) \psi_-
=
- m \boldsymbol\beta \psi_+
+ \mathbf\Pi_-\left( A_\mu^{\mathrm{hom.}} \boldsymbol\alpha^\mu \psi - \mathcal N(\psi,\psi,\psi) \right).\end{aligned}$$
Corresponding to the operators on the left, we define the following spaces.
\[A:Def\] For $s,b \in {\mathbb{R}}$, $X_\pm^{s,b}$ is the completion of the Schwartz space $\mathcal{S}({\mathbb{R}}^{1+3})$ with respect to the norm $$\label{B:24}
{\left\Vert u \right\Vert}_{X_\pm^{s,b}} = {\bigl\Vert {\langle \xi \rangle}^s
{\langle \tau\pm{| \xi |} \rangle}^b \,\widetilde
u(\tau,\xi) \bigr\Vert}_{L^2_{\tau,\xi}},$$ where $\widetilde u(\tau,\xi)$ denotes the Fourier transform of $u(t,x)$, and ${\langle \xi \rangle} = (1+{| \xi |}^2)^{1/2}$.
Given $T > 0$, we denote by $X_\pm^{s,b}(S_T)$ the restriction of $X_\pm^{s,b}$ to the time-slab $$S_T = (-T,T) \times {\mathbb{R}}^3.$$ We recall the well-known fact that $X_\pm^{s,b}(S_T) \hookrightarrow C([-T,T]; H^s)$, for $b > 1/2$.
Our first main result is that is locally well-posed almost down to the critical regularity determined by scaling. To see what this regularity is, observe that in the massless case $m=0$, M-D is invariant under the rescaling $$\psi(t,x) \longrightarrow \frac{1}{L^{3/2}} \psi\left(\frac{t}{L},\frac{x}{L}\right),
\qquad
(\mathbf E,\mathbf B)(t,x) \longrightarrow \frac{1}{L^2} (\mathbf E,\mathbf B)\left(\frac{t}{L},\frac{x}{L}\right),$$ hence the scale invariant data space is $$\psi_0 \in L^2
\qquad
(\mathbf E_0,\mathbf B_0) \in \dot H^{-1/2} \times \dot H^{-1/2},$$ and one does not expect well-posedness with less regularity than this. Our first main result is that local well-posedness holds with only slightly more regularity:
\[A:Thm1\] Let $s > 0$. Assume given initial data with the regularity $$\label{A:100}
\psi_0 \in H^s({\mathbb{R}}^3;{\mathbb{C}}^4),
\qquad
\mathbf E_0, \mathbf B_0 \in H^{s-1/2}({\mathbb{R}}^3;{\mathbb{R}}^3),$$ and satisfying the constraints . Then:
*(a)* (Lorenz data.) There exist $\{a_\mu,\dot a_\mu\}_{\mu=0,1,2,3}$ with $a_0=\dot a_0 = 0$ and $$\label{A:102}
\mathbf a \in {| D |}^{-1} H^{s-1/2}({\mathbb{R}}^3;{\mathbb{R}}^3),
\qquad
\dot{\mathbf a} \in H^{s-1/2}({\mathbb{R}}^3;{\mathbb{R}}^3),$$ and such that the constraint is satisfied.
*(b)* (Local existence.) Use the data $\{a_\mu,\dot a_\mu\}$ from part *(a)* to define $A_\mu^{\mathrm{hom.}}$ as in . Then there exists a time $T > 0$, depending continuously on the norms of the data , and there exists a $$\psi \in C([-T,T];H^s({\mathbb{R}}^3;{\mathbb{C}}^4))$$ which solves on $S_T = (-T,T) \times {\mathbb{R}}^3$ with initial data $\psi(0) = \psi_0$.
*(c)* (Uniqueness.) The solution has the regularity, with notation as in , $$\label{A:104}
\psi_+ \in X^{s,b}_+(S_T),
\qquad
\psi_- \in X^{s,b}_-(S_T),$$ where $b=1/2+\varepsilon$ for $\varepsilon > 0$ sufficiently small, depending on $s$. Moreover, the solution is unique in this regularity class.
The proof is by iteration in the space , so the main challenge is to prove closed estimates in this space; then existence and uniqueness follow by standard arguments (which we do not repeat here), as do persistence of higher regularity and continuous dependence of the solution on the data, which we did not explicitly include in statement of Theorem \[A:Thm1\]. (The latter two properties guarantee that our solutions are limits of smooth solutions, a fact which was used above to reduce the Lorenz gauge condition to constraints on the initial data.)
Since $\psi$ and $A_\mu^{\mathrm{hom.}}$ have very little regularity, it is far from obvious that the nonlinear terms in make sense as distributions (so that it is meaningful to talk about a solution of ). The fact that they do make sense follows from the very estimates that we use to close the iteration.
Let us mention some earlier results for M-D. Local existence of smooth solutions was proved by Gross [@Gross:1966]. Georgiev [@Georgiev:1991] proved global existence for small, smooth data. Bournaveas [@Bournaveas:1996] proved local well-posedness for data with $s > 1/2$; this was improved to $s=1/2$ by Masmoudi and Nakanishi [@Masmoudi:2003].
We remark also that our work leaves open the important question whether M-D is well-posed (globally, for small-norm data) for some scale invariant data space.
Having obtained the solution $\psi$ of , we can immediately construct the full four-potential by defining $A_\mu^{\mathrm{inh.}}$ as in . This has poor regularity properties, however, due to the lack of null structure in the right hand side of , and the fact that $\psi$ only has slightly more than $L^2$ regularity. We do prove that $$\label{A:110}
A_\mu^{\mathrm{inh.}} \in C([-T,T];H^{s-1/2}),$$ but this is a full degree lower than the regularity of the data for $A_\mu$ (cf. ).
But the regularity of $A_\mu$ is not of interest; what matters is the electromagnetic field $(\mathbf E, \mathbf B)$, and this turns out to have much better regularity properties, due to the structure of Maxwell’s equations. In fact, the data regularity (see ) persists throughout the time interval of existence, as our second main result shows:
\[A:Thm2\] Assume that the hypotheses of Theorem \[A:Thm1\] are satisfied, and let $\psi$ be the solution of on $S_T = (-T,T) \times {\mathbb{R}}^3$ obtained in that theorem. Then there exists a unique solution $$(\mathbf E, \mathbf B) \in C([-T,T];H^{s-1/2})$$ of Maxwell’s equations , with the Dirac four-current induced by $\psi$, and with data as in . Moreover, $A_\mu^{\mathrm{inh.}}$, defined by , has the regularity .
Notation {#B}
========
Absolute constants
------------------
In estimates we use the shorthand $X \lesssim Y$ for $X \le CY$, where $C \gg 1$ is some absolute constant; $X=O(R)$ is short for ${| X |} \lesssim R$; $X \sim Y$ means $X \lesssim Y \lesssim X$; $X \ll Y$ stands for $X \le C^{-1} Y$, with $C$ as above. We write $\simeq$ for equality up to multiplication by an absolute constant (typically factors involving $2\pi$). Constants which are not absolute are always denoted explicitly, often with the parameters they depend on as subscripts or arguments.
Fourier transforms and multipliers
----------------------------------
We write $$\begin{aligned}
\mathcal F_x f(\xi) = \widehat f(\xi) &= \int_{{\mathbb{R}}^3} e^{-ix\cdot\xi} f(x) \, dx,
\\
\mathcal F u(\tau,\xi) = \widetilde u(\tau,\xi) &= \int_{{\mathbb{R}}^{1+3}} e^{-i(t\tau+x\cdot\xi)} u(t,x) \, dt \, dx,\end{aligned}$$ where $\tau \in {\mathbb{R}}$ and $\xi \in {\mathbb{R}}^3$; we call $\xi$ the spatial frequency. We write $X=(\tau,\xi)$, and multiple frequencies are numbered by subscript, as in $X_j = (\tau_j,\xi_j)$. Coordinates, on the other hand, are always denoted by superscripts, as in $\xi_j = (\xi_j^1,\xi_j^2,\xi_j^3)$.
If $A \subset {\mathbb{R}}^3, B \subset {\mathbb{R}}^{1+3}$, then ${\mathbf{P}}_A, {\mathbf{P}}_B$ are the multipliers given by $$\widehat{{\mathbf{P}}_A f}(\xi) = \chi_A(\xi)\widetilde f(\xi),
\qquad
\widetilde{{\mathbf{P}}_B u}(X) = \chi_B(X)\widetilde u(X),$$ where $\chi_A, \chi_B$ are the characteristic functions of $A, B$. To simplify the notation when a set is given by some condition, we also write, for example, ${\mathbf{P}}_{{\langle \xi \rangle} \sim N}$ and $\chi_{{\langle \xi \rangle} \sim N}$ instead of ${\mathbf{P}}_{\{ \xi \colon {\langle \xi \rangle} \sim N \}}$ and $\chi_{\{ \xi \colon {\langle \xi \rangle} \sim N \}}$.
Let $D=-i\nabla$, where $\nabla=(\partial_1,\partial_2,\partial_3)$. Given $h : {\mathbb{R}}^3 \to {\mathbb{C}}$, we denote by $h(D)$ the multiplier given by $\widehat{h(D)f}(\xi) = h(\xi)\widehat f(\xi)$.
Bilinear interactions
---------------------
Note the convolution identity $$\label{B:2}
\widetilde{u_1\overline{u_2}}(X_0)
\simeq \int
\widetilde{u_1}(X_1)\,
\widetilde{u_2}(X_2)
\, d\mu^{12}_{X_0},
\quad
d\mu^{12}_{X_0} \equiv \delta(X_0-X_1+X_2) \, dX_1 \, dX_2.$$ Here $\delta$ is the point mass at zero, hence $$\label{B:2:2}
X_0 = X_1 - X_2
\qquad
\left( \iff \text{$\tau_0 = \tau_1-\tau_2$ and $\xi_0 = \xi_1-\xi_2$} \right),$$ motivating the following terminology: A triple $(X_0,X_1,X_2)$ of vectors in ${\mathbb{R}}^{1+3}$ is said to be a *bilinear interaction* if is verified.
Similarly, for a product without conjugation (then $X_0 = X_1 + X_2$) $$\label{B:2:4}
\widetilde{u_1u_2}(X_0)
\simeq \int \widetilde{u_1}(X_1)\widetilde{u_2}(X_2) \, d\nu^{12}_{X_0},
\quad
d\nu^{12}_{X_0} \equiv \delta(X_0-X_1-X_2) \, dX_1 \, dX_2.$$
$L^p$ and Sobolev norms
-----------------------
All $L^p$ norms are taken with respect to Lebesgue measure. In fact, we use almost exclusively $L^2$ norms, hence we reserve the notation ${\left\Vert \cdot \right\Vert}$ for the $L^2$ norm over ${\mathbb{R}}^n$, where $n=1$, $3$ or $1+3$, depending on the context. For example, if $u=u(t,x)$ is a space-time function, then ${\left\Vert u \right\Vert}$ is understood to be taken over ${\mathbb{R}}^{1+3}$. If we are taking the norm of an expression, we indicate by a subscript which variable or variables the norm is taken over, as in ${\Vert F(\tau,\xi) \Vert}_{L^2_{\tau,\xi}}$. The $n$-dimensional Lebesgue-measure of a set $A \subset {\mathbb{R}}^n$ is denoted ${| A |}$, where the value of $n$ will always be clear from the context.
Let $H^s$ be the completion of $\mathcal S({\mathbb{R}}^3)$ with respect to ${\left\Vert f \right\Vert}_{H^s} = {\bigl\Vert {\langle \xi \rangle}^s\widehat f\, \bigr\Vert}_{L^2_\xi}$. Here, and throughout the paper, we use the shorthand $${\langle \xi \rangle} = \left(1 + {| \xi |}^2\right)^{1/2}.$$ An alternative, direct characterization is $H^s = \mathcal F_x^{-1} L^2 \left( {\langle \xi \rangle}^{2s} \, d\xi \right)$. In the statement of Theorem \[A:Thm1\] we also refer to the space ${| D |}^{-1} H^s$, which is defined by $${| D |}^{-1} H^s
= \mathcal F_x^{-1} \left\{ \frac{\widehat g(\xi)}{{| \xi |}{\langle \xi \rangle}^s} \colon g \in L^2({\mathbb{R}}^3) \right\}
= \mathcal F_x^{-1} L^2 \left( {| \xi |}^2{\langle \xi \rangle}^{2s} \, d\xi \right)$$ with norm ${\bigl\Vert {| \xi |}{\langle \xi \rangle}^s\widehat f\, \bigr\Vert}_{L^2_\xi}$. Equivalently, ${| D |}^{-1} H^s$ is the completion of $\mathcal S({\mathbb{R}}^3)$ with respect to this norm.
With this definition, the regularity statement in part (a) of Theorem \[A:Thm1\] holds by the following lemma (with $f=0$ and $\mathbf u = \mathbf B_0$):
\[P:Lemma\] Let $s \in {\mathbb{R}}$ and assume that $f \in H^s({\mathbb{R}}^3;{\mathbb{R}})$ and $\mathbf u \in H^s({\mathbb{R}}^3,{\mathbb{R}}^3)$, with $\nabla \cdot \mathbf u = 0$. Then there exists a unique $\mathbf v \in {| D |}^{-1}H^s( {\mathbb{R}}^3; {\mathbb{R}}^3)$ such that $$\label{P:2}
\nabla \cdot \mathbf v = f,
\qquad
\nabla \times \mathbf v = \mathbf u.$$ Moreover, $\partial_j v^k \in H^s({\mathbb{R}}^3;{\mathbb{R}})$ for $j,k=1,2,3$.
The identity $\nabla \times \nabla \times \mathbf v = \nabla (\nabla \cdot \mathbf v) - \Delta \mathbf v$ tells us that $\mathbf v$ must solve the Poisson equation $\Delta \mathbf v = - \nabla \times \mathbf u + \nabla f$. We therefore define $\mathbf v$ in Fourier space: $$\widehat{\mathbf v}(\xi) = \frac{ \xi \times \widehat{\mathbf u}(\xi)}
{{| \xi |}^2} - \frac{\xi}{{| \xi |}^2} \widehat f(\xi).$$ Then clearly, $\mathbf v \in {| D |}^{-1}H^s( {\mathbb{R}}^3; {\mathbb{R}}^3)$ (see section \[B\] for the definition of this space) and it is easy to check that is satisfied (here the assumption $\nabla \cdot \mathbf u = 0$ is needed). The uniqueness reduces to the fact that if $w \in {| D |}^{-1}H^s$ and $\Delta w = 0$, then $w=0$. To prove this, note that ${| \xi |}^2 \widehat w(\xi) = 0$ in the sense of tempered distributions. But since $w \in {| D |}^{-1}H^s$, we know that $\widehat w$ is a measurable function, and it follows that $\widehat w = 0$ pointwise a.e., hence $w=0$.
Angles
------
Let ${\theta}(a,b)$ be the angle between nonzero $a,b \in {\mathbb{R}}^3$. Then (see [@Selberg:2008a]) $$\begin{aligned}
\label{B:4}
{| a |}+{| b |}-{| a+b |} &\sim \min({| a |},{| b |}) {\theta}(a,b)^2,
\\
\label{B:6}
{| a-b |}-{\bigl\vert {| a |}-{| b |} \bigr\vert} &\sim \frac{{| a |}{| b |}}{{| a-b |}} {\theta}(a,b)^2 \qquad(a \neq b).\end{aligned}$$
Special sets
------------
The characteristic set of the operator $-i\partial_t\pm{| D |}$ appearing in is the null cone component $$\label{B:40}
K^\pm = \left\{ (\tau,\xi) \in {\mathbb{R}}^{1+3} : \tau\pm{| \xi |} = 0 \right\}.$$ The union $K = K^+ \cup K^-$ is the full null cone; we say that a vector $X \in {\mathbb{R}}^{1+3}$ is *null* if it belongs to $K$. The geometry of interactions of null cones plays a fundamental role in the analysis of the system .
For $N,L \ge 1$, $r,\gamma > 0$ and $\omega \in \mathbb S^2$, where $\mathbb S^2 \subset {\mathbb{R}}^3$ is the unit sphere, define $$\begin{aligned}
\label{B:101}
\Gamma_\gamma(\omega) &= \left\{ \xi \in {\mathbb{R}}^3 \colon \theta(\xi,\omega) \le \gamma \right\}
\\
\label{B:108}
T_r(\omega) &= \left\{ \xi \in {\mathbb{R}}^3 \colon {| P_{\omega^{\perp}} \xi |} \lesssim r \right\},
\\
\label{B:102}
K^\pm_{N,L}
&= \left\{ (\tau,\xi) \in {\mathbb{R}}^{1+3} \colon {\langle \xi \rangle} \sim N, \; {\langle \tau\pm{| \xi |} \rangle} \sim L \right\},
\\
\label{B:104}
K^{\pm}_{N,L,\gamma,\omega}
&= \left\{ (\tau,\xi) \in {\mathbb{R}}^{1+3} \colon {\langle \xi \rangle} \sim N, \; \pm\xi \in \Gamma_{\gamma}(\omega), \; {\langle \tau\pm{| \xi |} \rangle} \sim L \right\},
\\
\label{B:106}
H_d(\omega) &= \left\{ (\tau,\xi) \in {\mathbb{R}}^{1+3} \colon {| \tau+\xi\cdot\omega |} \lesssim d \right\},\end{aligned}$$ where $P_{\omega^{\perp}}$ denotes the projection onto the orthogonal complement $\omega^\perp$ of $\omega$ in ${\mathbb{R}}^3$. Thus, $\Gamma_\gamma(\omega)$ is a conical sector around $\omega$, $T_r(\omega)$ is a tube of radius comparable to $r$ around ${\mathbb{R}}\omega$, $K^\pm_{N,L}$ and $K^{\pm}_{N,L,\gamma,\omega}$ consist of pieces of thickened null cones, and $H_d(\omega)$ is an $O(d)$-thickening of the null hyperplane $\tau+\xi\cdot\omega = 0$. Implicit absolute constants are used to make the notation more flexible.
Clearly, $$\label{B:110}
K^{\pm}_{N,L,\gamma,\omega}
\subset {\mathbb{R}}\times T_{N\gamma}(\omega).$$ We also claim that $$\label{B:112}
K^{\pm}_{N,L,\gamma,\omega}
\subset H_{\max(L,N\gamma^2)}(\omega).$$ Indeed, if $(\tau,\xi) \in K^{\pm}_{N,L,\gamma,\omega}$, then $\tau+\xi \cdot \omega$ equals $$\left(\tau\pm{| \xi |}\right)
- \left( \pm {| \xi |} - \xi \cdot \omega \right)
= O(L)
- \frac{{| \xi |}^2\bigl(1-\cos^2\theta(\pm\xi,\omega)\bigr)}{\pm\left({| \xi |}\pm\xi \cdot \omega\right)}
= O(L) + O(N\gamma^2),$$ where we used the fact that $\theta(\pm\xi,\omega) \le \gamma < 1$, hence $\pm\xi \cdot \omega \ge 0$.
Angular decompositions
----------------------
Given $\gamma \in (0,\pi]$ and $\omega \in \mathbb S^2$, define $\Gamma_\gamma(\omega)$ as in . For the purpose of decomposing ${\mathbb{R}}^3$ into such sectors without too much overlap, let $\Omega(\gamma)$ be a maximal $\gamma$-separated subset of the unit sphere $\mathbb S^2$. Then $$\label{B:66}
1 \le \sum_{\omega \in \Omega(\gamma)} \chi_{\Gamma_\gamma(\omega)}(\xi)
\le 5^2
\qquad (\forall \xi \neq 0),$$ where the left inequality holds by the maximality of $\Omega(\gamma)$, and the right inequality by the $\gamma$-separation, since the latter implies (we omit the proof):
\[B:Lemma3\] For $k \in {\mathbb{N}}$ and $\omega \in \mathbb S^2$, $\#\left\{ \omega' \in \Omega(\gamma) : {\theta}(\omega',\omega) \le k\gamma \right\}\le (2k+1)^2$.
The following will be used for angular decomposition in bilinear estimates.
\[B:Lemma4\] We have $$1 \sim \sum_{\genfrac{}{}{0pt}{1}{0 < \gamma < 1}{\text{$\gamma$ \emph{dyadic}}}}
\sum_{\genfrac{}{}{0pt}{1}{\omega_1,\omega_2 \in \Omega(\gamma)}{3\gamma \le {\theta}(\omega_1,\omega_2) \le 12\gamma}}
\chi_{\Gamma_\gamma(\omega_1)}(\xi_1) \chi_{\Gamma_\gamma(\omega_2)}(\xi_2),$$ for all $\xi_1,\xi_2 \in {\mathbb{R}}^3 \setminus \{0\}$ with ${\theta}(\xi_1,\xi_2) > 0$.
We omit the straightforward proof. The condition ${\theta}(\omega_1,\omega_2) \ge 3\gamma$ implies that the minimum angle between vectors in $\Gamma_{\gamma}(\omega_1)$ and $\Gamma_{\gamma}(\omega_2)$ is greater than or equal to $\gamma$, so the sectors are well-separated. If separation is not needed, the following variation may be preferable (again, we skip the easy proof):
\[B:Lemma5\] For any $0 < \gamma < 1$ and $k \in {\mathbb{N}}$, $$\chi_{{\theta}(\xi_1,\xi_2) \le k\gamma}(\xi_1,\xi_2) \lesssim \sum_{\genfrac{}{}{0pt}{1}{\omega_1,\omega_2 \in \Omega(\gamma)}{{\theta}(\omega_1,\omega_2) \le (k+2)\gamma}}
\chi_{\Gamma_\gamma(\omega_1)}(\xi_1) \chi_{\Gamma_\gamma(\omega_2)}(\xi_2),$$ for all $\xi_1,\xi_2 \in {\mathbb{R}}^3 \setminus \{0\}$.
Dyadic decompositions
---------------------
Later we shall bound certain multilinear integral forms in terms of the norms , and we will use an index $j$ to number the functions appearing in the multilinear forms. We rely on a dyadic decomposition based on the size of the weights in . Throughout, $N$’s and $L$’s (indexed by $j$) will denote dyadic numbers greater than or equal to one, i.e., they are of the form $2^m$ for some nonnegative integer $m$. We then assign sizes ${\langle \xi \rangle} \sim N$ and ${\langle \tau\pm{| \xi |} \rangle} \sim L$ to the weights in the norm , and again this will be indexed by $j$. We call the $N$’s and the $L$’s *elliptic* and *hyperbolic* weights, respectively. To have a uniform notation, we shall assume throughout that the $F$’s are arbitrary nonnegative functions in $L^2({\mathbb{R}}^{1+3})$, and that the $u$’s are given by $$\label{B:200}
\widetilde u(X) = \chi_{K^{\pm}_{N,L}}(X) F(X) \qquad (X=(\tau,\xi) \in {\mathbb{R}}^{1+3}, \; F \in L^2({\mathbb{R}}^{1+3}), \; F \ge 0),$$ all of which is indexed by $j$. Often we also use angular decompositions, and for this we use the shorthand, for $\gamma > 0$ and $\omega \in \mathbb S^2$, $$\label{B:202}
u^{\gamma,\omega}
= {\mathbf{P}}_{\pm\xi \in \Gamma_{\gamma}(\omega)} u
\qquad
\left( \iff
\widetilde{u^{\gamma,\omega}}
= \chi_{K^{\pm}_{N,L,\gamma,\omega}} F
\right),$$ where everything except $\gamma$ is subject to indexation by $j$ ($\gamma$ is excepted because it relates two different $j$’s, as in Lemmas \[B:Lemma4\] and \[B:Lemma5\]). Then by , $$\label{B:204}
{\left\Vert u \right\Vert}
\sim
\left( \sum_{\omega \in \Omega(\gamma)}
{\left\Vert u^{\gamma,\omega} \right\Vert}^2 \right)^{1/2}.$$ Finally, we note the following fact, which is used to sum $\omega_1,\omega_2$ in an angularly decomposed bilinear estimate: If $\gamma > 0$, then $$\label{B:208}
\begin{aligned}
\sum_{\genfrac{}{}{0pt}{1}{\omega_1,\omega_2 \in \Omega(\gamma)}
{\theta(\omega_1,\omega_2) \lesssim \gamma}}
{\left\Vert u_1^{\gamma,\omega_1} \right\Vert} {\left\Vert u_2^{\gamma,\omega_2} \right\Vert}
&\le \left( \sum_{\omega_1,\omega_2}
{\left\Vert u_1^{\gamma,\omega_1} \right\Vert}^2 \right)^{1/2}
\left( \sum_{\omega_1,\omega_2}
{\left\Vert u_2^{\gamma,\omega_2} \right\Vert}^2 \right)^{1/2}
\\
&\lesssim {\left\Vert u_1 \right\Vert}{\left\Vert u_2 \right\Vert}.
\end{aligned}$$ Here we applied the Cauchy-Schwarz inequality, then Lemma \[B:Lemma3\] and .
Main estimates, duality and time cut-off {#C}
========================================
We shall iterate the Maxwell-Dirac-Lorenz system in the spaces $$\psi_\pm \in X_\pm^{s,1/2+\varepsilon}(S_T),$$ where $\varepsilon > 0$ is sufficiently small, and $0 < T \le 1$ will depend on the data norm $$\label{C:3}
\mathcal I_0
= {\left\Vert (\psi_0,\mathbf E_0,\mathbf B_0) \right\Vert}_{H^s \times H^{s-1/2} \times H^{s-1/2}}.$$ By a standard argument (see [@Selberg:2007d]) the problem of obtaining closed estimates for the iterates reduces to proving the nonlinear estimates $$\begin{aligned}
\label{C:1}
{\bigl\Vert \mathbf\Pi_{\pm_2}\left( A_\mu^{\mathrm{hom.}} \boldsymbol\alpha^\mu \mathbf\Pi_{\pm_1} \psi_1 \right) \bigr\Vert}_{X_{\pm_2}^{s,-1/2+2\varepsilon}(S_T)}
&\le C
\, \mathcal I_0 {\left\Vert \psi_1 \right\Vert}_{X_{\pm_1}^{s,1/2+\varepsilon}(S_T)},
\\
\label{C:2}
{\bigl\Vert \mathbf\Pi_{\pm_4}\mathcal N\bigl(\mathbf\Pi_{\pm_1}\psi_1,\mathbf\Pi_{\pm_2}\psi_2,\mathbf\Pi_{\pm_3}\psi_3\bigr) \bigr\Vert}_{X_{\pm_4}^{s,-1/2+2\varepsilon}(S_T)}
&\le C \prod_{j=1}^3 {\left\Vert \psi_j \right\Vert}_{X_{\pm_j}^{s,1/2+\varepsilon}(S_T)},\end{aligned}$$ where $C = C_{s,\varepsilon}$. By a standard argument, it suffices to prove these without the restriction to $S_T$, for all $\psi_j \in \mathcal S({\mathbb{R}}^{1+3})$, and of course we can insert a smooth time cut-off $\rho : {\mathbb{R}}\to [0,1]$ with $\rho(t)=1$ for ${| t |} \le 1$ and $\rho(t)=0$ for ${| t |} \ge 2$.
Using also the fact that the dual of $X^{s,-1/2+2\varepsilon}_\pm$ is $X^{-s,1/2-2\varepsilon}_\pm$, we then reduce and to the respective integral estimates $$\begin{aligned}
\label{C:10}
{\left\vert I^{\pm_1,\pm_2} \right\vert}
&\le C_{s,\varepsilon}
\mathcal I_0 {\left\Vert \psi_1 \right\Vert}_{X_{\pm_1}^{s,1/2+\varepsilon}}
{\left\Vert \psi_2 \right\Vert}_{X_{\pm_2}^{-s,1/2-2\varepsilon}},
\\
\label{C:12}
{\left\vert J^{\pm_1,\dots,\pm_4} \right\vert}
&\le C_{s,\varepsilon} {\left\Vert \psi_1 \right\Vert}_{X_{\pm_1}^{s,1/2+\varepsilon}} {\left\Vert \psi_2 \right\Vert}_{X_{\pm_2}^{s,1/2+\varepsilon}} {\left\Vert \psi_3 \right\Vert}_{X_{\pm_3}^{s,1/2+\varepsilon}} {\left\Vert \psi_4 \right\Vert}_{X_{\pm_4}^{-s,1/2-2\varepsilon}},\end{aligned}$$ where $$\begin{aligned}
\label{C:14}
I^{\pm_1,\pm_2}
&=
\iint \rho A_\mu^{\mathrm{hom.}}
{\left\langle \, \boldsymbol\alpha^\mu \mathbf\Pi_{\pm_1} \psi_1 , \mathbf\Pi_{\pm_2} \psi_2 \, \right\rangle} \, dt \, dx,
\\
\label{C:16}
J^{\pm_1,\dots,\pm_4}
&=
\iint \rho \square^{-1} {\left\langle \, \boldsymbol\alpha^\mu\mathbf\Pi_{\pm_1}\psi_1 , \mathbf\Pi_{\pm_2}\psi_2 \, \right\rangle} \cdot{\left\langle \, \boldsymbol\alpha_\mu\mathbf\Pi_{\pm_3}\psi_3 , \mathbf\Pi_{\pm_4}\psi_4 \, \right\rangle} \, dt \, dx,\end{aligned}$$ and the $\psi_j \in \mathcal S({\mathbb{R}}^{1+3})$ are ${\mathbb{C}}^4$-valued.
The time cut-off $\rho$ is included in order to smooth out a singularity in the operator $\square^{-1}$ appearing in $\mathcal N$. In fact, the following holds (see [@Klainerman:1995b]):
\[C:Lemma1\] Given $G \in \mathcal S({\mathbb{R}}^{1+3})$, set $u = \square^{-1} G$. Then $u = u_+ - u_-$, where $$\widehat{u_\pm}(t,\xi) = \frac{e^{\mp it {| \xi |}}}{4\pi{| \xi |}} \int_{-\infty}^\infty \frac{e^{it(\tau'\pm{| \xi |})}-1}{\tau'\pm{| \xi |}} \widetilde G(\tau',\xi) \, d\tau'.$$ Moreover, multiplying by the time cut-off and taking Fourier transform also in time, $$\widetilde{\rho u_\pm}(\tau,\xi)
\\
= \int_{-\infty}^\infty
\frac{\kappa_{\pm}(\tau,\tau';\xi)}{4\pi{| \xi |}}
\widetilde G(\tau',\xi) \, d\tau',$$ where $$\kappa_{\pm}(\tau,\tau';\xi)
= \frac{\widehat\rho(\tau-\tau')-\widehat\rho(\tau\pm{| \xi |})}{\tau'\pm{| \xi |}}$$ and $\widehat \rho(\tau)$ denotes the Fourier transform of $\rho(t)$.
We focus first on the quadrilinear estimate , which is by far the most difficult of the two. The trilinear estimate is proved at the end of the paper.
Write $$\begin{gathered}
s_1=s_2=s_3=s, \quad s_4=-s,
\quad
b_1=b_2=b_3=\frac12 + \varepsilon, \quad b_4=\frac12-2\varepsilon,
\\
\widetilde{\psi_j}
= z_j {\bigl\vert \widetilde{\psi_j} \bigr\vert},
\qquad
{\bigl\vert \widetilde{\psi_j}(X_j) \bigr\vert}
= \frac{F_j(X_j)}{{\langle \xi_j \rangle}^{s_j} {\langle \tau_j\pm_j{| \xi_j |} \rangle}^{b_j}}
\qquad (X_j = (\tau_j,\xi_j)), \end{gathered}$$ where $z_j : {\mathbb{R}}^{1+3} \to {\mathbb{C}}^4$ is measurable, ${| z_j |}=1$ and $F_j \in L^2({\mathbb{R}}^{1+3)}$, $F_j \ge 0$. Applying Plancherel’s theorem, Lemma \[C:Lemma1\] and to , we see that it is enough to prove for $$\label{C:26}
J^{\boldsymbol\Sigma}
=
\int
\frac{\kappa_{\pm_0}(\tau_0,\tau_0';\xi_0)}{{| \xi_0 |}}
\cdot
\frac{q_{1234} \prod_{j=1}^4 F_j(X_j)}
{\prod_{j=1}^4 {\langle \xi_j \rangle}^{s_j} {\langle \tau_j\pm_j{| \xi_j |} \rangle}^{b_j}}
\, d\mu^{12}_{X_0'} \, d\mu^{43}_{X_0}
\, d\tau_0 \, d\tau_0' \, d\xi_0,$$ where $$\begin{gathered}
\boldsymbol\Sigma = (\pm_0,\pm_1,\pm_2,\pm_3,\pm_4),
\\
X_0' = (\tau_0',\xi_0), \qquad X_0 = (\tau_0,\xi_0),
\qquad X_j = (\tau_j,\xi_j), \;\; j=1,\dots,4,
\\
\label{C:32}
e_j = \pm_j \frac{\xi_j}{{| \xi_j |}} \in \mathbb S^2 \qquad (j=0,\dots,4),
\\
\label{C:30}
q_{1234} = {\langle \, \boldsymbol\alpha^\mu\mathbf\Pi(e_1)z_1(X_1) , \mathbf\Pi(e_2)z_2(X_2) \, \rangle}
{\langle \, \boldsymbol\alpha_\mu\mathbf\Pi(e_3)z_3(X_3) , \mathbf\Pi(e_4)z_4(X_4) \, \rangle}.\end{gathered}$$ Note the implicit summation over $\mu=0,1,2,3$ in . The convolution measures $d\mu^{12}_{X_0'}, d\mu^{43}_{X_0}$ are defined as in , hence, in , $$\begin{gathered}
\label{C:35:1}
X_0' = X_1 - X_2, \qquad
X_0 = X_4 - X_3,
\\
\label{C:35}
\tau_0'=\tau_1-\tau_2, \qquad \tau_0=\tau_4-\tau_3,
\qquad
\xi_0=\xi_1-\xi_2=\xi_4-\xi_3.\end{gathered}$$ We may restrict the integration in to the region where $\xi_j \neq 0$ for $j=0,\dots,4$, hence the unit vectors $e_j$ are well-defined, as are the angles $$\label{C:40}
\theta_{jk} = {\theta}(e_j,e_k) = {\theta}(\pm_j\xi_j,\pm_k\xi_k) \qquad (j,k=0,\dots,4),$$ which play a key role in our analysis.
The part of corresponding to ${| \xi_0 |} \le 1$ is easy to treat (see section \[R\]), so for the moment we shall restrict to ${| \xi_0 |} \ge 1$, hence we replace the weight ${| \xi_0 |}$ in by ${\langle \xi_0 \rangle}$. Now assign dyadic sizes to the weights in (recall the convention that $N$’s and $L$’s are dyadic numbers greater than or equal to one) $${\langle \tau_0'\pm_0{| \xi_0 |} \rangle} \sim L_0',
\qquad
{\langle \tau_j\pm_j{| \xi_j |} \rangle} \sim L_j,
\qquad
{\langle \xi_j \rangle} \sim N_j
\qquad (j=0,\dots,4),$$ and define $${\boldsymbol{N}}= (N_0,\dots,N_4),
\qquad
{\boldsymbol{L}}= (L_0,L_0',L_1,\dots,L_4).$$ We will use the shorthand $${N_{\mathrm{min}}}^{012} = \min(N_0,N_1,N_2),
\qquad
{N_{\mathrm{max}}}^{012} = \max(N_0,N_1,N_2),$$ and similarly for the $L$’s and for other index sets than $012$. In particular, an index $0'$ will refer to $L_0'$, so $${L_{\mathrm{min}}}^{0'12} = \min(L_0',L_1,L_2)$$ and $${L_{\mathrm{min}}}^{00'} = \min(L_0,L_0'),$$ for example. In the case of a three-index such as $0'12$ we also let ${L_{\mathrm{med}}}^{0'12}$ denote the median.
By , $\xi_0=\xi_1-\xi_2$ in , so by the triangle inequality, $N_j \lesssim N_k+N_l$ for all permutations $(j,k,l)$ of $(0,1,2)$, hence one of the following must hold:
\[C:41\] $$\begin{aligned}
{2}
\label{C:41a}
N_0 &\ll N_1 \sim N_2& \qquad &(\text{``low output''}),
\\
\label{C:41c}
N_0 &\sim {N_{\mathrm{max}}}^{12} \ge {N_{\mathrm{min}}}^{12}& \qquad &(\text{``high output''}),\end{aligned}$$
and similarly for the index 034. In view of , ${N_{\mathrm{med}}}^{012} \sim {N_{\mathrm{max}}}^{012}$ and $$\label{C:43}
{N_{\mathrm{min}}}^{012} {N_{\mathrm{max}}}^{012} \sim N_0{N_{\mathrm{min}}}^{12},$$ and similarly for the index 034.
We now pull out the dyadic weights in , using the fact that $$\label{C:48}
\kappa_{\pm}(\tau_0,\tau_0';\xi_0)
\lesssim
\frac{\sigma_{L_0,L_0'}(\tau_0-\tau_0')}{(L_0L_0')^{1/2}},$$ where $$\label{C:60}
\sigma_{L_0,L_0'}(r)
=
\left\{
\begin{alignedat}{2}
&\frac{1}{{\langle r \rangle}^{2}}& \quad &\text{if $L_0 \sim L_0'$},
\\
&\frac{1}{(L_0L_0')^{1/2}}& \quad &\text{otherwise}.
\end{alignedat}
\right.$$ To prove we note that $$\label{C:49}
\tau_0-\tau_0' = (\tau_0\pm_0{| \xi_0 |}) - (\tau_0'\pm_0{| \xi_0 |}),$$ and apply the following with $p=\tau_0\pm_0{| \xi_0 |}$, $q=\tau_0-\tau_0'$:
\[C:Lemma2\] For any $M \in {\mathbb{N}}$, $$\label{C:46}
{\left\vert \frac{\widehat\rho(p)-\widehat\rho(q)}{p-q} \right\vert}
\le \frac{C_M}{{\langle p-q \rangle} \min({\langle p \rangle},{\langle q \rangle})^M}.$$
Since $\widehat\rho$ is a Schwartz function, ${| \widehat\rho(\tau) |} \le C_M{\langle \tau \rangle}^{-M}$. This immediately implies if ${| p-q |} > 1$. If, on the other hand, ${| p-q |} \le 1$, then ${\langle p-q \rangle} \sim 1$ and ${\langle p \rangle} \sim {\langle q \rangle}$, and using ${| \widehat\rho\,'(\tau) |} \le C_M{\langle \tau \rangle}^{-M}$ we get $${\left\vert \frac{\widehat\rho(p)-\widehat\rho(q)}{p-q} \right\vert}
= {\left\vert \int_0^1 \widehat\rho\,'\bigl(q + \lambda(p-q)\bigr) \, d\lambda \right\vert}
\le \sup_{0 \le \lambda \le 1} \frac{C_M}{{\langle q + \lambda(p-q) \rangle}^M}
\lesssim \frac{C_M}{{\langle q \rangle}^M}.$$
By and , $$\label{C:56}
{| J^{\mathbf\Sigma} |}
\lesssim \sum_{{\boldsymbol{N}},{\boldsymbol{L}}} \frac{N_4^s J_{{\boldsymbol{N}},{\boldsymbol{L}}}^{\mathbf\Sigma}}{N_0(L_0L_0')^{1/2}(N_1N_2N_3)^s
(L_1 L_2 L_3)^{1/2+\varepsilon}L_4^{1/2-2\varepsilon}},$$ where $$\begin{gathered}
\label{C:58}
J_{{\boldsymbol{N}},{\boldsymbol{L}}}^{\mathbf\Sigma}
=
\int
{\left\vert q_{1234} \right\vert}
\sigma_{L_0,L_0'}(\tau_0-\tau_0') \,
\chi_{K^{\pm_0}_{N_0,L_0}}\!\!(X_0) \,
\chi_{K^{\pm_0}_{N_0,L_0'}}\!(X_0')
\\
\times
\widetilde{u_1}(X_1) \widetilde{u_2}(X_2)
\widetilde{u_3}(X_3) \widetilde{u_4}(X_4)
\, d\mu^{12}_{X_0'} \, d\mu^{43}_{X_0}
\, d\tau_0 \, d\tau_0' \, d\xi_0,\end{gathered}$$ with notation as in and (indexed by 1,2,3,4).
Note that the $\psi_j$ do not appear explicitly in . In fact, from now on we can let the $F_j$ be arbitrary, nonnegative $L^2$-functions and the $z_j$ arbitrary measurable ${\mathbb{C}}^4$-valued functions with ${| z_j |}=1$.
We have now reduced to proving estimates for $J_{{\boldsymbol{N}},{\boldsymbol{L}}}^{\mathbf\Sigma}$. These should of course be independent of $s$ and $\varepsilon$, so we would like to have the estimate which exactly balances the weights in when $s=\varepsilon=0$: $$\label{C:80}
J_{{\boldsymbol{N}},{\boldsymbol{L}}}^{\mathbf\Sigma}
\lesssim \bigl( N_0^2 L_0L_0'L_1L_2L_3L_4\bigr)^{1/2}
{\left\Vert u_1 \right\Vert}{\left\Vert u_2 \right\Vert}{\left\Vert u_3 \right\Vert}{\left\Vert u_4 \right\Vert}.$$ In fact, this holds in almost all the interactions, but for a certain case we have only been able to prove it up to a factor $\log{\langle L_0 \rangle}$ on the right hand side.
We have the following result:
\[C:Thm2\] The following estimate holds for all combinations of signs: $$\label{C:80:2}
J_{{\boldsymbol{N}},{\boldsymbol{L}}}^{\mathbf\Sigma}
\lesssim \bigl( N_0^2 L_0L_0'L_1L_2L_3L_4\bigr)^{1/2}
\log{\langle L_0 \rangle}
{\left\Vert u_1 \right\Vert}{\left\Vert u_2 \right\Vert}{\left\Vert u_3 \right\Vert}{\left\Vert u_4 \right\Vert}.$$
The proof of this theorem takes up a large part of the paper (sections \[D:60\]–\[E\]). The logarithmic loss can likely be removed, but for our purposes it is harmless, since we assume $s,\varepsilon > 0$; once Theorem \[C:Thm2\] has been proved, the main quadrilinear estimate follows by a straightforward summation argument.
For later use we define the operator $T^{\pm_0}_{L_0,L_0'}$ by $$\label{C:90}
T^{\pm_0}_{L_0,L_0'}F(\tau_0,\xi_0)
=
\int a^{\pm_0}_{L_0,L_0'}(\tau_0,\tau_0',\xi_0)
F(\tau_0',\xi_0) \, d\tau_0',$$ where $$\label{C:90:2}
a^{\pm_0}_{L_0,L_0'}(\tau_0,\tau_0',\xi_0)
=
\left\{
\begin{alignedat}{2}
&\frac{1}{{\langle \tau_0-\tau_0' \rangle}^{2}}& \quad &\text{if $L_0 \sim L_0'$},
\\
&\frac{\chi_{\tau_0\pm_0{| \xi_0 |}=O(L_0)}
\chi_{\tau_0'\pm_0{| \xi_0 |}=O(L_0')}}{(L_0L_0')^{1/2}}&
\quad &\text{otherwise}.
\end{alignedat}
\right.$$
\[D:Lemma3\] ${\bigl\Vert T^{\pm_0}_{L_0,L_0'}F \bigr\Vert} \lesssim {\left\Vert F \right\Vert}$ for $F \in L^2({\mathbb{R}}^{1+3})$.
By duality, this is equivalent to $$\label{C:92}
{\left\vert \iiint a^{\pm_0}_{L_0,L_0'}(\tau_0,\tau_0',\xi_0)
F(\tau_0',\xi_0) G(\tau_0,\xi_0) \, d\tau_0 \, d\tau_0' \, d\xi_0 \right\vert}
\lesssim {\left\Vert F \right\Vert} {\left\Vert G \right\Vert}.$$ If $L_0 \sim L_0'$, we use the Cauchy-Schwarz inequality with respect to the measure ${\langle \tau-\tau_0 \rangle}^{-2} d\tau_0 \, d\tau_0' \, d\xi_0$, obtaining $$\text{l.h.s.}\eqref{C:92}
\le
\left( \iiint \frac{F^2(\tau_0',\xi_0)}{{\langle \tau_0-\tau_0' \rangle}^{2}} d\tau_0 \, d\tau_0' \, d\xi_0 \right)^{1/2}
\left( \iiint \frac{G^2(\tau_0,\xi_0)}{{\langle \tau_0-\tau_0' \rangle}^{2}} d\tau_0 \, d\tau_0' \, d\xi_0 \right)^{1/2}.$$ If $L_0 \ll L_0'$ or $L_0' \ll L_0$, then by the Cauchy-Schwarz inequality, $$\label{C:92:2}
\begin{aligned}
\text{l.h.s.}\eqref{C:92}
\le
\frac{1}{(L_0L_0')^{1/2}}
&\left( \iiint \chi_{\tau_0\pm_0{| \xi_0 |}=O(L_0)} F^2(\tau_0',\xi_0) d\tau_0 \, d\tau_0' \, d\xi_0 \right)^{1/2}
\\
\times
&\left( \iiint \chi_{\tau_0'\pm_0{| \xi_0 |}=O(L_0')} G^2(\tau_0,\xi_0) d\tau_0 \, d\tau_0' \, d\xi_0 \right)^{1/2}.
\end{aligned}$$
In some situations we use the following variant of the last lemma:
\[D:Lemma5\] Assume that $L_0 \ll L_0'$ or $L_0' \ll L_0$. Let $\omega, \omega' \in \mathbb S^2$, $c,c' \in {\mathbb{R}}$ and $d,d' > 0$. Assume that $F,G \in L^2({\mathbb{R}}^{1+3})$ satisfy $$\begin{aligned}
\operatorname{supp}F &\subset \left\{ (\tau_0',\xi_0) \colon
\tau_0'+\xi_0\cdot\omega' = c' + O(d') \right\},
\\
\operatorname{supp}G &\subset \left\{ (\tau_0,\xi_0) \colon
\tau_0+\xi_0\cdot\omega = c + O(d) \right\}.\end{aligned}$$ Then $${\bigl\Vert T^{\pm_0}_{L_0,L_0'}F \bigr\Vert}
\lesssim \left(\frac{d'}{L_0'}\right)^{1/2} {\left\Vert F \right\Vert},$$ and $${\left\vert \iint T^{\pm_0}_{L_0,L_0'}F(\tau_0,\xi_0) G(\tau_0,\xi_0) \, d\tau_0 \, d\xi_0 \right\vert}
\lesssim
\left(\frac{dd'}{L_0L_0'}\right)^{1/2} {\left\Vert F \right\Vert} {\left\Vert G \right\Vert}.$$
Replace $\chi_{\tau_0'\pm_0{| \xi_0 |}=O(L_0')}$ in by $\chi_{\tau_0'+\xi_0\cdot\omega'=c'+O(d')}$ to get the first estimate above. To get the second estimate we replace also $\chi_{\tau_0\pm_0{| \xi_0 |}=O(L_0)}$ in by $\chi_{\tau_0+\xi_0\cdot\omega=c+O(d)}$.
Bilinear null structure: A review {#D}
=================================
Since bilinear $L^2$ estimates for the spaces $X^{s,b}_\pm$ are well understood, it is natural to try to reduce directly to such estimates by applying the Cauchy-Schwarz inequality in the most obvious way. This approach fails, but it is worthwhile to go through the argument, since it leads us to the quadrilinear null structure.
Given a bounded, measurable $\sigma : {\mathbb{R}}^{1+3} \times {\mathbb{R}}^{1+3} \to {\mathbb{C}}$, we define the bilinear operator $\mathfrak B_\sigma(u_1,u_2)$ by inserting $\sigma(X_1,X_2)$ in the convolution integral : $$\label{D:1}
\mathcal F \mathfrak B_\sigma(u_1,u_2)
(X_0)
= \int
\sigma(X_1,X_2)\,
\widetilde{u_1}(X_1)\,
\widetilde{u_2}(X_2)
\, d\mu^{12}_{X_0}.$$ Given signs $\pm_0,\pm_1,\pm_2$, we say that the bilinear interaction $(X_0,X_1,X_2)$ is *null* if the hyperbolic weights $$\label{D:3}
{\mathfrak h}_j \equiv \tau_j \pm_j {| \xi_j |} \qquad (j=0,1,2)$$ all vanish. This is dangerous, since we then get no help from the hyperbolic weights in the integral (there we actually have two bilinear interactions, and the worst case would be when both are null simultaneously).
In the bilinear null interaction, $X_0,X_1,X_2$ are all null, and since $X_0 = X_1 - X_2$, it is clear geometrically that they must be collinear (otherwise $X_0$ could not end up lying on the null cone). Therefore, the angle $\theta_{12}=\theta(\pm_1\xi_1,\pm_2\xi_2)$ must vanish. It is therefore not surprising that if $\sigma(X_1,X_2) = O(\theta_{12})$, then we have better $L^2$ estimates for $\mathfrak B_\sigma(u_1,u_2)$ than for a generic product like $u_1\overline{u_2}$.
In fact, the following holds (we prove this below):
\[D:Thm\] If the symbol $\sigma$ satisfies $\sigma(X_1,X_2) = O(\theta_{12})$, then with notation as in (indexed by $1,2$), $${\Bigl\Vert {\mathbf{P}}_{K^{\pm_0}_{N_0,L_0}} \mathfrak B_\sigma(u_1,u_2) \Bigr\Vert}
\lesssim (N_0L_0L_1L_2)^{1/2} {\left\Vert u_1 \right\Vert}{\left\Vert u_2 \right\Vert}.$$
This estimate fails to hold for a generic product like $u_1\overline{u_2}$.
Now compare Theorem \[D:Thm\] with the estimate that we want to prove. Let us for simplicity replace $\sigma_{L_0,L_0'}(\tau_0-\tau_0')$ by $\delta(\tau_0-\tau_0')$, so that $\tau_0'=\tau_0$ and $L_0=L_0'$. Clearly, if we could estimate the absolute value of the symbol $q_{1234}$ by a product $$\label{D:4:2}
\sigma_{12}(X_1,X_2) \sigma_{34}(X_3,X_4)$$ such that $$\label{D:4:4}
\sigma_{12}(X_1,X_2) = O(\theta_{12})
\qquad
\sigma_{34}(X_3,X_4) = O(\theta_{34}),$$ where $\theta_{12},\theta_{34}$ are defined as in , then we could apply the Cauchy-Schwarz inequality in , and deduce from Theorem \[D:Thm\].
Let us see if this works. Clearly, the absolute value of the symbol $q_{1234}$ is bounded by the sum over $\mu=0,1,2,3$ of the terms with $$\begin{aligned}
\sigma_{12}(X_1,X_2)
&=
{\left\vert {\left\langle \, \boldsymbol\alpha^\mu\mathbf\Pi(e_1)z_1(X_1) , \mathbf\Pi(e_2)z_2(X_2) \, \right\rangle} \right\vert},
\\
\sigma_{34}(X_3,X_4)
&=
{\left\vert {\left\langle \, \boldsymbol\alpha_\mu\mathbf\Pi(e_3)z_3(X_3) , \mathbf\Pi(e_4)z_4(X_4) \, \right\rangle} \right\vert}.\end{aligned}$$ But these symbols fail to satisfy (take $\mu=0$ and recall that $\boldsymbol\alpha^0 = \mathbf I_{4 \times 4}$).
This is not quite the end of the story, however. Let us see what happens for $\mu = 1,2,3$. Then we can use the commutation identity $$\label{D:54}
\boldsymbol\alpha^j \mathbf\Pi(e) = \mathbf\Pi(-e)\boldsymbol\alpha^j + e^j \mathbf I_{4\times4} \qquad (j=1,2,3; \; e \in \mathbb S^2).$$ If it were not for the remainder term $e^j \mathbf I_{4\times4}$, we could apply the following:
\[D:Lemma4\] Let $\boldsymbol \gamma$ be a $4 \times 4$ matrix. A sufficient condition for the symbol $$\sigma_{12}^{\boldsymbol\gamma}(X_1,X_2)
=
{\left\vert {\left\langle \, \boldsymbol\gamma\mathbf\Pi(e_1)z_1(X_1) , \mathbf\Pi(e_2)z_2(X_2) \, \right\rangle} \right\vert}$$ to satisfy , is that $$\label{D:50}
\boldsymbol\gamma\mathbf\Pi(\xi)=\mathbf\Pi(-\xi)\boldsymbol\gamma
\qquad (\forall \xi \in {\mathbb{R}}^3).$$
By , and , $${\left\vert {\langle \, \boldsymbol\gamma \mathbf\Pi(e_1) z_1 , \mathbf\Pi(e_2) z_2 \, \rangle} \right\vert}
=
{\left\vert {\langle \, \mathbf\Pi(e_2)\mathbf\Pi(-e_1) \boldsymbol\gamma z_1 , z_2 \, \rangle} \right\vert}
\lesssim \theta(e_1,e_2) {| {\boldsymbol{\gamma}}z_1 |}{| z_2 |},$$ so follows, since ${| z_1 |}={| z_2 |}=1$.
For the simpler Dirac-Klein-Gordon system (D-K-G), the analogue of is obtained by replacing the $\boldsymbol\alpha$’s in by $\boldsymbol\beta$, which does satisfy . This was used in [@Selberg:2007d] to prove almost optimal local well-posedness for D-K-G in 3D. (Essentially this reduces to Theorem \[D:Thm\].)
For M-D, the part of which corresponds to the sum over $\mu=1,2,3$, and which does not take into account the remainder term in , can be treated in the same way as D-K-G. The crucial point, however, is that the remainder term can be combined with the term corresponding to $\mu=0$ in , to produce a more complicated null structure, which is not bilinear, but quadrilinear; see section \[D:60\]. In view of this, we cannot just rely on standard $L^2$ bilinear estimates such as Theorem \[D:Thm\], although these certainly play an important role. In addition, we will use a number of modified bilinear estimates proved by the third author in [@Selberg:2008a]. These estimates are recalled in section \[J\].
To end this section, we recall the standard $L^2$ bilinear estimates of the form $$\label{D:10}
{\Bigl\Vert {\mathbf{P}}_{K^{\pm_0}_{N_0,L_0}} \left( u_1 \overline{u_2} \right) \Bigr\Vert}
\le C
{\left\Vert u_1 \right\Vert} {\left\Vert u_2 \right\Vert}.$$
We have the following:
\[M:Thm\] With notation as in (indexed by $1,2$), the estimate holds with $$\begin{aligned}
\label{M:10}
C
&\sim \bigl( {N_{\mathrm{min}}}^{012}{N_{\mathrm{min}}}^{12} L_1 L_2 \bigr)^{1/2},
\\
\label{M:14}
C
&\sim \bigl( {N_{\mathrm{min}}}^{012}{N_{\mathrm{min}}}^{0j} L_0 L_j \bigr)^{1/2} \qquad (j=1,2),
\\
\label{M:17}
C
&\sim \bigl( N_0{N_{\mathrm{min}}}^{12} {L_{\mathrm{min}}}^{012} {L_{\mathrm{med}}}^{012} \bigr)^{1/2},
\\
\label{M:18}
C
&\sim \bigl( ({N_{\mathrm{min}}}^{012})^3 {L_{\mathrm{min}}}^{012} \bigr)^{1/2},\end{aligned}$$ for any choice of signs $\pm_0,\pm_1,\pm_2$.
By a standard argument (see, for example, [@Selberg:2007d Lemmas 3 and 4]), follows from the analogous estimates for two solutions of the homogeneous wave equation, proved in [@Foschi:2000 Theorem 12.1]. Alternatively, see [@Selberg:2008a] for a short, direct proof of . By duality, follows from . Combining and , and recalling , we then get . Finally, via the Cauchy-Schwarz inequality, reduces to a trivial volume estimate; see [@Tao:2001 Eq. (37)].
It is now easy to prove Theorem \[D:Thm\]. The only other ingredient needed is the following more or less standard result, which generalizes the observation, made above, that the angle $\theta_{12}$ must vanish in the null interaction.
\[D:Lemma1\] Consider a bilinear interaction $(X_0,X_1,X_2)$, with $\xi_j \neq 0$ for $j=1,2$. Given signs $(\pm_0,\pm_1,\pm_2)$, define the hyperbolic weights as in , and define the angle $\theta_{12} = \theta(\pm_1\xi_1,\pm_2\xi_2)$. Then $$\label{D:30}
\max\left( {| {\mathfrak h}_0 |}, {| {\mathfrak h}_1 |}, {| {\mathfrak h}_2 |} \right)
\gtrsim
\min\left({| \xi_1 |},{| \xi_2 |}\right)\theta_{12}^2.$$ Moreover, if $$\label{D:34}
{| \xi_0 |} \ll {| \xi_1 |} \sim {| \xi_2 |}
\qquad \text{and} \qquad
\pm_1 \neq \pm_2,$$ then $$\label{D:31}
\theta_{12} \sim 1
\qquad \text{and} \qquad
\max\left( {| {\mathfrak h}_0 |}, {| {\mathfrak h}_1 |}, {| {\mathfrak h}_2 |} \right)
\gtrsim
\min\left({| \xi_1 |},{| \xi_2 |}\right),$$ whereas if does not hold, then $$\label{D:32}
\max\left( {| {\mathfrak h}_0 |}, {| {\mathfrak h}_1 |}, {| {\mathfrak h}_2 |} \right)
\gtrsim
\frac{{| \xi_1 |}{| \xi_2 |}\theta_{12}^2}{{| \xi_0 |}}.$$
See [@Selberg:2008a] for a proof.
We can now prove Theorem \[D:Thm\]. By Lemma \[D:Lemma1\], $\theta_{12} \lesssim ({L_{\mathrm{max}}}^{012}/{N_{\mathrm{min}}}^{12})^{1/2}$, hence $$\label{D:40}
\begin{aligned}
{\Bigl\Vert {\mathbf{P}}_{K^{\pm_0}_{N_0,L_0}} \mathfrak B_{\theta_{12}}(u_1,u_2) \Bigr\Vert}
&\lesssim \left( \frac{{L_{\mathrm{max}}}^{012}}{{N_{\mathrm{min}}}^{12}} \right)^{1/2}
{\Bigl\Vert {\mathbf{P}}_{K^{\pm_0}_{N_0,L_0}} \left( u_1 \overline{u_2} \right) \Bigr\Vert}
\\
&\lesssim \left( \frac{{L_{\mathrm{max}}}^{012}}{{N_{\mathrm{min}}}^{12}} \right)^{1/2}
\bigl( N_0{N_{\mathrm{min}}}^{12} {L_{\mathrm{min}}}^{012} {L_{\mathrm{med}}}^{012} \bigr)^{1/2}
{\left\Vert u_1 \right\Vert}{\left\Vert u_2 \right\Vert},
\end{aligned}$$ where we used from Theorem \[M:Thm\] to get the last inequality. Simplifying, we get Theorem \[D:Thm\].
We remark that Theorem \[M:Thm\] is related to the so-called null form estimates first investigated in [@Klainerman:1993], and subsequently in numerous other papers by various authors; see [@Foschi:2000] for a survey. For this reason, we call $\mathfrak B_{\theta_{12}}$ a *null form*.
Later, we shall also use the related null form $\mathfrak B_{\theta_{12}}'$ corresponding to a product without conjugation. Let us write out both definitions here for easy reference: $$\begin{aligned}
\label{D:1:1}
\mathcal F \mathfrak B_{\theta_{12}}(u_1,u_2)
(X_0)
&= \int
\theta(\pm_1\xi_1,\pm_2\xi_2)\,
\widetilde{u_1}(X_1)\,
\widetilde{u_2}(X_2)
\, d\mu^{12}_{X_0},
\\
\label{D:1:2}
\mathcal F \mathfrak B'_{\theta_{12}}(u_1,u_2)
(X_0)
&= \int
\theta(\pm_1\xi_1,\pm_2\xi_2)\,
\widetilde{u_1}(X_1)\,
\widetilde{u_2}(X_2)
\, d\nu^{12}_{X_0},\end{aligned}$$ with notation as in and .
Quadrilinear null structure in Maxwell-Dirac {#D:60}
============================================
Consider the symbol $$q_{1234} = {\langle \, \boldsymbol\alpha^\mu\mathbf\Pi(e_1)z_1 , \mathbf\Pi(e_2)z_2 \, \rangle}
{\langle \, \boldsymbol\alpha_\mu\mathbf\Pi(e_3)z_3 , \mathbf\Pi(e_4)z_4 \, \rangle},$$ appearing in . Here $e_1,\dots,e_4 \in {\mathbb{R}}^3$ and $z_1,\dots,z_4 \in {\mathbb{C}}^4$ are unit vectors. The null structure will be expressed in terms of the angles $$\theta_{jk} = {\theta}(e_j,e_k),$$ six of which are distinct. We shall refer to the index pairs 12 and 34 as the *internal pairs*, and the angles $\theta_{12}$ and $\theta_{34}$ as the *internal angles*. Angles between vectors from different internal pairs are then called *external angles*. So the external angles are $\theta_{13}$, $\theta_{14}$, $\theta_{23}$ and $\theta_{24}$. Let us denote their minimum by $$\label{D:70}
\phi = \min \left\{ \theta_{13}, \theta_{14}, \theta_{23}, \theta_{24} \right\}.$$
\[D:Lemma2\] With notation as above, $$\label{D:74}
{| q_{1234} |}
\lesssim
\theta_{12}\theta_{34} + \phi \max(\theta_{12},\theta_{34}) + \phi^2,$$ for all unit vectors $e_1,\dots,e_4 \in {\mathbb{R}}^3$ and $z_1,\dots,z_4 \in {\mathbb{C}}^4$.
By idempotency we may replace $z_j$ in $q_{1234}$ by $z_j' = \mathbf\Pi(e_j)z_j$, for $j=1,\dots,4$. We do this for notational convenience, the advantage being that $$\label{D:76}
z_j' = \mathbf\Pi(e_j)z_j' \qquad (j=1,\dots,4).$$ Note also that ${| z_j' |} \le {| z_j |} = 1$.
Let us first assume $\phi = \theta_{13}$. Apply in $q_{1234}$, and make use of and the identities –. Note the implicit summation in $q_{1234}$, and recall that $\boldsymbol\alpha_0 = - \boldsymbol\alpha^0 = \boldsymbol I_{4 \times 4}$. We thus get $$q_{1234}
= A + B + C + D,$$ where $$\begin{aligned}
A &= (e_1 \cdot e_3-1) {\langle \, z'_1 , z'_2 \, \rangle}
{\langle \, z'_3 , z'_4 \, \rangle},
\\
B &=
{\langle \, \mathbf\Pi(-e_1)\boldsymbol\alpha^j z'_1 , z'_2 \, \rangle}
{\langle \, \mathbf\Pi(-e_3)\boldsymbol\alpha_j z'_3 , z'_4 \, \rangle},
\\
C &=
{\langle \, e_1^j z'_1 , z'_2 \, \rangle}
{\langle \, \mathbf\Pi(-e_3)\boldsymbol\alpha_j z'_3 , z'_4 \, \rangle},
\\
D &=
{\langle \, \mathbf\Pi(-e_1)\boldsymbol\alpha_j z'_1 , z'_2 \, \rangle}
{\langle \, e_3^j z'_3 , z'_4 \, \rangle}.\end{aligned}$$ Here $e_1^j$ are the coordinates of $e_1$, and we implicitly sum over $j=1,2,3$.
Clearly, $$\label{D:80}
{| A |} \le {| 1-\cos\theta_{13} |}
\lesssim \theta_{13}^2.$$ By and , $$\label{D:82}
{| B |} \lesssim \theta_{12}\theta_{34}.$$ Using we write $C = C_1-C_2$, where $$\begin{aligned}
C_1 &= {\langle \, z'_1 , z'_2 \, \rangle}
{\langle \, \mathbf\Pi(-e_3) \mathbf\Pi(e_1) z'_3 , z'_4 \, \rangle},
\\
C_2 &= {\langle \, z'_1 , z'_2 \, \rangle}
{\langle \, \mathbf\Pi(-e_3)\mathbf\Pi(-e_1) z'_3 , z'_4 \, \rangle}.\end{aligned}$$ By and , $$\begin{aligned}
{| C_1 |} \le {| {\langle \, \mathbf\Pi(-e_3) \mathbf\Pi(e_1) z'_3 , \mathbf\Pi(-e_3) \mathbf\Pi(e_4) z'_4 \, \rangle} |}
\lesssim \theta_{13} \theta_{34},
\\
{| C_2 |} \le {| {\langle \, \mathbf\Pi(-e_1) \mathbf\Pi(e_3) z'_3 , \mathbf\Pi(-e_3) \mathbf\Pi(e_4) z'_4 \, \rangle} |}
\lesssim \theta_{13} \theta_{34},\end{aligned}$$ hence $$\label{D:84}
{| C |} \lesssim \theta_{13}\theta_{34}.$$ By symmetry with $C$, $$\label{D:86}
{| D |} \lesssim \theta_{13}\theta_{12}.$$
Combining – gives ${| q_{1234} |} \lesssim \theta_{12}\theta_{34} + \theta_{13}\theta_{12} + \theta_{13}\theta_{34} + \theta_{13}^2$, which proves the lemma under the assumption $\phi=\theta_{13}$. To remove that assumption, we observe that $q_{1234}$ can be written in four different ways, since the $\boldsymbol\alpha^\mu$ are hermitian: $$\begin{aligned}
q_{1234} &= {\langle \, \boldsymbol\alpha^\mu\mathbf\Pi(e_1)z'_1 , \mathbf\Pi(e_2)z'_2 \, \rangle}
{\langle \, \boldsymbol\alpha_\mu\mathbf\Pi(e_3)z'_3 , \mathbf\Pi(e_4)z'_4 \, \rangle}
\\
&= {\langle \, \boldsymbol\alpha^\mu\mathbf\Pi(e_1)z'_1 , \mathbf\Pi(e_2)z'_2 \, \rangle}
{\langle \, \mathbf\Pi(e_3)z'_3 , \boldsymbol\alpha_\mu\mathbf\Pi(e_4)z'_4 \, \rangle}
\\
&= {\langle \, \mathbf\Pi(e_1)z'_1 , \boldsymbol\alpha^\mu\mathbf\Pi(e_2)z'_2 \, \rangle}
{\langle \, \boldsymbol\alpha_\mu\mathbf\Pi(e_3)z'_3 , \mathbf\Pi(e_4)z'_4 \, \rangle}
\\
&= {\langle \, \mathbf\Pi(e_1)z'_1 , \boldsymbol\alpha^\mu\mathbf\Pi(e_2)z'_2 \, \rangle}
{\langle \, \mathbf\Pi(e_3)z'_3 , \boldsymbol\alpha_\mu\mathbf\Pi(e_4)z'_4 \, \rangle},\end{aligned}$$ and by the same argument as above, these expressions can be used to prove the lemma in the cases where the minimum $\phi$ of the external angles is $\theta_{13}$, $\theta_{14}$, $\theta_{23}$ or $\theta_{24}$, respectively.
When we apply this lemma, it is natural to distinguish the cases $$\begin{gathered}
\label{D:97}
\phi \lesssim \min(\theta_{12},\theta_{34}),
\\
\label{D:98}
\min(\theta_{12},\theta_{34}) \ll \phi \lesssim \max(\theta_{12},\theta_{34}),
\\
\label{D:99}
\max(\theta_{12},\theta_{34}) \ll \phi,\end{gathered}$$ which imply, respectively, that the first, second or third term in dominates.
In certain situations, the last two cases can be treated simultaneously, by virtue of the following simplified version of Lemma \[D:Lemma2\].
\[D:Lemma6\] In the cases and , $${| q_{1234} |}
\lesssim
\theta_{13}\theta_{24}.$$
Note that in cases and , $$\label{D:127}
\min(\theta_{12},\theta_{34}) \ll \phi \le \theta_{13}, \theta_{14}, \theta_{23}, \theta_{24}.$$ In case the dominant term in is $\phi^2$, hence the lemma follows from . To handle case , note that by symmetry we may assume that $\theta_{12} \le \theta_{34}$. Then we claim that implies $$\label{D:100}
\theta_{13} \sim \theta_{23},
\qquad
\theta_{14} \sim \theta_{24}.$$ Granting this, we next notice that since we are in case , and since $\theta_{12} \le \theta_{34}$, the dominant term in is $\phi\theta_{34}$, but $\theta_{34} \le \theta_{13} + \theta_{14} \lesssim \max(\theta_{13},\theta_{24})$ by , and $\phi \le \min(\theta_{13},\theta_{24})$ by , so the lemma holds.
It remains to prove . By the triangle inequality for distances on the unit sphere, $\theta_{13} \le \theta_{12} + \theta_{23}$, and since $\theta_{12} \ll \theta_{13}$, it follows that $\theta_{13} \lesssim \theta_{23}$. Repeating this argument with the index 13 replaced by $23$, $14$ and $24$, we get , so the claim is proved. This completes the proof of the lemma.
Our general strategy is now to apply the Cauchy-Schwarz inequality in various ways to reduce to bilinear $L^2$ estimates. The standard estimates suffice in the particularly easy case , which we dispose of straight away: The term $\theta_{12}\theta_{34}$ then dominates in , hence, with notation as in , $$\label{D:102}
\begin{aligned}
J_{{\boldsymbol{N}},{\boldsymbol{L}}}^{\mathbf\Sigma}
&\lesssim
{\Bigl\Vert T^{\pm_0}_{L_0,L_0'}
\mathcal F \, {\mathbf{P}}_{K^{\pm_0}_{N_0,L_0'}}
\mathfrak B_{\theta_{12}}(u_1,u_2) \Bigr\Vert}
{\Bigl\Vert {\mathbf{P}}_{K^{\pm_0}_{N_0,L_0}}
\mathfrak B_{\theta_{34}}(u_3,u_4) \Bigr\Vert}
\\
&\lesssim
{\Bigl\Vert {\mathbf{P}}_{K^{\pm_0}_{N_0,L_0'}}
\mathfrak B_{\theta_{12}}(u_1,u_2) \Bigr\Vert}
{\Bigl\Vert {\mathbf{P}}_{K^{\pm_0}_{N_0,L_0}}
\mathfrak B_{\theta_{34}}(u_3,u_4) \Bigr\Vert}
\\
&\lesssim
\left(N_0L_0'L_1L_2\right)^{1/2}\left(N_0L_0L_3L_4\right)^{1/2}
{\left\Vert u_1 \right\Vert} {\left\Vert u_2 \right\Vert} {\left\Vert u_3 \right\Vert} {\left\Vert u_4 \right\Vert},
\end{aligned}$$ where we applied the Cauchy-Schwarz inequality, Lemma \[D:Lemma3\] and Theorem \[D:Thm\].
This proves Theorem \[C:Thm2\] in the case , so henceforth we assume that or holds. Then the standard $L^2$ bilinear estimates do not suffice. In addition we need a number of modified bilinear estimates proved by the third author in [@Selberg:2008a], which we recall in the next section. In section \[F\] we fill in some details about bilinear interactions, and then in sections \[G\] and \[E\] we finally prove Theorem \[C:Thm2\].
Additional bilinear estimates {#J}
=============================
Here we recall a number of bilinear estimates proved in [@Selberg:2008a]. For more about the motivation behind these estimates, see [@Selberg:2008a].
Anisotropic bilinear estimate
-----------------------------
\[J:Thm3\] Let $\omega \in \mathbb S^2$, $I \subset {\mathbb{R}}$ a compact interval. In addition to the usual assumption , assume now $$\operatorname{supp}\widehat{u_1} \subset
\left\{ (\tau,\xi) \colon
\theta(\xi,\omega^\perp) \ge \alpha \right\}
\qquad \text{for some $\;0 < \alpha \ll 1$}.$$ Then $${\left\Vert {\mathbf{P}}_{\xi_0 \cdot \omega \in I}
(u_1 u_2) \right\Vert}
\lesssim \left( \frac{{| I |}{N_{\mathrm{min}}}^{12}L_1L_2 }{\alpha} \right)^{1/2}
{\left\Vert u_1 \right\Vert}
{\left\Vert u_2 \right\Vert}.$$ The same estimate holds for ${\left\Vert {\mathbf{P}}_{\xi_1 \cdot \omega \in I} u_1 \cdot u_2 \right\Vert}$ and ${\left\Vert u_1 \cdot {\mathbf{P}}_{\xi_2 \cdot \omega \in I} u_2 \right\Vert}$.
Here $\omega^\perp \subset {\mathbb{R}}^3$ is the orthogonal complement of $\omega$, and ${| I |}$ is the length of $I$.
Null form estimate with tube restricition
-----------------------------------------
Recall that $T_r(\omega) \subset {\mathbb{R}}^3$, for $r > 0$ and $\omega \in \mathbb S^2$, denotes a tube of radius comparable to $r$ around ${\mathbb{R}}\omega$. Recall also the definition of the null forms in and .
\[G:Thm\] Let $r > 0$ and $\omega \in \mathbb S^2$. Then with notation as in , $$\label{G:16}
{\bigl\Vert \mathfrak B_{\theta_{12}}({\mathbf{P}}_{{\mathbb{R}}\times T_r(\omega)} u_1,u_2) \bigr\Vert}
\lesssim \left( r^2 L_1 L_2 \right)^{1/2}
{\left\Vert u_1 \right\Vert}{\left\Vert u_2 \right\Vert}.$$ Moreover, the same estimate holds for $\mathfrak B_{\theta_{12}}'$.
The key point here is that we are able to exploit concentration of the Fourier supports near a null ray, which is not possible for a standard product like $u_1 \overline{u_2}$.
In [@Selberg:2008a], is proved under the hypothesis that ${| \xi_j |} \sim N_j$ on the support of $\widetilde{u_j}$, for $j=1,2$, instead of ${\langle \xi_j \rangle} \sim N_j$, as we have here. This only makes as difference if ${N_{\mathrm{min}}}^{12} \sim 1$, however, but in that case, recalling also the standing assumption $L_1,L_2 \ge 1$, follows from $${\bigl\Vert {\mathbf{P}}_{{\mathbb{R}}\times T_r(\omega)} u_1 \cdot \overline{u_2} \bigr\Vert}
\lesssim \left( r^2 {N_{\mathrm{min}}}^{12} {L_{\mathrm{min}}}^{12} \right)^{1/2}
{\left\Vert u_1 \right\Vert}{\left\Vert u_2 \right\Vert}.$$ This estimate is also proved in [@Selberg:2008a], but requires only ${| \xi_j |} \lesssim N_j$ for $j=1,2$.
In [@Selberg:2008a], is proved for $\mathfrak B_{\theta_{12}}'$, but it is easy to see that this implies the same estimate for $\mathfrak B_{\theta_{12}}$.
Concentration/nonconcentration null form estimate
-------------------------------------------------
In the following refinement of Theorem \[G:Thm\] we limit attention to interactions which are nearly null, by restricting the symbol in and to $\theta_{12} \ll 1$; we denote these modified null forms by $\mathfrak B_{\theta_{12} \ll 1}$ and $\mathfrak B'_{\theta_{12} \ll 1}$.
\[J:Thm1\] Let $r > 0$, $\omega \in \mathbb S^2$ and $I_0 \subset {\mathbb{R}}$ a compact interval. Assume that $N_1, N_2 \gg 1$ and that $$\label{J:0:2}
r \ll {N_{\mathrm{min}}}^{12}.$$ Then with notation as in , $${\left\Vert {\mathbf{P}}_{\xi_0 \cdot \omega \in I_0} \mathfrak B_{\theta_{12} \ll 1}({\mathbf{P}}_{{\mathbb{R}}\times T_r(\omega)} u_1,u_2) \right\Vert}
\lesssim \left( r^2 L_1 L_2 \right)^{1/2}
\left( \sup_{I_1} {\left\Vert {\mathbf{P}}_{\xi_1 \cdot \omega \in I_1} u_1 \right\Vert} \right)
{\left\Vert u_2 \right\Vert},$$ where the supremum is over all translates $I_1$ of $I_0$. The same holds for $\mathfrak B'_{\theta_{12} \ll 1}$.
Here the condition $N_1,N_2 \gg 1$ serves to ensure that the spatial Fourier supports, given by the conditions ${\langle \xi_j \rangle} \sim N_j$ for $j=1,2$, do not degenerate to balls.
Nonconcentration low output estimate
------------------------------------
The following result improves in certain situations.
\[J:Thm2\] Assume $N_0 \ll N_1 \sim N_2$, and define $r = (N_0{L_{\mathrm{max}}}^{012})^{1/2}$. Then with notation as in , $${\Bigl\Vert {\mathbf{P}}_{K^{\pm_0}_{N_0,L_0}}\!\!(u_1u_2) \Bigr\Vert}
\lesssim \left( N_1^2 {L_{\mathrm{min}}}^{012} {L_{\mathrm{med}}}^{012} \right)^{1/2} {\left\Vert u_1 \right\Vert}
\sup_{\omega \in \mathbb S^2} {\left\Vert {\mathbf{P}}_{{\mathbb{R}}\times T_{r}(\omega)}u_2 \right\Vert}.$$
In fact, we do not use Theorem \[J:Thm2\] as stated, but we use the main ideas from its proof, a key ingredient of which is the following partial orthogonality estimate for a family of thickened null hyperplanes corresponding to a set of well-separated directions on the unit sphere.
\[J:Lemma\] Suppose $N,d > 0$, $\omega_0 \in \mathbb S^2$ and $0 < \gamma < \gamma' < 1$. The estimate $$\label{J:40}
\sum_{\genfrac{}{}{0pt}{1}{\omega \in \Omega(\gamma)}{\theta(\omega,\omega_0) \le \gamma'}} \chi_{H_d(\omega)}(\tau,\xi)
\lesssim \frac{\gamma'}{\gamma} + \frac{d}{N\gamma^2}$$ holds for all $(\tau,\xi) \in {\mathbb{R}}^{1+3}$ with ${| \xi |} \sim N$.
Null form estimate with ball restriction
----------------------------------------
In the following analogue of Theorem \[G:Thm\], the tube is replaced by a ball. Then the symbol $\theta_{12}$ in and can be replaced by $\sqrt{\theta_{12}}$, defining $\mathfrak B_{\sqrt{\theta_{12}}}$ and $\mathfrak B'_{\sqrt{\theta_{12}}}$.
\[G:Thm2\] Let $r > 0$, and let $B \subset {\mathbb{R}}^3$ be a ball of radius $r$ and arbitrary center. Then with notation as in , $${\bigl\Vert {\mathbf{P}}_{{\mathbb{R}}\times B} \mathfrak B_{\sqrt{\theta_{12}}}(u_1,u_2) \bigr\Vert}
\lesssim \left( r^2 L_1 L_2 \right)^{1/2}
{\left\Vert u_1 \right\Vert}{\left\Vert u_2 \right\Vert}.$$ The same holds if ${\mathbf{P}}_{{\mathbb{R}}\times B}$ is placed in front of either $u_1$ or $u_2$, instead of outside the product. Moreover, the same estimates hold for $\mathfrak B'_{\sqrt{\theta_{12}}}$
Some properties of bilinear interactions {#F}
========================================
Here we fill in some details about the bilinear interaction $X_0=X_1-X_2$, where $X_j=(\tau_j,\xi_j)$. Given signs $(\pm_0,\pm_1,\pm_2)$, we define the hyperbolic weights ${\mathfrak h}_j$ as in , and we assume $\xi_j \neq 0$ for $j=0,1,2$, so that the unit vectors $e_j$ and the angles $\theta_{jk}$ are well-defined, as in and .
In Lemma \[D:Lemma1\] we related the angle $\theta_{12}$ to the size of the hyperbolic weights ${\mathfrak h}_j$ and the elliptic weights ${| \xi_j |}$. The sign $\pm_0$ was arbitrary, but by keeping track of the sign $\pm_0$ we can get additional information. In fact, since $\tau_0=\tau_1-\tau_2$, $$\label{F:0}
{\mathfrak h}_0 - {\mathfrak h}_1 + {\mathfrak h}_2
= \pm_0{| \xi_0 |}-\pm_1{| \xi_1 |} \pm_2{| \xi_2 |},$$ so by , and the fact that $\xi_0=\xi_1-\xi_2$, we get the information in Table \[F:Table\].
-------------------------------------------------------------------------------------------------------
$(\pm_0,\pm_1,\pm_2) =$ ${| {\mathfrak h}_0 |} + {| {\mathfrak h}_1 |} + {| {\mathfrak h}_2 |} \ge$
------------------------- -----------------------------------------------------------------------------
$(+,+,+)$ or $(-,-,-)$ ${| \xi_0 |} + {| \xi_2 |} - {| \xi_1 |}
\sim \min({| \xi_0 |},{| \xi_2 |}) \theta_{02}^2$
$(-,+,+)$ or $(+,-,-)$ ${| \xi_0 |} + {| \xi_1 |} - {| \xi_2 |}
\sim \min({| \xi_0 |},{| \xi_1 |}) \theta_{01}^2$
$(+,+,-)$ or $(-,-,+)$ ${| \xi_0 |} + {| \xi_1 |} + {| \xi_2 |}$
$(-,+,-)$ or $(+,-,+)$ ${| \xi_1 |} + {| \xi_2 |} - {| \xi_0 |}
\sim \min({| \xi_1 |},{| \xi_2 |}) \theta_{12}^2$
-------------------------------------------------------------------------------------------------------
: Estimates for the bilinear interaction.[]{data-label="F:Table"}
From this table we see that for certain bilinear interactions, $$\label{F:2}
\max\left({| {\mathfrak h}_0 |},{| {\mathfrak h}_1 |},{| {\mathfrak h}_2 |}\right)
\gtrsim {| \xi_0 |},$$ which is good because it excludes the null interaction. In general, holds if $\pm_0$ is *different from* the following sign:
\[F:Def\] We define the sign, depending on $(\pm_1,\pm_2)$ and $({| \xi_1 |},{| \xi_2 |})$, $$\pm_{12}
=
\begin{cases}
+ \qquad &\text{if $(\pm_1,\pm_2) = (+,+)$ and ${| \xi_1 |} > {| \xi_2 |}$},
\\
- \qquad &\text{if $(\pm_1,\pm_2) = (+,+)$ and ${| \xi_1 |} \le {| \xi_2 |}$},
\\
+ \qquad &\text{if $(\pm_1,\pm_2) = (+,-)$}.
\end{cases}$$ The definitions in the remaining cases $(\pm_1,\pm_2) = (-,-), (-,+)$ are then obtained by reversing all three signs $\pm_{12},\pm_1,\pm_2$ above.
As remarked already we then have, by the above table:
\[F:Lemma2\] If $\pm_0 \neq \pm_{12}$, then holds.
Now consider the case of equal signs. The possible interactions are illustrated in Figures \[F:Fig1\] and \[F:Fig2\]. From the sine rule we obtain:
\[F:Lemma1\] If $\pm_0 = \pm_{12}$, then $$\label{F:4}
\min\left(\theta_{01},\theta_{02}\right)
\sim
\frac{\min\left({| \xi_1 |},{| \xi_2 |}\right)}{{| \xi_0 |}} \sin \theta_{12}.$$ Moreover, if $\pm_0 = \pm_{12}$ and $\pm_1 \neq \pm_2$, then $$\label{F:5}
\max\left(\theta_{01},\theta_{02}\right)
\sim
\theta_{12}.$$
It suffices to consider the cases $(\pm_1,\pm_2) = (+,+), (+,-)$. In both cases, the law of sines yields (see Figures \[F:Fig1\] and \[F:Fig2\]) $$\begin{aligned}
\label{F:6}
{| \xi_2 |}\sin\theta_{12} &= {| \xi_0 |}\sin\theta_{01},
\\
\label{F:8}
{| \xi_1 |}\sin\theta_{12} &= {| \xi_0 |}\sin\theta_{02}.\end{aligned}$$ Now we claim that $$\label{F:10}
\min\left(\theta_{01},\theta_{02}\right)
= \begin{cases}
\theta_{01} \quad &\text{if ${| \xi_1 |} > {| \xi_2 |}$},
\\
\theta_{02} \quad &\text{if ${| \xi_1 |} \le {| \xi_2 |}$},
\end{cases}$$ and $$\label{F:12}
0 \le \min\left(\theta_{01},\theta_{02}\right) \le \frac{\pi}{2}.$$ Clearly, – imply , so it remains to prove the claim.
First, assume that $(\pm_1,\pm_2) = (+,+)$. If ${| \xi_1 |} > {| \xi_2 |}$, then from Figure \[F:Fig1\](a) we see that $\theta_{02} = \theta_{01} + \theta_{12}$ and $\theta_{01} \in [0,\pi/2]$, whereas if ${| \xi_1 |} \le {| \xi_2 |}$, then as in Figure \[F:Fig1\](b) we have $\theta_{01} = \theta_{02} + \theta_{12}$ and $\theta_{02} \in [0,\pi/2]$. Thus, and hold.
Second, consider $(\pm_1,\pm_2) = (+,-)$. From Figure \[F:Fig2\] we see that $\theta_{12} = \theta_{01} + \theta_{02}$, so the minimum of $\theta_{01}$ and $\theta_{02}$ must belong to $[0,\pi/2]$, proving . If ${| \xi_1 |} > {| \xi_2 |}$, the minimum is $\theta_{01}$, whereas if ${| \xi_1 |} \le {| \xi_2 |}$, the minimum is $\theta_{02}$, so holds. Since $\theta_{12} = \theta_{01} + \theta_{02}$, is immediate.
![$(\pm_1,\pm_2) = (+,+)$ and $\pm_0=\pm_{12}$.[]{data-label="F:Fig1"}](figures-1.pdf)
![$(\pm_1,\pm_2) = (+,-)$ and $\pm_0=\pm_{12}=+$. Here $\theta_{12} = \theta_{01} + \theta_{02}$.[]{data-label="F:Fig2"}](figures-2.pdf)
The following is reminiscent of in Lemma \[D:Lemma1\].
\[F:Lemma4\] For all signs, $$\label{F:14}
\max\left( {| {\mathfrak h}_0 |}, {| {\mathfrak h}_1 |}, {| {\mathfrak h}_2 |} \right)
\gtrsim {| \xi_0 |} \min\left(\theta_{01},\theta_{02}\right)^2.$$
If $\pm_0\neq\pm_{12}$, holds by Lemma \[F:Lemma2\], so we may assume $\pm_0=\pm_{12}$, hence is valid. We split into the cases $\theta_{12} \ll 1$ and $\theta_{12} \sim 1$.
If $\theta_{12} \ll 1$, Lemma \[D:Lemma1\] implies , and using also we get $${| \xi_0 |} \min\left(\theta_{01},\theta_{02}\right)^2
\sim
\frac{\min\left({| \xi_1 |},{| \xi_2 |}\right)^2}{{| \xi_0 |}} \theta_{12}^2
\lesssim
\frac{\min\left({| \xi_1 |},{| \xi_2 |}\right)\max\left( {| {\mathfrak h}_0 |}, {| {\mathfrak h}_1 |}, {| {\mathfrak h}_2 |} \right)}{\max\left({| \xi_1 |},{| \xi_2 |}\right)},$$ so holds.
Next, assume $\theta_{12} \sim 1$. Then by in Lemma \[D:Lemma1\], $$\min\left({| \xi_1 |},{| \xi_2 |}\right)
\lesssim\max\left( {| {\mathfrak h}_0 |}, {| {\mathfrak h}_1 |}, {| {\mathfrak h}_2 |} \right),$$ and combining this with we get $${| \xi_0 |} \min\left(\theta_{01},\theta_{02}\right)^2
\lesssim
{| \xi_0 |} \min\left(\theta_{01},\theta_{02}\right)
\lesssim
\min\left({| \xi_1 |},{| \xi_2 |}\right)
\lesssim
\max\left( {| {\mathfrak h}_0 |}, {| {\mathfrak h}_1 |}, {| {\mathfrak h}_2 |} \right),$$ so again holds.
Based on the previous lemmas we now prove the following.
\[F:Lemma3\] If $\pm_0 = \pm_{12}$ and $\pm_1=\pm_2$, then $$\label{F:20}
\frac{{| \xi_1 |}{| \xi_2 |}\theta_{12}^2}{{| \xi_0 |}} \sim \min\left( {| \xi_0 |}, {| \xi_1 |}, {| \xi_2 |}\right) \max(\theta_{01},\theta_{02})^2
\lesssim
\max\left( {| {\mathfrak h}_0 |}, {| {\mathfrak h}_1 |}, {| {\mathfrak h}_2 |} \right),$$ whereas if $\pm_0 = \pm_{12}$ and $\pm_1 \neq \pm_2$, then $$\label{F:22}
\max\left(\theta_{01},\theta_{02}\right)
\sim
\theta_{12}.$$
The last statement was proved in Lemma \[F:Lemma1\], so we only prove . It suffices to consider $(\pm_1,\pm_2) = (+,+)$. If ${| \xi_1 |} > {| \xi_2 |}$, then $\pm_0 = \pm_{12} = +$, and by Table \[F:Table\], $$\max\left( {| {\mathfrak h}_0 |}, {| {\mathfrak h}_1 |}, {| {\mathfrak h}_2 |} \right)
\gtrsim
{| \xi_0 |} + {| \xi_2 |} - {| \xi_1 |}
\sim \min\left( {| \xi_0 |}, {| \xi_2 |}\right) \theta_{02}^2,$$ but also $$\max\left( {| {\mathfrak h}_0 |}, {| {\mathfrak h}_1 |}, {| {\mathfrak h}_2 |} \right)
\gtrsim
{| \xi_0 |} + {| \xi_2 |} - {| \xi_1 |}
= {| \xi_0 |} - {\bigl\vert {| \xi_1 |} - {| \xi_2 |} \bigr\vert}
\sim \frac{{| \xi_1 |}{| \xi_2 |}\theta_{12}^2}{{| \xi_0 |}},$$ where we used . Combining these estimates, and noting that $\theta_{01} \le \theta_{02}$ by , we get the desired estimate. The case ${| \xi_1 |} \le {| \xi_2 |}$ is treated similarly.
Proof of the dyadic quadrilinear estimate, Part I {#G}
=================================================
Here we prove Theorem \[C:Thm2\] under the assumption $$\label{G:2}
L_0 \sim L_0'.$$ By symmetry, we may assume $$\label{G:6}
\theta_{12} \le \theta_{34},$$ and $$\label{G:6:2}
L_1 \le L_2,
\qquad
L_3 \le L_4,$$ We distinguish the cases
\[G:4\] $$\begin{aligned}
{2}
\label{G:4a}
&L_2 \le L_0',&&
\\
\label{G:4b}
&&
&L_4 \le L_0,
\\
\label{G:4c}
&L_2 > L_0',&
\qquad
&L_4 > L_0,\end{aligned}$$
each of which may be split further into the subcases and (recall that has been completely dealt with). In each case it is of course understood that the region of integration in is restricted accordingly.
In view of , the cases , simplify to
\[G:8\] $$\begin{aligned}
{2}
\label{G:8a}
&\theta_{12} \ll \phi \lesssim \theta_{34},&
\qquad &\bigl( \implies {| q_{1234} |} \lesssim \phi\theta_{34} \bigr)
\\
\label{G:8b}
&\theta_{12} \le \theta_{34} \ll \phi,&
\qquad &\bigl( \implies {| q_{1234} |} \lesssim \phi^2 \bigr),\end{aligned}$$
where the symbol estimates on the right hold by Lemma \[D:Lemma2\]. By Lemma \[D:Lemma1\], $$\label{D:134:10}
\theta_{12} \lesssim \gamma \equiv \min \left( \gamma^*, \biggl(\frac{N_0{L_{\mathrm{max}}}^{0'12}}{N_1N_2}\biggr)^{1/2} \right),
\qquad \text{for some $0 < \gamma^* \ll 1$}.$$ In fact, here we can choose any $0 < \gamma^* \ll 1$ that we want, by adjusting the implicit constant in . By Lemma \[D:Lemma1\] we also have $$\label{E:4}
\theta_{34} \lesssim \gamma' \equiv \left( \frac{{L_{\mathrm{max}}}^{034}}{{N_{\mathrm{min}}}^{34}} \right)^{1/2}.$$
Observe that, with notation as in and , $$\label{E:2}
\phi \le \min(\theta_{01},\theta_{02}) + \min(\theta_{03},\theta_{04}),$$ since $\theta_{jk} \le \theta_{0j} + \theta_{0k}$. By Lemma \[F:Lemma4\], $$\label{E:2:2}
\min(\theta_{01},\theta_{02})
\lesssim \biggl(\frac{{L_{\mathrm{max}}}^{0'12}}{N_0} \biggr)^{1/2},
\qquad
\min(\theta_{03},\theta_{04})
\lesssim \biggl(\frac{{L_{\mathrm{max}}}^{034}}{N_0} \biggr)^{1/2}.$$
To simplify, we introduce the shorthand $$\label{E:3:1}
u_{0'12}
= {\mathbf{P}}_{K^{\pm_0}_{N_0,L_0'}}\left( u_1 \overline{u_2} \right),
\qquad
u_{043}
= {\mathbf{P}}_{K^{\pm_0}_{N_0,L_0}}\left( u_4 \overline{u_3} \right),$$ with notation as in . We define $\pm_{12}$ and $\pm_{43}$ as in Definition \[F:Def\], recalling that $\xi_0 = \xi_1-\xi_2 = \xi_4-\xi_3$ by . Note the following important relations between the angles $\theta_{12}, \theta_{01}, \theta_{02}$ in the low and high output interactions (recall ): $$\begin{aligned}
{2}
\label{E:3:2}
&\pm_0=\pm_{12}, \; \theta_{12} \ll 1, \; N_0 \ll N_1 \sim N_2&
\quad \implies \quad
&\theta_{01} \sim \theta_{02} \sim \frac{N_1}{N_0} \theta_{12},
\\
\label{E:3:3}
&\pm_0=\pm_{12}, \; \theta_{12} \ll 1, \; N_1 \lesssim N_0 \sim N_2&
\quad \implies \quad
&\theta_{12} \sim \theta_{01} \sim \frac{N_0}{N_1} \theta_{02}.\end{aligned}$$ This follows by Lemma \[F:Lemma1\], in the proof of that lemma, and Lemma \[F:Lemma3\]. Note also that can only happen if $\pm_1=\pm_2$. Of course, applies symmetrically when $N_2 \lesssim N_0 \sim N_1$. Analogous estimates apply to the index 043.
Case
-----
Then we can treat the cases and simultaneously by using Lemma \[D:Lemma6\] and pairing up $u_1$ with $u_3$, and $u_2$ with $u_4$. After an angular decomposition we then apply the null form estimate with tube restriction.
Discard the characteristic functions of $K^{\pm_0}_{N_0,L_0}, K^{\pm_0}_{N_0,L_0'}$ in (this is reasonable in view of and ), and estimate ${| q_{1234} |}$ by Lemma \[D:Lemma6\]. To achieve the desired pairing of the $u$’s, we change the variables $(\tau_0',\tau_0,\xi_0)$ in to $$\tilde\tau_0' = \tau_1+\tau_3,
\qquad
\tilde\tau_0 = \tau_2+\tau_4,
\qquad
\tilde\xi_0 = \xi_1+\xi_3 = \xi_2+\xi_4.$$ By and , $\tilde\tau_0' - \tilde\tau_0 = \tau_0' - \tau_0$, so the symbol is invariant under the change of variables: $$\label{G:5}
a_{L_0,L_0'}^{\pm_0}(\tau_0,\tau_0',\xi_0) =
a_{L_0,L_0'}^{\pm_0}(\tilde\tau_0,\tilde\tau_0',\tilde\xi_0).$$ Note that this relies on the assumption $L_0 \sim L_0'$. We conclude that $$\label{D:134}
\begin{aligned}
J_{{\boldsymbol{N}},{\boldsymbol{L}}}^{\mathbf\Sigma}
&\lesssim
\int
T_{L_0,L_0'} \mathcal F\mathfrak B'_{\theta_{13}}(u_1,u_3)(\tilde X_0)
\cdot \mathcal F \mathfrak B'_{\theta_{24}}(u_2,u_4)(\tilde X_0)
\, d\tilde X_0
\\
&\lesssim
{\left\Vert \mathfrak B'_{\theta_{13}}(u_1,u_3) \right\Vert}
{\left\Vert \mathfrak B'_{\theta_{24}}(u_2,u_4) \right\Vert},
\end{aligned}$$ where we used Lemma \[D:Lemma3\].
The question is now how to estimate the last two factors in . Observe that there is no point in applying Lemma \[D:Lemma1\] with respect to the angles $\theta_{13}, \theta_{24}$, since we know nothing about the associated ${\mathfrak h}_0$-weights. In other words, Theorem \[D:Thm\] is not useful here. Instead, we use the upper bound for $\theta_{12}$. Although this angle does not explicitly appear in , it can still make its influence felt if we use Lemma \[B:Lemma4\] to angularly decompose $(\pm_1\xi_1,\pm_2\xi_2)$ before we use . The angular decomposition restricts the spatial Fourier supports to tubes, so we can apply Theorem \[G:Thm\]. Let us turn to the details.
Making use of the restriction , we apply Lemma \[B:Lemma5\] to the pair $(\pm_1\xi_1,\pm_2\xi_2)$, and then we use , thus obtaining $$\label{D:134:2}
J_{{\boldsymbol{N}},{\boldsymbol{L}}}^{\mathbf\Sigma}
\lesssim
\sum_{\omega_1,\omega_2}
{\left\Vert \mathfrak B'_{\theta_{13}}(u_1^{\gamma,\omega_1},u_3) \right\Vert}
{\left\Vert \mathfrak B'_{\theta_{24}}(u_2^{\gamma,\omega_2},u_4) \right\Vert},$$ where the sum is over $\omega_1,\omega_2 \in \Omega(\gamma)$ satisfying $\theta(\omega_1,\omega_2) \lesssim \gamma$, and we use the notation (indexed by $1,2$, except for $\gamma$). Since the spatial frequency $\xi_j$ of $u_j^{\gamma,\omega_j}$ is restricted to a tube of radius comparable to $N_j\gamma$ about ${\mathbb{R}}\omega_j$, we can apply Theorem \[G:Thm\] to . The result is $$\label{D:134:3}
\begin{aligned}
J_{{\boldsymbol{N}},{\boldsymbol{L}}}^{\mathbf\Sigma}
&\lesssim
N_1 N_2 \gamma^2
\left( L_1 L_2 L_3 L_4 \right)^{1/2}
\sum_{\omega_1,\omega_2}{\left\Vert u_1^{\gamma,\omega_1} \right\Vert}{\left\Vert u_2^{\gamma,\omega_2} \right\Vert}{\left\Vert u_3 \right\Vert}{\left\Vert u_4 \right\Vert}
\\
&\lesssim
N_0 L_0'
\left( L_1 L_2 L_3 L_4 \right)^{1/2}
{\left\Vert u_1 \right\Vert}{\left\Vert u_2 \right\Vert}{\left\Vert u_3 \right\Vert}{\left\Vert u_4 \right\Vert},
\end{aligned}$$ where we summed $\omega_1,\omega_2$ as in , and used the definition of $\gamma$, taking into account the assumption . In view of , this proves .
This concludes case .
Case
-----
If $\theta_{34} \ll 1$, then we have the analogue of , so by symmetry the argument used for case above applies, with the roles of the indices 12 and 34 reversed.
It then remains to consider $\theta_{34} \sim 1$. Then ${N_{\mathrm{min}}}^{34} \lesssim L_0$, by Lemma \[D:Lemma1\]. Moreover, we may assume $L_2 > L_0'$, since the case $L_2 \le L_0'$ was completely dealt with above. Now trivially estimate ${| q_{1234} |} \lesssim 1$. Then with notation as in , $$\begin{aligned}
J_{{\boldsymbol{N}},{\boldsymbol{L}}}^{\mathbf\Sigma}
&\lesssim
{\left\Vert u_{0'12} \right\Vert}
{\left\Vert u_{043} \right\Vert}
\\
&\lesssim
\left(N_0^2L_0'L_1\right)^{1/2}\left(({N_{\mathrm{min}}}^{34})^2L_3L_4\right)^{1/2}
{\left\Vert u_1 \right\Vert} {\left\Vert u_2 \right\Vert} {\left\Vert u_3 \right\Vert} {\left\Vert u_4 \right\Vert}
\\
&\lesssim
\left(N_0^2L_1L_2\right)^{1/2}\left(L_0^2L_3L_4\right)^{1/2}
{\left\Vert u_1 \right\Vert} {\left\Vert u_2 \right\Vert} {\left\Vert u_3 \right\Vert} {\left\Vert u_4 \right\Vert},\end{aligned}$$ where the first inequality holds by the Cauchy-Schwarz inequality and Lemma \[D:Lemma3\], the second by Theorem \[M:Thm\], and the third by the assumption $L_2 > L_0'$ and the fact that ${N_{\mathrm{min}}}^{34} \lesssim L_0$. In view of , this implies .
This concludes the case .
Case
-----
So far we were able to treat and simultaneously, but from now on we need to separate the two, and we further divide into subcases depending on which term dominates in the right hand side of :
\[G:10\] $$\begin{aligned}
{2}
\label{G:10a}
&\theta_{12} \ll \phi \lesssim \theta_{34},&
\qquad
&\min(\theta_{01},\theta_{02}) \ge \min(\theta_{03},\theta_{04}),
\\
\label{G:10b}
&\theta_{12} \ll \phi \lesssim \theta_{34},&
\qquad
&\min(\theta_{01},\theta_{02}) < \min(\theta_{03},\theta_{04}),
\\
\label{G:10c}
&\theta_{12} \le \theta_{34} \ll \phi,&
\qquad
&\min(\theta_{01},\theta_{02}) < \min(\theta_{03},\theta_{04}),
\\
\label{G:10d}
&\theta_{12} \le \theta_{34} \ll \phi,&
\qquad
&\min(\theta_{01},\theta_{02}) \ge \min(\theta_{03},\theta_{04}).\end{aligned}$$
Subcase is by far the most difficult, and will be split further into subcases.
Case , subcase {#G:12:0}
---------------
Then by and –, $$\label{G:12}
{| q_{1234} |}
\lesssim \phi\theta_{34}
\lesssim
\biggl( \frac{L_2}{N_0} \biggr)^{1/2}
\left( \frac{L_4}{{N_{\mathrm{min}}}^{34}} \right)^{1/2},$$ hence $$\label{G:12:2}
\begin{aligned}
J_{{\boldsymbol{N}},{\boldsymbol{L}}}^{\mathbf\Sigma}
&\lesssim
\biggl( \frac{L_2}{N_0} \biggr)^{1/2}
\left( \frac{L_4}{{N_{\mathrm{min}}}^{34}} \right)^{1/2}
{\left\Vert u_{0'12} \right\Vert} {\bigl\Vert u_{043} \bigr\Vert}
\\
&\lesssim
\biggl( \frac{L_2}{N_0}
\cdot \frac{L_4}{{N_{\mathrm{min}}}^{34}}
\cdot N_0^2 L_0' L_1
\cdot N_0 {N_{\mathrm{min}}}^{034} L_0 L_3 \biggr)^{1/2}
\prod_{j=1}^4 {\left\Vert u_j \right\Vert},
\end{aligned}$$ where we used the Cauchy-Schwarz inequality, Lemma \[D:Lemma3\] and Theorem \[M:Thm\].
Case , subcase {#G:14:0}
---------------
Then by and , $$\label{G:14:2}
{| q_{1234} |} \lesssim \phi\theta_{34}
\lesssim \min(\theta_{03},\theta_{04}) \left( \frac{L_4}{{N_{\mathrm{min}}}^{34}} \right)^{1/2},$$ so by and Lemma \[B:Lemma5\], applied to the pair $(\pm_1\xi_1,\pm_2\xi_2)$, $$\label{G:14:4}
J_{{\boldsymbol{N}},{\boldsymbol{L}}}^{\mathbf\Sigma}
\lesssim
\sum_{\omega_1,\omega_2} \left( \frac{L_4}{{N_{\mathrm{min}}}^{34}} \right)^{1/2}
{\left\Vert \mathfrak B_{\theta_{03}}' \left( {\mathbf{P}}_{K^{\pm_0}_{N_0,L_0}} \mathcal F^{-1} T_{L_0,L_0'}^{\pm_0} \mathcal F u_{0'12}^{\gamma,\omega_1,\omega_2}, u_3 \right) \right\Vert} {\left\Vert u_4 \right\Vert},$$ where $$\label{G:14:6}
u^{\gamma,\omega_1,\omega_2}_{0'12}
=
{\mathbf{P}}_{K^{\pm_0}_{N_0,L_0'}}
\left( u_1^{\gamma,\omega_1}
\overline{ u_2^{\gamma,\omega_2} } \right)$$ and the sum is over $\omega_1,\omega_2 \in \Omega(\gamma)$ with $\theta(\omega_1,\omega_2) \lesssim \gamma$. Since the spatial Fourier support of $u_j^{\gamma,\omega_j}$ is contained in a tube of radius comparable to $N_j\gamma$ about ${\mathbb{R}}\omega_j$, it follows that the spatial Fourier support of $u^{\gamma,\omega_1,\omega_2}_{0'12}$ is contained in a tube of radius comparable to ${N_{\mathrm{max}}}^{12}\gamma$ around ${\mathbb{R}}\omega_1$. Therefore, by Theorem \[G:Thm\], Lemma \[D:Lemma3\] and Theorem \[M:Thm\], $$\label{G:14:8}
\begin{aligned}
J_{{\boldsymbol{N}},{\boldsymbol{L}}}^{\mathbf\Sigma}
&\lesssim
\sum_{\omega_1,\omega_2} \left( \frac{L_4}{{N_{\mathrm{min}}}^{34}} \right)^{1/2}
{N_{\mathrm{max}}}^{12}\gamma \left(L_0L_3\right)^{1/2}
{\left\Vert u_{0'12}^{\gamma,\omega_1,\omega_2} \right\Vert}
{\left\Vert u_3 \right\Vert} {\left\Vert u_4 \right\Vert}
\\
&\lesssim
\left( \frac{L_4}{{N_{\mathrm{min}}}^{34}} \right)^{1/2}
{N_{\mathrm{max}}}^{12} \left( \frac{N_0L_2}{N_1N_2} \right)^{1/2} \left(L_0L_3
\cdot N_0 {N_{\mathrm{min}}}^{012} L_0'L_1 \right)^{1/2}
\prod_{j=1}^4 {\left\Vert u_j \right\Vert}
\\
&\lesssim
\left( \frac{N_0}{{N_{\mathrm{min}}}^{34}} N_0^2 L_0L_0'L_1L_2L_3L_4 \right)^{1/2}
\prod_{j=1}^4 {\left\Vert u_j \right\Vert}.
\end{aligned}$$ Here we summed $\omega_1,\omega_2$ as in , in the second step we used the definition of $\gamma$ from , recalling the assumption , and in the last step we used .
From we get the desired estimate if $N_0 \lesssim {N_{\mathrm{min}}}^{34}$, but also whenever we are able to gain an extra factor $({N_{\mathrm{min}}}^{34}/N_0)^{1/2}$. In particular, this happens if $\pm_0 \neq \pm_{43}$, since then $N_0 \lesssim {L_{\mathrm{max}}}^{034}$ by Lemma \[F:Lemma2\], so instead of we can use the estimate $\theta_{34} \lesssim 1 \lesssim ({L_{\mathrm{max}}}^{034}/N_0)^{1/2}$, thereby gaining the desired factor.
Thus, we may assume $\pm_0 = \pm_{43}$. Moreover, we can assume $\pm_0=\pm_{12}$, since if this fails to hold, then Lemma \[F:Lemma2\] implies $N_0 \lesssim L_2$, hence the argument in section \[G:12:0\] applies.
It remains to consider $N_3 \ll N_0 \sim N_4$ and $N_4 \ll N_0 \sim N_3$, but the case $N_3 \ll N_0 \sim N_4$ is easy: By and , $$\label{G:14:10}
N_3 \ll N_0 \sim N_4 \implies
\theta_{04} \lesssim \frac{N_3}{N_0} \theta_{34}, \quad \theta_{03} \sim \theta_{34},
\quad
{| q_{1234} |} \lesssim \frac{N_3}{N_0} \theta_{03} \theta_{34},$$ hence we gain a factor $N_3/N_0$ in , which is more than enough.
That leaves the interaction $N_4 \ll N_0 \sim N_3$, which is much harder; we split it further into $N_0 \lesssim N_2$ and $N_2 \ll N_0$, treated in the next two sections. Recall also that $\pm_0=\pm_{12}=\pm_{34}$ by the above reductions, and that we are assuming , by symmetry.
Case , subcase , $\pm_0=\pm_{12}=\pm_{43}$, $N_4 \ll N_0 \sim N_3$, $N_0 \lesssim N_2$ {#G:20:0}
--------------------------------------------------------------------------------------
We can insert ${\mathbf{P}}_{{| \xi_4 |} \lesssim N_4}$ in front of $\mathfrak B_{\theta_{03}}'$ in , and instead of Theorem \[G:Thm\] we then apply Theorem \[J:Thm1\]. Let us check that the hypotheses of the latter are satisfied. We have $N_4 \ll N_0 \sim N_3$, hence $N_0,N_3 \gg 1$ and $\theta_{03} \ll 1$ (by the analogue of ), so $\mathfrak B_{\theta_{03}}$ can be replaced by $\mathfrak B_{\theta_{03} \ll 1}$ in . The hypothesis translates to, in the present situation, $$\label{G:20:2}
{N_{\mathrm{max}}}^{12} \gamma \ll N_0,$$ where $\gamma$ is given by .
But if fails, then $N_0 \ll N_1 \sim N_2$, so by we see that the failure of is equivalent to $N_0 \lesssim L_2$, hence the argument in section \[G:12:0\] applies.
Thus, we can assume , hence Theorem \[J:Thm1\] applies, so in we can replace ${\left\Vert u_{0'12}^{\gamma,\omega_1,\omega_2} \right\Vert}$ by $$\sup_{I} {\left\Vert {\mathbf{P}}_{\xi_0 \cdot \omega_1 \in I} u_{0'12}^{\gamma,\omega_1,\omega_2} \right\Vert},$$ where the supremum is over all intervals $I \subset {\mathbb{R}}$ with length ${| I |} = N_4$. But since $\gamma \ll 1$, Theorem \[J:Thm3\] implies, via duality, $$\label{G:20:6}
\sup_I {\left\Vert {\mathbf{P}}_{\xi_0 \cdot \omega_1 \in I}
u_{0'12}^{\gamma,\omega_1,\omega_2} \right\Vert}
\lesssim
\bigl(N_4 {N_{\mathrm{min}}}^{01} L_0'L_1 \bigr)^{1/2}
{\left\Vert u_1^{\gamma,\omega_1} \right\Vert} {\left\Vert u_2^{\gamma,\omega_2} \right\Vert}.$$ Thus, in the second line of , the combination $N_0{N_{\mathrm{min}}}^{012}L_0'L_1$ can be replaced by $N_4{N_{\mathrm{min}}}^{01}L_0'L_1$, and since we are assuming $N_0 \lesssim N_2$, this means that we gain the desired factor $(N_4/N_0)^{1/2}$.
Case , subcase , $\pm_0=\pm_{12}=\pm_{43}$, $N_4 \ll N_0 \sim N_3$, $N_2 \ll N_0$ {#G:20:7}
---------------------------------------------------------------------------------
To make the argument from the previous section work, we need to somehow gain a factor $(N_2/N_0)^{1/2}$. We shall apply an argument which essentially is the same as the one introduced in [@Selberg:2008a] to prove the result stated here as Theorem \[J:Thm2\].
Let us first dispose of the easy case $N_2 \sim 1$. Then we simply estimate $$\label{G:20:8:2}
J_{{\boldsymbol{N}},{\boldsymbol{L}}}^{\mathbf\Sigma}
\lesssim
{\left\Vert u_{0'12} \right\Vert} {\bigl\Vert u_{043} \bigr\Vert}
\lesssim
\left( N_2^2 L_1 L_2
\cdot N_0^2 L_0 L_3 \right)^{1/2}
\prod_{j=1}^4 {\left\Vert u_j \right\Vert},$$ where we used Lemma \[D:Lemma3\] and Theorem \[M:Thm\].
For the rest of this section we therefore assume $1 \ll N_2 \ll N_0 \sim N_1$. This ensures that the region described by ${\langle \xi_2 \rangle} \sim N_2$ does not degenerate to a ball.
By and , $$\label{G:20:8}
\theta_{12} \sim \theta_{02} \sim \frac{N_0}{N_2} \theta_{01},
\qquad \text{hence}
\qquad
\theta_{01} \lesssim \alpha \equiv \frac{N_2}{N_0} \gamma.$$ Now modify by applying Lemma \[B:Lemma5\] again, this time to $(\pm_0\xi_0,\pm_1\xi_1)$: $$\begin{gathered}
\label{G:20:10}
J_{{\boldsymbol{N}},{\boldsymbol{L}}}^{\mathbf\Sigma}
\lesssim
\sum_{\omega_1,\omega_2} \sum_{\omega_0',\omega_1'} \left( \frac{L_4}{N_4} \right)^{1/2}
\\
\times
{\left\Vert {\mathbf{P}}_{{| \xi_4 |} \lesssim N_4}
\mathfrak B_{\theta_{03} \ll 1}' \left( {\mathbf{P}}_{K^{\pm_0}_{N_0,L_0}} \mathcal F^{-1} T_{L_0,L_0'}^{\pm_0} \mathcal F u_{0'12}^{\gamma,\omega_1,\omega_2;\alpha,\omega_0',\omega_1'}, u_3 \right) \right\Vert} {\left\Vert u_4 \right\Vert},\end{gathered}$$ where the second sum is over $\omega_0',\omega_1' \in \Omega(\alpha)$ satisfying $\theta(\omega_0',\omega_1') \lesssim \alpha$, and $$\begin{aligned}
\label{G:20:12}
u_{0'12}^{\gamma,\omega_1,\omega_2;\alpha,\omega_0',\omega_1'}
&= {\mathbf{P}}_{\pm_0\xi_0 \in \Gamma_{\alpha}(\omega_0')}
{\mathbf{P}}_{K^{\pm_0}_{N_0,L_0'}} \left( u_1^{\gamma,\omega_1;\alpha,\omega_1'}
u_2^{\gamma,\omega_2} \right)
\\
\label{G:20:14}
u_1^{\gamma,\omega_1;\alpha,\omega_1'}
&=
{\mathbf{P}}_{\pm_1\xi_1 \in \Gamma_{\alpha}(\omega_1')} u_1^{\gamma,\omega_1}.\end{aligned}$$ Observe that the spatial Fourier support of is contained in a tube of radius comparable to $N_0\alpha \sim N_2\gamma$ around ${\mathbb{R}}\omega_0'$; this tube is much thinner than the one for , which is of radius comparable to $N_1\gamma$ around ${\mathbb{R}}\omega_1$, hence we gain a factor $N_2/N_0$ when we apply Theorem \[J:Thm1\], compared to our estimates in section \[G:20:0\]. On the other hand, we now have the additional sum over $\omega_0',\omega_1'$. To come out on top, we have to make sure that this sum does not cost us more than a factor $(N_0/N_2)^{1/2}$; this requires some orthogonality, which is supplied by the following crucial information, to be fed into Lemma \[J:Lemma\].
As in , let $(X_0',X_1,X_2)$ denote the bilinear interaction for , so that $X_0'=X_1-X_2$. By , $$X_0' \in H_{\max(L_0',N_0\alpha^2)}(\omega_1'),
\qquad
X_1 \in H_{\max(L_1,N_0\alpha^2)}(\omega_1'),$$ where the latter relies on the assumption $\theta(\omega_0',\omega_1') \lesssim \alpha$. Therefore, $$\label{G:20:16}
X_2 = X_1-X_0' \in H_d(\omega_1'),
\qquad \text{where}
\qquad
d = \max\bigl({L_{\mathrm{max}}}^{0'1},N_0\alpha^2\bigr),$$ so we can insert ${\mathbf{P}}_{H_d(\omega_1')}$ in front of $u_2^{\gamma,\omega_2}$ in .
With this information in hand, we estimate . Apply Theorem \[J:Thm1\], recalling the crucial fact that the tube radius is now $N_2\gamma$. Repeating also the argument from section \[G:20:0\], but now with $u_2^{\gamma,\omega_2}$ replaced by ${\mathbf{P}}_{H_d(\omega_1')}u_2^{\gamma,\omega_2}$, we get $$\label{G:20:18}
\begin{aligned}
J_{{\boldsymbol{N}},{\boldsymbol{L}}}^{\mathbf\Sigma}
&\lesssim
\left( \frac{L_4}{N_4} \right)^{1/2}
N_2\gamma \left(L_0L_3\right)^{1/2}
\bigl(N_4 N_0 L_0'L_1 \bigr)^{1/2}
\\
&\qquad
\times
\sum_{\omega_1,\omega_2} \sum_{\omega_0',\omega_1'}
{\bigl\Vert u_1^{\gamma,\omega_1;\alpha,\omega_1'} \bigr\Vert}
{\left\Vert {\mathbf{P}}_{H_d(\omega_1')} u_2^{\gamma,\omega_2} \right\Vert}
{\bigl\Vert u_3 \bigr\Vert} {\bigl\Vert u_4 \bigr\Vert}
\\
&\lesssim
\left( \frac{L_4}{N_4} \right)^{1/2}
N_2 \left( \frac{L_2}{N_2} \right)^{1/2} \left(L_0L_3\right)^{1/2}
\bigl(N_4 N_0 L_0'L_1 \bigr)^{1/2}
\\
&\qquad
\times
\sum_{\omega_1,\omega_2}
\sqrt{B(\omega_1)}
{\bigl\Vert u_1^{\gamma,\omega_1} \bigr\Vert}
{\left\Vert u_2^{\gamma,\omega_2} \right\Vert}
{\bigl\Vert u_3 \bigr\Vert} {\bigl\Vert u_4 \bigr\Vert},
\end{aligned}$$ where $$\label{G:20:20}
B(\omega_1)
=
\sup_{(\tau,\xi), \; {| \xi |} \sim N_2}
\sum_{\genfrac{}{}{0pt}{1}{\omega_1' \in \Omega(\alpha)}{\theta(\omega_1',\omega_1) \lesssim \gamma}} \chi_{H_d(\omega_1')}(\tau,\xi)$$ The second inequality in was obtained by the Cauchy-Schwarz inequality, and we used also Lemma and .
If we can prove that $$\label{G:20:22}
\sup_{\omega \in \mathbb S^2} B(\omega)
\lesssim
\frac{N_0}{N_2},$$ then summing $\omega_1,\omega_2$ as in we get the desired estimate from .
By Lemma \[J:Lemma\], $$\sup_{\omega \in \mathbb S^2} B(\omega)
\lesssim
\frac{\gamma}{\alpha} + \frac{d}{N_2\alpha^2}.$$ The first term on the right hand side is comparable to $N_0/N_2$, by . As for the second term, this is also comparable to $N_0/N_2$ if $d = N_0\alpha^2$. In view of the definition of $d$, in , it then remains to consider $d={L_{\mathrm{max}}}^{0'1}$, which happens when $N_0\alpha^2 \le {L_{\mathrm{max}}}^{0'1}$. Then instead of we only get $$\label{G:20:24}
\sup_{\omega \in \mathbb S^2} B(\omega)
\lesssim
\frac{{L_{\mathrm{max}}}^{0'1}}{N_2\alpha^2},$$ but to compensate we can use the following replacement for : $$\label{G:20:26}
{\left\Vert {\mathbf{P}}_{\xi_0 \cdot \omega_1 \in I}
u_{0'12}^{\gamma,\omega_1,\omega_2} \right\Vert}
\lesssim
\bigl(N_4 (N_2\gamma)^2 {L_{\mathrm{min}}}^{0'1} \bigr)^{1/2}
{\left\Vert u_1^{\gamma,\omega_1} \right\Vert} {\left\Vert u_2^{\gamma,\omega_2} \right\Vert},$$ which in fact reduces to a more or less trivial volume estimate; a proof is given in [@Selberg:2008a Sect. 8]. If we now apply instead of , but of course with $u_2^{\gamma,\omega_2}$ replaced by ${\mathbf{P}}_{H_d(\omega_1')}u_2^{\gamma,\omega_2}$, then we must multiply by the square root of $$\frac{(N_2 \gamma)^2}{N_0{L_{\mathrm{max}}}^{0'1}}$$ But multiplying this by the right hand side of gives the factor in the right hand side of , so the net effect is the same.
The proof for case , subcase is now complete. Note that we did not use the assumption $L_0 \sim L_0'$ for this.
Case , subcase {#case-subcase}
---------------
Then by , and , $$\label{G:22:2}
{| q_{1234} |} \lesssim \phi^2
\lesssim \min(\theta_{03},\theta_{04}) \left(\frac{L_4}{N_0} \right)^{1/2}.$$ Comparing with , we then we get with an extra factor $({N_{\mathrm{min}}}^{34}/N_0)^{1/2}$, implying the desired estimate.
Case , subcase {#case-subcase-1}
---------------
This follows by the argument from the previous section, by symmetry (reverse the roles of 12 and 34). This works because we know that $\theta_{12},\theta_{34} \ll 1$, hence holds, as does its analogue for $\theta_{34}$.
This concludes the proof of Theorem \[C:Thm2\] for $L_0 \sim L_0'$.
Proof of the dyadic quadrilinear estimate, Part II {#E}
==================================================
It remains to prove Theorem \[C:Thm2\] when $$\label{G:24}
L_0 \ll L_0' \qquad \text{or} \qquad L_0 \gg L_0'.$$ By symmetry we may assume $$\label{G:24:2}
L_1 \le L_2,
\qquad
L_3 \le L_4.$$ Unlike in section \[G\], we do not use – as our main cases, the reason being that the argument used in cases and only works when $L_0 \sim L_0'$, since it relies on , and moreover $L_0,L_0'$ do not appear symmetrically in . Instead we use and as our main cases. In the case , we assume without loss of generality that $\theta_{12} \le \theta_{34}$. This we could also assume in case , of course, but it serves no purpose. Instead, in case we use the symmetry to assume without loss of generality that the first term in dominates. We also use the latter to split into two subcases. Thus, it suffices to consider the following cases:
\[E:1\] $$\begin{aligned}
{2}
\label{E:1a}
&\theta_{12} \ll \phi \lesssim \theta_{34},&
\qquad
&\min(\theta_{01},\theta_{02}) \ge \min(\theta_{03},\theta_{04}),
\\
\label{E:1c}
&\theta_{12} \ll \phi \lesssim \theta_{34},&
\qquad
&\min(\theta_{01},\theta_{02}) < \min(\theta_{03},\theta_{04}),
\\
\label{E:1e}
&\theta_{12}, \theta_{34} \ll \phi,&
\qquad
&\min(\theta_{01},\theta_{02}) \le \min(\theta_{03},\theta_{04}).\end{aligned}$$
The latter two we will further split into
\[E:0\] $$\begin{aligned}
{2}
\label{E:0a}
&L_2 > L_0',&
\qquad
&L_4 > L_0,
\\
\label{E:0b}
&L_2 \le L_0',&
\qquad
&L_4 > L_0,
\\
\label{E:0c}
&L_2 \le L_0',&
\qquad
&L_4 \le L_0,
\\
\label{E:0d}
&L_2 > L_0',&
\qquad
&L_4 \le L_0.\end{aligned}$$
We may assume that $$\label{G:24:4:2}
{N_{\mathrm{min}}}^{12}, {N_{\mathrm{min}}}^{34} \gg 1,$$ since otherwise trivial estimates analogous to apply.
Case
-----
Then by –, $${| q_{1234} |}
\lesssim \phi\theta_{34}
\lesssim
\biggl( \frac{{L_{\mathrm{max}}}^{0'12}}{N_0} \biggr)^{1/2}
\left( \frac{{L_{\mathrm{max}}}^{034}}{{N_{\mathrm{min}}}^{34}} \right)^{1/2},$$ so with notation as in , $$\label{E:6}
J_{{\boldsymbol{N}},{\boldsymbol{L}}}^{\mathbf\Sigma}
\lesssim
\biggl( \frac{{L_{\mathrm{max}}}^{0'12}}{N_0} \biggr)^{1/2}
\left( \frac{{L_{\mathrm{max}}}^{034}}{{N_{\mathrm{min}}}^{34}} \right)^{1/2}
{\bigl\Vert T_{L_0,L_0'}^{\pm_0}\mathcal F u_{0'12} \bigr\Vert} {\bigl\Vert u_{043} \bigr\Vert}.$$ If we apply Lemma \[D:Lemma3\] followed by Theorem \[M:Thm\], we get the desired estimate except in the case $N_0 \ll N_1 \sim N_2$. But we can do better by applying the following:
\[E:Lemma1\] Assuming holds, we have $$\label{E:12}
{\bigl\Vert T_{L_0,L_0'}^{\pm_0}\mathcal F u_{0'12} \bigr\Vert}
\lesssim
\left(N_0 {N_{\mathrm{min}}}^{012} {L_{\mathrm{min}}}^{0'12} {L_{\mathrm{med}}}^{0'12} \right)^{1/2}
{\left\Vert u_1 \right\Vert} {\left\Vert u_2 \right\Vert}.$$
The main point, comparing with in Theorem \[M:Thm\], is that ${N_{\mathrm{min}}}^{12}$ there is replaced by ${N_{\mathrm{min}}}^{012}$ in . Plugging the latter into , and estimating ${\left\Vert u_{043} \right\Vert}$ by the analogue of , we get .
If ${L_{\mathrm{max}}}^{0'12} = {L_{\mathrm{max}}}^{12}$, then holds by Lemma \[D:Lemma3\] and , so we assume ${L_{\mathrm{max}}}^{0'12} = L_0'$ for the rest of the proof.
Since $\theta_{12} \ll 1$, we have $\theta_{12} \lesssim \gamma$ with $\gamma$ as in . Then by Lemma \[B:Lemma5\], the left hand side of is dominated by the sum, with notation as in , $$S = \sum_{\omega_1,\omega_2}
{\left\Vert T_{L_0,L_0'}^{\pm_0} \mathcal F u^{\gamma,\omega_1,\omega_2}_{0'12} \right\Vert},$$ where the sum is over $\omega_1,\omega_2 \in \Omega(\gamma)$ with $\theta(\omega_1,\omega_2) \lesssim \gamma$. By , $$\label{E:12:2}
\mathcal F u_1^{\gamma,\omega_1} \subset H_{\max(L_1,N_1\gamma^2)}(\omega_1),
\qquad
\mathcal F u_2^{\gamma,\omega_2} \subset H_{\max(L_2,N_2\gamma^2)}(\omega_1),$$ where the latter relies on the assumption $\theta(\omega_1,\omega_2) \lesssim \gamma$. We conclude that $$\label{E:12:4}
\operatorname{supp}\mathcal F u^{\gamma,\omega_1,\omega_2}_{0'12}
\subset H_{d'}(\omega_1),
\qquad \text{where}
\qquad
d' = \max\left({L_{\mathrm{max}}}^{12},{N_{\mathrm{max}}}^{12}\gamma^2\right).$$ Then by Lemma \[D:Lemma5\], $$\label{E:12:5}
S
\lesssim
\sum_{\omega_1,\omega_2}
\left(\frac{d'}{L_0'}\right)^{1/2}
{\left\Vert u^{\gamma,\omega_1,\omega_2}_{0'12} \right\Vert}.$$ If $d' = {L_{\mathrm{max}}}^{12}$, use and sum $\omega_1,\omega_2$ as in to obtain $$\label{E:12:6}
S
\lesssim
\left(\frac{{L_{\mathrm{max}}}^{12}}{L_0'}\right)^{1/2}
\left( N_0 {N_{\mathrm{min}}}^{012} L_0' {L_{\mathrm{min}}}^{12} \right)^{1/2}
{\left\Vert u_1 \right\Vert} {\left\Vert u_2 \right\Vert},$$ proving . The other possibility is $d' = {N_{\mathrm{max}}}^{12}\gamma^2 \sim N_0L_0'/{N_{\mathrm{min}}}^{12}$, where the estimate holds by . Then by and , $$\label{E:12:8}
S
\lesssim
\left(\frac{N_0L_0'}{{N_{\mathrm{min}}}^{12}L_0'}\right)^{1/2}
\left( {N_{\mathrm{min}}}^{012} {N_{\mathrm{min}}}^{12} L_1 L_2 \right)^{1/2}
{\left\Vert u_1 \right\Vert} {\left\Vert u_2 \right\Vert},$$ completing the proof of Lemma \[E:Lemma1\].
This concludes case .
Case , subcase {#case-subcase-2}
---------------
This is covered by the proof for $L_0 \sim L_0'$, where it corresponds to case , subcase ; see sections \[G:14:0\]–\[G:20:7\].
Case , subcase {#E:0b:0}
---------------
Again we use the argument in sections \[G:14:0\]–\[G:20:7\], but with certain modifications. Observe that holds. Now we repeat the argument leading to , but instead of using Lemma \[D:Lemma3\] and Theorem \[M:Thm\] to estimate $T_{L_0,L_0'}^{\pm_0} \mathcal F u_{0'12}^{\gamma,\omega_1,\omega_2}$, we use Lemma \[E:Lemma1\]. Then we get $$\label{G:14:8:2}
\begin{aligned}
J_{{\boldsymbol{N}},{\boldsymbol{L}}}^{\mathbf\Sigma}
&\lesssim
\sum_{\omega_1,\omega_2} \left( \frac{L_4}{{N_{\mathrm{min}}}^{34}} \right)^{1/2}
{N_{\mathrm{max}}}^{12}\gamma \left(L_0L_3\right)^{1/2}
{\bigl\Vert T_{L_0,L_0'}^{\pm_0} \mathcal F u_{0'12}^{\gamma,\omega_1,\omega_2} \bigr\Vert}
{\left\Vert u_3 \right\Vert} {\left\Vert u_4 \right\Vert}
\\
&\lesssim
\left( \frac{L_4}{{N_{\mathrm{min}}}^{34}} \right)^{1/2}
{N_{\mathrm{max}}}^{12} \left( \frac{N_0L_0'}{N_1N_2} \right)^{1/2} \left(L_0L_3
\cdot N_0 {N_{\mathrm{min}}}^{012} L_1L_2 \right)^{1/2}
\prod_{j=1}^4 {\left\Vert u_j \right\Vert}
\\
&\lesssim
\left( \frac{N_0}{{N_{\mathrm{min}}}^{34}} N_0^2 L_0L_0'L_1L_2L_3L_4 \right)^{1/2}
\prod_{j=1}^4 {\left\Vert u_j \right\Vert},
\end{aligned}$$ and the last line is exactly as in .
Now we continue as in section \[G:14:0\]. We are done if $N_0 \lesssim {N_{\mathrm{min}}}^{34}$, or whenever we are able to gain an extra factor $({N_{\mathrm{min}}}^{34}/N_0)^{1/2}$ in ; the latter happens when $N_3 \ll N_0 \sim N_4$, in view of .
Thus, we are left with $N_4 \ll N_0 \sim N_3$. Then we proceed as in section \[G:20:0\]. We may assume (otherwise $N_0 \lesssim L_0'$, and then holds), hence Theorem \[J:Thm1\] applies, so in we can replace ${\bigl\Vert T_{L_0,L_0'}^{\pm_0} \mathcal F u_{0'12}^{\gamma,\omega_1,\omega_2} \bigr\Vert}$ by $$\label{G:14:8:4}
\sup_{I} {\left\Vert T_{L_0,L_0'}^{\pm_0} \mathcal F
{\mathbf{P}}_{\xi_0 \cdot \omega_1 \in I}
u_{0'12}^{\gamma,\omega_1,\omega_2} \right\Vert},$$ where the supremum is over $I \subset {\mathbb{R}}$ with ${| I |} = N_4$. By Theorem \[J:Thm3\], $$\label{E:42}
\sup_{I} {\left\Vert {\mathbf{P}}_{\xi_0 \cdot \omega_1 \in I} u_{0'12}^{\gamma,\omega_1,\omega_2} \right\Vert}
\lesssim
\left(N_4 {N_{\mathrm{min}}}^{12} L_1L_2 \right)^{1/2}
{\left\Vert u_1^{\gamma,\omega_1} \right\Vert} {\left\Vert u_2^{\gamma,\omega_2} \right\Vert}.$$ If we combine this with Lemma \[D:Lemma3\], we get $$\label{E:42:0}
\text{l.h.s.\eqref{G:14:8:4}}
\lesssim
\left(N_4 {N_{\mathrm{min}}}^{12} L_1L_2 \right)^{1/2}
{\left\Vert u_1^{\gamma,\omega_1} \right\Vert} {\left\Vert u_2^{\gamma,\omega_2} \right\Vert},$$ but this is not enough: It allows us to replace the combination $N_0{N_{\mathrm{min}}}^{012}L_1L_2$ in the second line of by $N_4{N_{\mathrm{min}}}^{12}L_1L_2$, but what we need is $N_4{N_{\mathrm{min}}}^{012}L_1L_2$. That is, we need $$\label{E:42:1}
\text{l.h.s.\eqref{G:14:8:4}}
\lesssim
\left(N_4 {N_{\mathrm{min}}}^{012} L_1L_2 \right)^{1/2}
{\left\Vert u_1^{\gamma,\omega_1} \right\Vert} {\left\Vert u_2^{\gamma,\omega_2} \right\Vert}.$$ If this holds, then we gain the necessary factor $(N_4/N_0)^{1/2}$ in .
So let us prove . We assume $N_0 \ll N_1 \sim N_2$, since otherwise reduces to . Recalling from the proof of Lemma \[E:Lemma1\], we use Lemma \[D:Lemma5\] to estimate the norm inside the supremum in , and then we apply either or the variation $$\label{E:42:2}
\sup_I {\left\Vert {\mathbf{P}}_{\xi_0 \cdot \omega_1 \in I}
u_{0'12}^{\gamma,\omega_1,\omega_2} \right\Vert}
\lesssim
\bigl(N_4 {N_{\mathrm{min}}}^{01} L_0'L_1 \bigr)^{1/2}
{\left\Vert u_1^{\gamma,\omega_1} \right\Vert} {\left\Vert u_2^{\gamma,\omega_2} \right\Vert},$$ which follows from Theorem \[J:Thm3\] via duality, again using the fact that $\gamma \ll 1$. Specifically, if $d'={L_{\mathrm{max}}}^{12}$, we use , whereas is used if $d'={N_{\mathrm{max}}}^{12}\gamma^2$; cf. and in the proof of Lemma \[E:Lemma1\]. Then follows.
This concludes case , subcase .
Case , subcase {#G:40:0:0}
---------------
The argument from section \[E:0b:0\] does not work, since the $L_4$ in becomes $L_0$, so there are two $L_0$’s; one of them is due to , and is unavoidable. The other, unwanted factor comes from estimating the null form in by Theorem \[G:Thm\] or Theorem \[J:Thm1\], so we have to avoid using those theorems. Instead we shall use an extra decomposition of the spatial frequencies into cubes to gain better control. Then we use Theorems \[J:Thm3\] and \[G:Thm2\], and also Lemma \[D:Lemma5\].
For the remainder of section \[E\], we change the notation from , writing now $$\label{D:134:10:2}
\theta_{12} \lesssim \gamma \equiv \biggl(\frac{N_0{L_{\mathrm{max}}}^{0'12}}{N_1N_2}\biggr)^{1/2}.$$
By , and , $$\label{G:40:0:2}
\phi
\lesssim \min(\theta_{03},\theta_{04})
\lesssim
\left( \frac{L_0}{N_0} \right)^{1/2},$$ hence $$\label{G:40:0}
{| q_{1234} |} \lesssim \phi\theta_{34}
\lesssim
\left( \frac{L_0}{N_0} \right)^{1/2}
\theta_{34}.$$ Therefore, applying Lemma \[B:Lemma5\] to the pair $(\pm_1\xi_1,\pm_2\xi_2)$ and Lemma \[B:Lemma4\] to the pair $(\pm_3\xi_3,\pm_4\xi_4)$, and recalling and , $$\begin{gathered}
\label{G:40:2}
J_{{\boldsymbol{N}},{\boldsymbol{L}}}^{\mathbf\Sigma}
\lesssim
\sum_{\omega_1,\omega_2}
\sum_{0 < \gamma_{34} \lesssim \gamma'}
\sum_{\omega_3,\omega_4}
\left( \frac{L_0}{N_0} \right)^{1/2}
\gamma_{34}
\\
\times
\int T_{L_0,L_0'}^{\pm_0} \mathcal F u_{0'12}^{\gamma,\omega_1,\omega_2}(X_0)
\cdot
\mathcal F u_{043}^{\gamma_{34},\omega_4,\omega_3}(X_0) \, dX_0,\end{gathered}$$ where $\gamma'$ is defined as in , $u_{0'12}^{\gamma,\omega_1,\omega_2}$ is defined as in , and similarly $$\label{G:14:6:2}
u^{\gamma_{34},\omega_4,\omega_3}_{043}
=
{\mathbf{P}}_{K^{\pm_0}_{N_0,L_0}}
\left( u_4^{\gamma_{34},\omega_4}
\overline{ u_3^{\gamma_{34},\omega_3} } \right)$$ The sums in are over $\omega_1,\omega_2 \in \Omega(\gamma)$ with $\theta(\omega_1,\omega_2) \lesssim \gamma$, over dyadic $\gamma_{34}$, and over $\omega_3,\omega_4 \in \Omega(\gamma_{34})$ satisfying $$\label{G:14:6:4}
3\gamma_{34} \le \theta(\omega_3,\omega_4) \le 12\gamma_{34}.$$ Due to this separation, $\theta_{34} \sim \gamma_{34}$ in the bilinear interaction of .
Recall that the spatial Fourier support of $u^{\gamma,\omega_1,\omega_2}_{0'12}$ is contained in a tube of radius $$r \sim {N_{\mathrm{max}}}^{12} \gamma$$ around ${\mathbb{R}}\omega_1$, where $\gamma$ is given by . Cover ${\mathbb{R}}$ by almost disjoint intervals $I$ of length $r$, and write $$u_{0'12}^{\gamma,\omega_1,\omega_2}
=
\sum_I
{\mathbf{P}}_{\xi_0 \cdot \omega_1 \in I} u_{0'12}^{\gamma,\omega_1,\omega_2},$$ where the sum has cardinality $O(N_0/r)$. The spatial frequency $\xi_0$ of the summand is restricted to a cube $$Q_0=Q_0(I)$$ of side-length comparable to $r$. Let $\mathcal Q_0$ denote the cover of ${\mathbb{R}}^3$ by almost disjoint translates of $Q_0$, and restrict the spatial frequencies $\xi_1,\xi_2,\xi_3,\xi_4$ to cubes $Q_1,Q_2,Q_3,Q_4 \in \mathcal Q_0$, respectively. Since $\xi_0=\xi_1-\xi_2$, then once $Q_1$ has been chosen, the choice of $Q_2$ is restricted to the set $$\label{G:40:3}
\left\{ Q_2 \in \mathcal Q_0 \colon Q_2 \cap (Q_1-Q_0) \neq \emptyset \right\},$$ which has cardinality $O(1)$. This implies that, by the Cauchy-Schwarz inequality, $$\label{G:40:4}
\sum_{Q_1,Q_2}
{\bigl\Vert {\mathbf{P}}_{Q_1}u_1^{\gamma,\omega_1} \bigr\Vert}
{\bigl\Vert {\mathbf{P}}_{Q_2}u_2^{\gamma,\omega_2} \bigr\Vert}
\lesssim
{\bigl\Vert u_1^{\gamma,\omega_1} \bigr\Vert}
{\bigl\Vert u_2^{\gamma,\omega_2} \bigr\Vert},$$ where the sum over $Q_2$ of course is restricted to the set . The pair $(Q_3,Q_4)$ is similarly restricted, since $\xi_0=\xi_4-\xi_3$.
After this extra decomposition, is replaced by $$\begin{gathered}
\label{G:40:12}
J_{{\boldsymbol{N}},{\boldsymbol{L}}}^{\mathbf\Sigma}
\lesssim
\sum_{\omega_1,\omega_2}
\sum_{0 < \gamma_{34} \lesssim \gamma'}
\sum_{\omega_3,\omega_4}
\sum_I
\sum_{Q_1,Q_2}
\sum_{Q_3,Q_4}
\left( \frac{L_0}{N_0} \right)^{1/2} \gamma_{34}
\\
\times
\int T_{L_0,L_0'}^{\pm_0} \mathcal F u_{0'12}^{\gamma,\omega_1,\omega_2;Q_1,Q_2}(X_0)
\cdot
\mathcal F u_{043}^{\gamma_{34},\omega_4,\omega_3;Q_4,Q_3}(X_0) \, dX_0,\end{gathered}$$ where $$\label{G:40:13}
\begin{aligned}
u_{0'12}^{\gamma,\omega_1,\omega_2;Q_1,Q_2}
&= {\mathbf{P}}_{K^{\pm_0}_{N_0,L_0'}}
\left( {\mathbf{P}}_{Q_1} u_1^{\gamma,\omega_1}
\overline{ {\mathbf{P}}_{Q_2} u_2^{\gamma,\omega_2} } \right),
\\
u_{043}^{\gamma_{34},\omega_4,\omega_3;Q_4,Q_3}
&= {\mathbf{P}}_{K^{\pm_0}_{N_0,L_0}}
\left( {\mathbf{P}}_{Q_4} u_4^{\gamma_{34},\omega_4}
\overline{ {\mathbf{P}}_{Q_3} u_3^{\gamma_{34},\omega_3} } \right).
\end{aligned}$$ The sum over $Q_2$ in is restricted to the set determined by $Q_1$ and $I$, and similarly for the pair $(Q_3,Q_4)$.
Consider the integral in for a fixed choice of $\gamma_{34}$, the $\omega$’s, $I$ and the $Q$’s. We use the notation – for the bilinear interactions in this integral, so in particular $\xi_j \in Q_j$ and ${| \xi_j |} \sim N_j$ for $j=1,2,3,4$, recalling . Our plan is now to apply Lemma \[D:Lemma5\]. By we have $$\label{G:40:14}
\tau_0' + \xi_0 \cdot \omega_1 = O(d'),
\qquad
\text{where}
\qquad
d'=\max\left(L_2,{N_{\mathrm{max}}}^{12}\gamma^2\right).$$ We claim that also $$\label{G:40:16}
\tau_0 + \xi_0 \cdot \omega_3 = c + O(d),
\qquad
\text{where}
\qquad
d=\max\left(L_4,\frac{r^2}{{N_{\mathrm{min}}}^{34}},r\gamma_{34}\right),$$ and $c \in {\mathbb{R}}$ is a constant depending on $(Q_3,Q_4)$ and $(\omega_3,\omega_4)$.
Let us prove . Denote by $\xi_j^*$ the center of the cube $Q_j$, so that $$\label{G:40:17:0}
{| \xi_j-\xi_j^* |} \lesssim r.$$ Let $\omega_j^* = \pm_j\xi_j^*/{| \xi_j^* |}$. Replacing the side-length $r$ of the cubes by $2r$ if necessary, we may assume that $\omega_j^* \in \Gamma_{\gamma_{34}}(\omega_j)$ for $j=3,4$. Since $\theta(\omega_3,\omega_4) \lesssim \gamma_{34}$, we then have $$\label{G:40:17}
\theta(\omega_j^*,\omega_3) \lesssim \gamma_{34} \qquad \text{for $j=3,4$}.$$ Now write, for $j=3,4$, $$\label{G:40:18}
\tau_j + \xi_j \cdot \omega_3 =
\tau_j + \xi_j \cdot \omega_j^*
+ (\xi_j-\xi_j^*) \cdot (\omega_3 - \omega_j^*)
+ c_j,$$ where $$c_j = \xi_j^* \cdot (\omega_3 - \omega_j^*).$$ Since $\xi_j \in Q_j$ and ${| \xi_j |} \sim N_j$, we conclude from that $$\label{G:40:20}
\tau_j + \xi_j \cdot \omega_j^* = O\left(\max\left(L_j,\frac{r^2}{N_j}\right)\right)
\qquad (j=3,4).$$ By and , $$\label{G:40:22}
(\xi_j-\xi_j^*) \cdot (\omega_3 - \omega_j^*)
= O(r\gamma_{34}).$$ Now plug the estimates and into for $j=3,4$, and subtract. Since $\tau_0=\tau_4-\tau_3$ and $\xi_0=\xi_4-\xi_3$, this proves , with $c=c_4-c_3$.
In view of , and Lemma \[D:Lemma5\], we can dominate the integral in by the product of $$\label{G:40:30:2}
\left(\min\left(1,\frac{d'}{L_0'}\right) \right)^{1/2}
{\left\Vert u_{0'12}^{\gamma,\omega_1,\omega_2;Q_1,Q_2} \right\Vert}$$ and $$\label{G:40:30:4}
\left(\frac{d}{L_0} \right)^{1/2}
{\left\Vert u_{043}^{\gamma_{34},\omega_4,\omega_3;Q_4,Q_3} \right\Vert}.$$
By Theorem \[J:Thm3\], $$\label{G:40:32}
{\left\Vert u_{0'12}^{\gamma,\omega_1,\omega_2;Q_1,Q_2} \right\Vert}
\lesssim C\,
{\bigl\Vert {\mathbf{P}}_{Q_1}u_1^{\gamma,\omega_1} \bigr\Vert}
{\bigl\Vert {\mathbf{P}}_{Q_2}u_2^{\gamma,\omega_2} \bigr\Vert}$$ holds with $$\begin{aligned}
\label{G:40:32:4}
C^2 &\sim r{N_{\mathrm{min}}}^{01}L_0'L_1,
\\
\label{G:40:32:2}
C^2 &\sim r{N_{\mathrm{min}}}^{12}L_1L_2.\end{aligned}$$ From the definitions of $d'$ and $\gamma$ (see and ) and by the assumption , $$\label{G:40:32:6}
\frac{d'}{L_0'}
\sim \max\left(\frac{L_2}{L_0'},\frac{N_0}{{N_{\mathrm{min}}}^{12}}\right).$$ If $N_0 \sim {N_{\mathrm{max}}}^{12}$, we estimate by l.h.s. and use . If, on the other hand, $N_0 \ll N_1 \sim N_2$, then we combine with , observing that the product of with the minimum of and is dominated by $rN_0L_1L_2$. We conclude: $$\label{G:40:32:8}
\eqref{G:40:30:2} \lesssim \left(r{N_{\mathrm{min}}}^{012}L_1L_2\right)^{1/2}
{\bigl\Vert {\mathbf{P}}_{Q_1}u_1^{\gamma,\omega_1} \bigr\Vert}
{\bigl\Vert {\mathbf{P}}_{Q_2}u_2^{\gamma,\omega_2} \bigr\Vert}.$$
We further claim that $$\label{G:40:34}
{\left\Vert u_{043}^{\gamma_{34},\omega_4,\omega_3;Q_4,Q_3} \right\Vert}
\lesssim C\,
{\bigl\Vert {\mathbf{P}}_{Q_3}u_3^{\gamma_{34},\omega_3} \bigr\Vert}
{\bigl\Vert {\mathbf{P}}_{Q_4}u_4^{\gamma_{34},\omega_4} \bigr\Vert}$$ holds with $$\begin{aligned}
\label{G:40:36}
C^2
&\sim
r^3 L_3,
\\
\label{G:40:38}
C^2
&\sim
r {N_{\mathrm{min}}}^{34} L_3 L_4,
\\
\label{G:40:40}
C^2
&\sim
\frac{r^2 L_3 L_4}{\gamma_{34}}.\end{aligned}$$ In fact, holds by Theorem \[G:Thm2\], in view of the separation assumption ; holds by Theorem \[J:Thm3\], and reduces to a trivial volume estimate (see the proof of Theorem \[G:Thm2\] in [@Selberg:2008a]).
Now observe that $d$, defined by , times the minimum of –, is comparable to $r^3 L_3 L_4$. Therefore, $$\label{G:40:42}
\eqref{G:40:30:4}
\lesssim
\left(\frac{r^3L_3L_4}{L_0}\right)^{1/2}
{\bigl\Vert {\mathbf{P}}_{Q_3}u_3^{\gamma_{34},\omega_3} \bigr\Vert}
{\bigl\Vert {\mathbf{P}}_{Q_4}u_4^{\gamma_{34},\omega_4} \bigr\Vert}.$$
Now estimate the integral in by the product of and , and use and . The result is $$\label{G:40:42:2}
\begin{aligned}
J_{{\boldsymbol{N}},{\boldsymbol{L}}}^{\mathbf\Sigma}
&\lesssim
\sum_{\omega_1,\omega_2}
\sum_{0 < \gamma_{34} \lesssim \gamma'}
\sum_{\omega_3,\omega_4}
\sum_I
\left( \frac{L_0}{N_0} \right)^{1/2}
\gamma_{34}
\left(
\frac{r^3L_3L_4}{L_0}
\cdot
r{N_{\mathrm{min}}}^{012}L_1L_2 \right)^{1/2}
\\
&\quad
\times
\left(\sum_{Q_1,Q_2}
{\bigl\Vert {\mathbf{P}}_{Q_1}u_1^{\gamma,\omega_1} \bigr\Vert}
{\bigl\Vert {\mathbf{P}}_{Q_2}u_2^{\gamma,\omega_2} \bigr\Vert}\right)
\left(\sum_{Q_3,Q_4}
{\bigl\Vert {\mathbf{P}}_{Q_3}u_3^{\gamma_{34},\omega_3} \bigr\Vert}
{\bigl\Vert {\mathbf{P}}_{Q_4}u_4^{\gamma_{34},\omega_4} \bigr\Vert}\right)
\\
&\lesssim \left( \sum_I \frac{r}{N_0} \right)
\sum_{0 < \gamma_{34} \lesssim \gamma'} r\gamma_{34}
\left( \frac{{N_{\mathrm{min}}}^{012}}{N_0} \right)^{1/2}
\left( N_0^2 L_1L_2 L_3L_4 \right)^{1/2}
\\
&\quad
\times
\left(\sum_{\omega_1,\omega_2}
{\bigl\Vert u_1^{\gamma,\omega_1} \bigr\Vert}
{\bigl\Vert u_2^{\gamma,\omega_2} \bigr\Vert} \right)
\left( \sum_{\omega_3,\omega_4}
{\bigl\Vert u_3^{\gamma_{34},\omega_3} \bigr\Vert}
{\bigl\Vert u_4^{\gamma_{34},\omega_4} \bigr\Vert} \right)
\\
&\lesssim
\frac{{N_{\mathrm{max}}}^{12}\gamma\gamma'}{(L_0L_0')^{1/2}}
\left( \frac{{N_{\mathrm{min}}}^{012}}{N_0} \right)^{1/2}
\left( N_0^2 L_0L_0'L_1L_2 L_3L_4 \right)^{1/2}
{\left\Vert u_1 \right\Vert}{\left\Vert u_2 \right\Vert}{\left\Vert u_3 \right\Vert}{\left\Vert u_4 \right\Vert},
\end{aligned}$$ where to get the second inequality we summed the $Q$’s using and its analogue for $(Q_3,Q_4)$. In the final step we used the definition $r \sim {N_{\mathrm{max}}}^{12}\gamma$, we summed $I$ using the fact that the index set has cardinality $O(N_0/r)$, we summed the $\omega$’s as in , and finally we used the fact that $$\label{G:40:50}
\sum_{0 < \gamma_{34} \lesssim \gamma'} \gamma_{34}
\sim \gamma',$$ where the sum is over dyadic $\gamma_{34}$, of course.
Note that the above implies if the expression $$\label{G:40:52}
A = \frac{({N_{\mathrm{max}}}^{12}\gamma\gamma')^2{N_{\mathrm{min}}}^{012}}{N_0L_0L_0'}$$ is $O(1$). In view of , and , $$\label{G:40:54}
A
\lesssim
\frac{({N_{\mathrm{max}}}^{12})^2{N_{\mathrm{min}}}^{012}}{N_0L_0L_0'}
\cdot \frac{N_0L_0'}{N_1N_2}
\min\left(1,\frac{L_0}{{N_{\mathrm{min}}}^{34}}\right).$$ If we use the second factor in the last minimum, we get $$\label{G:40:56}
A
\lesssim
\frac{{N_{\mathrm{max}}}^{12}{N_{\mathrm{min}}}^{012}}{{N_{\mathrm{min}}}^{12}{N_{\mathrm{min}}}^{34}}
\lesssim \frac{N_0}{{N_{\mathrm{min}}}^{34}},$$ where we used . This proves except when $${N_{\mathrm{min}}}^{34} \ll N_0,$$ which we now assume. If $\pm_0\neq\pm_{43}$, then $N_0 \lesssim L_0$ by Lemma \[F:Lemma2\], so we can estimate the first factor in the minimum in by $1 \lesssim L_0/N_0$, thereby gaining a factor ${N_{\mathrm{min}}}^{34}/N_0$ compared to . If, on the other hand, $\pm_0=\pm_{43}$, then by Lemma \[F:Lemma1\] and , $$\min(\theta_{03},\theta_{04}) \lesssim \frac{{N_{\mathrm{min}}}^{34}}{N_0} \theta_{34}
\lesssim
\frac{{N_{\mathrm{min}}}^{34}}{N_0} \left( \frac{L_0}{{N_{\mathrm{min}}}^{34}} \right)^{1/2}
= \left(\frac{{N_{\mathrm{min}}}^{34}}{N_0} \right) \left( \frac{L_0}{N_0} \right)^{1/2},$$ which means that compared to we gain a factor $({N_{\mathrm{min}}}^{34}/N_0)^{1/2}$, which comes up squared in .
This completes case , subcase .
Case , subcase {#G:40:51}
---------------
This adds another layer of difficulty compared to the previous section, for a certain asymmetric interaction. As far as possible, however, we repeat the preceding argument.
The only difference from the previous case is that now $L_2 > L_0'$, instead of $L_2 \le L_0'$. This difference only shows up in the expression for $\gamma$, however, and this expression is not used explicitly in the previous section until the estimate . But in the present case, $d'/L_0' > 1$, hence is just equal to the left hand side of , so instead of we use with constant $C$ as in (using will not work now). Comparing with , we see that there will only be a problem if $N_2 \ll N_0 \sim N_1$.
To be precise, instead of we will now have $$\label{G:40:52:2}
A = \frac{({N_{\mathrm{max}}}^{12}\gamma\gamma')^2{N_{\mathrm{min}}}^{12}}{N_0L_0L_2},$$ leading to $$\label{G:40:54:2}
A
\lesssim
\frac{({N_{\mathrm{max}}}^{12})^2{N_{\mathrm{min}}}^{12}}{N_0L_0L_2}
\cdot \frac{N_0L_2}{N_1N_2}
\cdot \min\left(1,\frac{L_0}{{N_{\mathrm{min}}}^{34}}\right)
= \frac{{N_{\mathrm{min}}}^{12}}{{N_{\mathrm{min}}}^{012}} \times \text{r.h.s.}\eqref{G:40:54},$$ so we are done except in the case $$N_2 \ll N_0 \sim N_1,$$ which we now assume. Then we must somehow gain a factor $N_2/N_0$ in . We use the same idea as in section \[G:20:7\]. We may assume $$N_2 \gg 1,$$ since otherwise applies. We may further assume $$\pm_0=\pm_{012},$$ since otherwise applies.
Then holds, and we use this to make an extra angular decomposition in the analysis of the previous section, for the pair $(\pm_0\xi_0,\pm_1\xi_1)$. In view of , the effect of this extra decomposition is that we can replace ${\mathbf{P}}_{Q_1} u_1^{\gamma,\omega_1}$ and ${\mathbf{P}}_{Q_2} u_2^{\gamma,\omega_2}$ in by, respectively, $$\label{G:40:13:2}
{\mathbf{P}}_{Q_1} u_1^{\gamma,\omega_1;\alpha,\omega_1'},
\qquad
{\mathbf{P}}_{H_d(\omega_1')} {\mathbf{P}}_{Q_2} u_2^{\gamma,\omega_2},$$ where $u_1^{\gamma,\omega_1;\alpha,\omega_1'}$ is defined by and $d$ is given by . Here $\omega_1' \in \Omega(\alpha)$. There is also a vector $\omega_0' \in \Omega(\alpha)$, but since $\theta(\omega_0',\omega_1') \lesssim \alpha$ we know that only $O(1)$ $\omega_0'$’s can interact with a given $\omega_1'$, hence summing $\omega_0'$ is not a problem.
A key observation is that the spatial output $\xi_0$ is now restricted to a tube of radius $$r' \sim N_0\alpha \sim N_2\gamma$$ around ${\mathbb{R}}\omega_0'$, and the relation between this and the radius $r \sim N_0 \gamma$ used in the previous section is $$\frac{r'}{r} \sim \frac{N_2}{N_0}.$$
If we now repeat the decomposition into cubes as in the previous section, but now with $r$ replaced by $r'$, then apply with $C$ as in but with $r$ replaced by $r'$ and with the substitutions , and we apply also with $r$ replaced by $r'$, then we get $$\label{G:40:42:4}
\begin{aligned}
J_{{\boldsymbol{N}},{\boldsymbol{L}}}^{\mathbf\Sigma}
&\lesssim
\sum_{\omega_1,\omega_2}
\sum_{0 < \gamma_{34} \lesssim \gamma'}
\sum_{\omega_3,\omega_4}
\sum_{\omega_1'}
\sum_I
\left( \frac{L_0}{N_0} \right)^{1/2}
\gamma_{34}
\left( \frac{(r')^3L_3L_4}{L_0}
\cdot
r'N_0L_0'L_1 \right)^{1/2}
\\
&\quad
\times
{\bigl\Vert u_1^{\gamma,\omega_1;\alpha,\omega_1'} \bigr\Vert}
{\bigl\Vert {\mathbf{P}}_{H_d(\omega_1')}u_2^{\gamma,\omega_2} \bigr\Vert}
{\bigl\Vert u_3^{\gamma_{34},\omega_3} \bigr\Vert}
{\bigl\Vert u_4^{\gamma_{34},\omega_4} \bigr\Vert}
\\
&\lesssim \left( \sum_I \frac{r'}{N_0} \right)
\sum_{0 < \gamma_{34} \lesssim \gamma'}
\frac{r'\gamma_{34}}{(L_0L_2)^{1/2}}
\left( N_0^2 L_0 L_0' L_1 L_2 L_3 L_4 \right)^{1/2}
\\
&\quad
\times
\left(\sum_{\omega_1,\omega_2}
\sum_{\omega_1'}
{\bigl\Vert u_1^{\gamma,\omega_1;\alpha,\omega_1'} \bigr\Vert}
{\bigl\Vert {\mathbf{P}}_{H_d(\omega_1')}u_2^{\gamma,\omega_2} \bigr\Vert} \right)
\left( \sum_{\omega_3,\omega_4}
{\bigl\Vert u_3^{\gamma_{34},\omega_3} \bigr\Vert}
{\bigl\Vert u_4^{\gamma_{34},\omega_4} \bigr\Vert} \right)
\\
&\lesssim
\frac{r'\gamma'}{(L_0L_2)^{1/2}}
\left( N_0^2 L_0L_0'L_1L_2 L_3L_4 \right)^{1/2}
\sqrt{\sup_{\omega_1 \in \mathbb S^2} B(\omega_1)}
{\left\Vert u_1 \right\Vert} {\left\Vert u_2 \right\Vert} {\left\Vert u_3 \right\Vert} {\left\Vert u_4 \right\Vert},
\end{aligned}$$ where $B(\omega_1)$ is defined by . So now instead of we have $$A = \frac{(r'\gamma')^2}{L_0L_2} \sup_{\omega_1 \in \mathbb S^2} B(\omega_1),$$ and is replaced by $$\begin{aligned}
A
&\lesssim
\frac{N_2^2}{L_0L_2}
\cdot \frac{N_0L_2}{N_1N_2}
\min\left(1,\frac{L_0}{{N_{\mathrm{min}}}^{34}}\right) \sup_{\omega_1 \in \mathbb S^2} B(\omega_1)
\\
&\lesssim
\frac{N_2}{L_0}
\min\left(1,\frac{L_0}{{N_{\mathrm{min}}}^{34}}\right) \sup_{\omega_1 \in \mathbb S^2} B(\omega_1).\end{aligned}$$
When holds we are done, since then we get $$\label{G:40:56:2}
A \lesssim \frac{N_0}{L_0} \min\left(1,\frac{L_0}{{N_{\mathrm{min}}}^{34}}\right).$$ and by the same argument as at the end of the previous section we also know how to deal with the case ${N_{\mathrm{min}}}^{34} \ll N_0$.
If, on the other hand, does not hold, then as shown in section \[G:20:7\] we have instead . But to compensate we can use the fact that holds with $$C^2 \sim r' (N_2\gamma)^2 {L_{\mathrm{min}}}^{0'1},$$ as follows from . Then as observed in section \[G:20:7\], the net effect is the same, hence holds.
This completes case , subcase
Case , subcases and
--------------------
Then by , and , $${| q_{1234} |} \lesssim \phi^2
\lesssim \min(\theta_{03},\theta_{04})^2
\lesssim \min(\theta_{03},\theta_{04}) \left( \frac{L_4}{N_0} \right)^{1/2},$$ hence we can proceed as in section \[G:14:0\], but recalling also that we have Lemma \[E:Lemma1\] at our disposal. The result is that we can dominate $J_{{\boldsymbol{N}},{\boldsymbol{L}}}^{\mathbf\Sigma}$ by the last line of , but without the factor $N_0/{N_{\mathrm{min}}}^{34}$. Thus, holds.
Case , subcase {#case-subcase-3}
---------------
Here we would like to follow as closely as possible the argument for case , subcase , from section \[G:40:0:0\].
Since $\theta_{12},\theta_{34} \ll 1$, we have and similarly $$\label{D:134:10:4}
\theta_{34} \lesssim \gamma' \equiv \left(\frac{N_0L_0}{N_3N_4}\right)^{1/2},$$ which replaces .
We still have , but is replaced by $$\label{G:50:4}
{| q_{1234} |} \lesssim \phi^2,$$ hence the factor $\gamma_{34}$ in is replaced by the upper bound for $\phi$ in .
Since there is no $\gamma_{34}$, it may seem that we have a problem with the estimate with $C$ as in , since this is a null form estimate which requires that we have at least a square root of $\gamma_{34}$ (the dyadic size of $\theta_{34}$).
But the combination , is only used when we pick up the third factor $r\gamma_{34}$ in the maximum defining $d$ in , so we are still able to use .
Proceeding as in section \[G:40:0:0\], we then get with the following modifications: The factors $\gamma_{34}$ in the first and third lines are replaced by the upper bound in , and instead of the factor $\gamma'$ in the last line, which comes from the sum , we now have $$\label{G:50:6}
\left(\frac{L_0}{N_0}\right)^{1/2} \sum_{0 < \gamma_{34} \lesssim \gamma'} 1.$$ Of course the sum diverges, unless we can further restrict the range of the dyadic number $\gamma_{34}$.
The separation assumption is only needed when we apply the null form estimate , i.e., when the factor $r\gamma_{34}$ dominates in the definition of $d$ in ; then in particular, $$\label{D:134:10:6}
\gamma_{34} \gtrsim \frac{r}{{N_{\mathrm{min}}}^{34}} \sim \frac{{N_{\mathrm{max}}}^{12}}{{N_{\mathrm{min}}}^{34}}\gamma
= \frac{{N_{\mathrm{max}}}^{12}}{{N_{\mathrm{min}}}^{34}} \left( \frac{N_0L_0'}{N_1N_2} \right)^{1/2}.$$ On the other hand, we also have the upper bound for $\gamma_{34}$. The cardinality of the set of dyadic numbers $\gamma_{34}$ satisfying both and is comparable to $$\label{G:50:10}
\log
\left( \frac{\gamma'}{r/{N_{\mathrm{min}}}^{34}} \right)
\sim
\log \left( \frac{(N_1N_2)^{1/2}}{{N_{\mathrm{max}}}^{12}}
\cdot \frac{{N_{\mathrm{min}}}^{34}}{(N_3N_4)^{1/2}} \left(\frac{L_0}{L_0'}\right)^{1/2}
\right)
\lesssim \log L_0,$$ so the corresponding part of the sum in is $O(\log L_0)$.
It then remains to consider $$\theta_{34} \ll \frac{r}{{N_{\mathrm{min}}}^{34}},$$ but then we do not need the separation assumption , so here we can avoid a summation over $\gamma_{34}$ altogether by using Lemma \[B:Lemma5\] instead of Lemma \[B:Lemma4\].
Case , subcase {#case-subcase-4}
---------------
This follows by the argument from section with the same modifications as in the previous section. Now $L_0'$ in is replaced by $L_2$, but this does not change the final estimate in (recall that all the $L$’s are greater than or equal to one).
This completes the proof of Theorem \[C:Thm2\].
Summation of the dyadic pieces {#Y}
==============================
By summing the dyadic estimates from Theorem \[C:Thm2\], we prove that holds for any $s > 0$ and all sufficiently small $\varepsilon > 0$ (depending on $s$). Split the integral into two parts: $$J^{\boldsymbol\Sigma}
=
J^{\boldsymbol\Sigma}_{{| \xi_0 |} \ge 1}
+
J^{\boldsymbol\Sigma}_{{| \xi_0 |} < 1},$$ by restricting to the regions where ${| \xi_0 |} \ge 1$ and ${| \xi_0 |} < 1$, respectively.
The high frequency part
-----------------------
Recall that holds for $J^{\boldsymbol\Sigma}_{{| \xi_0 |} \ge 1}$. Now we combine the estimate from Theorem \[C:Thm2\] with the trivial estimate $$\label{Y:2}
J_{{\boldsymbol{N}},{\boldsymbol{L}}}^{\mathbf\Sigma}
\lesssim \left( \left({N_{\mathrm{min}}}^{012} {N_{\mathrm{min}}}^{034}\right)^3
{L_{\mathrm{min}}}^{0'12} {L_{\mathrm{min}}}^{034} \right)^{1/2}
\prod_{j=1}^4{\left\Vert u_j \right\Vert},$$ which is immediate from and Lemma \[D:Lemma3\]. Taking to the power $8\varepsilon$ and to the power $1-8\varepsilon$, we get $$J_{{\boldsymbol{N}},{\boldsymbol{L}}}^{\mathbf\Sigma}
\lesssim
\left( \left({N_{\mathrm{min}}}^{012} {N_{\mathrm{min}}}^{034}\right)^3
{L_{\mathrm{min}}}^{0'12} {L_{\mathrm{min}}}^{034} \right)^{4\varepsilon}
\left( N_0^2 L_0L_0'L_1L_2L_3L_4 \right)^{1/2-4\varepsilon}
\log{\langle L_0 \rangle}
\prod_{j=1}^4{\left\Vert u_j \right\Vert}.$$ Estimating $\log{\langle L_0 \rangle} \lesssim L_0^{\varepsilon}$ and $({L_{\mathrm{min}}}^{0'12} {L_{\mathrm{min}}}^{034})^{4\varepsilon} \lesssim (L_1L_3)^{4\varepsilon}$, inserting the above into , and recalling the notation , we see that is enough to prove $$\label{Y:10}
S \lesssim {\left\Vert F_1 \right\Vert}{\left\Vert F_2 \right\Vert}{\left\Vert F_3 \right\Vert}{\left\Vert F_4 \right\Vert},$$ where $$S = \sum_{{\boldsymbol{N}},{\boldsymbol{L}}} \frac{N_4^s \bigl( {N_{\mathrm{min}}}^{012} {N_{\mathrm{min}}}^{034} \bigr)^{12\varepsilon}}{N_0^{8\varepsilon}(N_1N_2N_3)^s (L_0L_0'L_1 L_2 L_3L_4)^{\varepsilon}}
\prod_{j=1}^4 {\bigl\Vert \chi_{K^{\pm_j}_{N_j,L_j}} F_j \bigr\Vert}.$$ The sum over ${\boldsymbol{N}}$ is restricted by the condition and its counterpart for the inded $034$. Recall that all the $N$’s and $L$’s are greater than or equal to one.
Summing ${\boldsymbol{L}}$ is trivial: $$\sum_{{\boldsymbol{L}}} \frac{1}{(L_0L_0'L_1 L_2 L_3L_4)^{\varepsilon}}
\prod_{j=1}^4 {\Bigl\Vert \chi_{K^{\pm_j}_{N_j,L_j}} F_j \Bigr\Vert}
\le C
\prod_{j=1}^4 {\left\Vert \chi_{{\langle \xi_j \rangle} \sim N_j} F_j \right\Vert},$$ where $C = \sum_{{\boldsymbol{L}}} (L_0L_0'L_1 L_2 L_3L_4)^{-\varepsilon} < \infty$, so it only remains to prove for the reduced sum $$\label{Y:20}
S' = \sum_{{\boldsymbol{N}}} \frac{N_4^s \bigl( {N_{\mathrm{min}}}^{012} {N_{\mathrm{min}}}^{034} \bigr)^{12\varepsilon}}{N_0^{8\varepsilon}(N_1N_2N_3)^s}
\prod_{j=1}^4 {\left\Vert \chi_{{\langle \xi_j \rangle} \sim N_j} F_j \right\Vert}.$$ Since $({N_{\mathrm{min}}}^{012} {N_{\mathrm{min}}}^{034})^{12\varepsilon} \le N_0^{24\varepsilon} \lesssim ({N_{\mathrm{max}}}^{12})^{24\varepsilon}$, $$\label{Y:22}
S' \lesssim \sum_{{\boldsymbol{N}}} \frac{N_4^s}{N_0^{8\varepsilon}(N_1N_2)^{s-24\varepsilon} N_3^s}
\prod_{j=1}^4 {\left\Vert \chi_{{\langle \xi_j \rangle} \sim N_j} F_j \right\Vert}.$$ To ensure that $s-25\varepsilon > 0$, we choose $\varepsilon > 0$ so small that $25\varepsilon \le s$.
We now split $S' = S_1 + S_2 + S_3$, corresponding to the cases $N_4 \lesssim N_0 \sim N_3$, $N_0 \ll N_3 \sim N_4$ and $N_3 \ll N_0 \sim N_4$, respectively.
### The case $N_4 \lesssim N_0 \sim N_3$
Since $\sum_{N_4 \lesssim N_0} N_4^s \sim N_0^s \sim N_3^s$, $$S_1 \lesssim \sum_{N_0,N_1,N_2,N_3} \frac{1}{N_0^{7\varepsilon}(N_1N_2)^{s-24\varepsilon} N_3^\varepsilon}
\prod_{j=1}^4 {\left\Vert \chi_{{\langle \xi_j \rangle} \sim N_j} F_j \right\Vert},$$ and this is trivially bounded by right hand side of .
### The case $N_0 \ll N_3 \sim N_4$
Then $$S_2 \lesssim \sum_{{\boldsymbol{N}}} \frac{\chi_{N_3 \sim N_4}}{N_0^{8\varepsilon}(N_1N_2)^{s-24\varepsilon}}
\prod_{j=1}^4 {\left\Vert \chi_{{\langle \xi_j \rangle} \sim N_j} F_j \right\Vert}.$$ Here $N_0,N_1,N_2$ sum outright, whereas $N_3 \sim N_4$ can be summed using the Cauchy-Schwarz inequality: $$\sum_{N_3 \sim N_4 \ge 1} {\left\Vert \chi_{{\langle \xi_3 \rangle} \sim N_3} F_3 \right\Vert} {\left\Vert \chi_{{\langle \xi_4 \rangle} \sim N_4} F_4 \right\Vert}
\lesssim {\left\Vert F_3 \right\Vert}{\left\Vert F_4 \right\Vert}.$$
### The case $N_3 \ll N_0 \sim N_4$
Then $N_4 \lesssim {N_{\mathrm{max}}}^{12}$. Now is too crude, but from we see that we can reduce to the sum $$\begin{aligned}
S_3 &= \sum_{{\boldsymbol{N}}} \frac{N_4^s \chi_{N_0 \sim N_4 \lesssim {N_{\mathrm{max}}}^{12}}}{N_0^{8\varepsilon}({N_{\mathrm{min}}}^{12})^{s-12\varepsilon}({N_{\mathrm{max}}}^{12})^s N_3^{s-12\varepsilon}}
\prod_{j=1}^4 {\left\Vert \chi_{{\langle \xi_j \rangle} \sim N_j} F_j \right\Vert}
\\
&\lesssim \sum_{{\boldsymbol{N}}} \frac{\chi_{N_0 \sim N_4}}{N_0^{7\varepsilon}({N_{\mathrm{min}}}^{12})^{s-12\varepsilon}({N_{\mathrm{max}}}^{12})^{\varepsilon} N_3^{s-12\varepsilon}}
\prod_{j=1}^4 {\left\Vert \chi_{{\langle \xi_j \rangle} \sim N_j} F_j \right\Vert},\end{aligned}$$ and this is trivial to sum.
The low frequency part {#R}
----------------------
Here we prove for $J^{\boldsymbol\Sigma}_{{| \xi_0 |} < 1}$, without any dyadic decomposition. For this, we need the estimate, for $f \in \mathcal S({\mathbb{R}}^3)$, $$\label{R:2}
{\left\Vert {\mathbf{P}}_{{| \xi |} < 1} f \right\Vert}_{L^\infty} \le {| B(0,1) |}^{1/2} {\left\Vert f \right\Vert}_{L^2},$$ or rather its dual, $$\label{R:4}
{\left\Vert {\mathbf{P}}_{{| \xi |} < 1} f \right\Vert}_{L^2} \le {| B(0,1) |}^{1/2} {\left\Vert f \right\Vert}_{L^1}.$$ Here $B(0,1)$ denotes the unit ball $\{ \xi \in {\mathbb{R}}^3 \colon {| \xi |} < 1 \}$. Note that follows from the Riemann-Lebesgue lemma and the Cauchy-Schwarz inequality.
We also need (this follows from the triangle inequality in Fourier space) $$\label{R:5}
{\left\Vert {\mathbf{P}}_{{| \xi |} < 1} (fg) \right\Vert} \le C_s {\left\Vert P_{{| \xi |} < 1}({\langle D \rangle}^s {\lfloor f \rfloor} \cdot {\langle D \rangle}^{-s} {\lfloor g \rfloor}) \right\Vert},$$ where we use the notation ${\lfloor f \rfloor} = \mathcal F_x^{-1}{| \widehat f\, |}$. Furthermore, we need the crude estimate $$\label{R:6}
{\bigl\Vert \rho{\square}^{-1}F \bigr\Vert} \lesssim {\left\Vert F \right\Vert},$$ which follows from a cut-off estimate proved in [@Klainerman:1995b], and we need $$\label{R:8}
{\left\Vert F \right\Vert}_{L_t^pL_x^2} \le C_{p,b} {\left\Vert F \right\Vert}_{X_\pm^{0,(1-2/p)b}} \qquad (2 \le p \le \infty, \; b > 1/2).$$ The latter is trivial for $p=2$, so by interpolation it suffices to prove it for $p=\infty$, but then by Minkowski’s integral inequality, the Riemann-Lebesgue lemma and Plancherel’s theorem, the left hand side is bounded by ${\Vert \widetilde F(\tau,\xi) \Vert}_{L_\xi^2L_\tau^1}$. Inserting ${\langle \tau\pm{| \xi |} \rangle}^b{\langle \tau\pm{| \xi |} \rangle}^{-b}$ and applying the Cauchy-Schwarz inequality in $\tau$, one easily obtains the desired estimate.
Now we estimate, for any $b > 1/2$, $$\begin{aligned}
J^{\boldsymbol\Sigma}_{{| \xi_0 |} < 1} &\le {\bigl\Vert \rho\square^{-1} {\mathbf{P}}_{{| \xi |} < 1}{\langle \, \boldsymbol\alpha^\mu\mathbf\Pi_{\pm_1}\psi_1 , \mathbf\Pi_{\pm_2}\psi_2 \, \rangle} \bigr\Vert} {\left\Vert {\mathbf{P}}_{{| \xi |} < 1} {\langle \, \boldsymbol\alpha_\mu\mathbf\Pi_{\pm_3}\psi_3 , \mathbf\Pi_{\pm_4}\psi_4 \, \rangle} \right\Vert}
\\
&\lesssim {\left\Vert {\mathbf{P}}_{{| \xi |} < 1}{\langle \, \boldsymbol\alpha^\mu\mathbf\Pi_{\pm_1}\psi_1 , \mathbf\Pi_{\pm_2}\psi_2 \, \rangle} \right\Vert} {\left\Vert {\mathbf{P}}_{{| \xi |} < 1} {\langle \, \boldsymbol\alpha_\mu\mathbf\Pi_{\pm_3}\psi_3 , \mathbf\Pi_{\pm_4}\psi_4 \, \rangle} \right\Vert}
\\
&\le C_s {\left\Vert {\mathbf{P}}_{{| \xi |} < 1}({\lfloor \psi_1 \rfloor}{\lfloor \psi_2 \rfloor}) \right\Vert} {\left\Vert {\mathbf{P}}_{{| \xi |} < 1} ({\langle D \rangle}^s{\lfloor \psi_3 \rfloor}\cdot{\langle D \rangle}^{-s}{\lfloor \psi_4 \rfloor}) \right\Vert}
\\
&\le C_s {\left\Vert {\lfloor \psi_1 \rfloor}{\lfloor \psi_2 \rfloor} \right\Vert}_{L_t^2L_x^1} {\left\Vert {\langle D \rangle}^s{\lfloor \psi_3 \rfloor}\cdot{\langle D \rangle}^{-s}{\lfloor \psi_4 \rfloor} \right\Vert}_{L_t^2L_x^1}
\\
&\le C_s {\left\Vert \psi_1 \right\Vert}_{L_t^4L_x^2} {\left\Vert \psi_2 \right\Vert}_{L_t^4L_x^2} {\left\Vert {\langle D \rangle}^s\psi_3 \right\Vert}_{L_t^4L_x^2} {\left\Vert {\langle D \rangle}^{-s}\psi_4 \right\Vert}_{L_t^4L_x^2}
\\
&\le C_{s,b} {\left\Vert \psi_1 \right\Vert}_{X_{\pm_1}^{0,b/2}} {\left\Vert \psi_2 \right\Vert}_{X_{\pm_2}^{0,b/2}} {\left\Vert \psi_3 \right\Vert}_{X_{\pm_3}^{s,b/2}} {\left\Vert \psi_4 \right\Vert}_{X_{\pm_4}^{-s,b/2}},\end{aligned}$$ where to get the second inequality we used , and then we used , , Hölder’s inequality and . Finally, if we write $b=1/2+\varepsilon$, where $\varepsilon > 0$, then we see that $b/2 \le 1/2-2\varepsilon$ for all $\varepsilon \le 1/10$, hence we have proved for the low frequency part. Notice that we did not need the null structure.
This concludes the proof of .
Proof of the trilinear estimate {#S}
===============================
Here we prove . Recall that $A_\mu^{\mathrm{hom.}}$ is the solution of , the data being determined by , , with regularity as in . Thus, $$A_0^{\mathrm{hom.}} = 0,$$ whereas $A_j^{\mathrm{hom.}}$ for $j=1,2,3$ splits in the usual way: $$A_j^{\mathrm{hom.}} = A_{j,+}^{\mathrm{hom.}} + A_{j,-}^{\mathrm{hom.}},$$ where $$\widetilde{A_{j,\pm_0}^{\mathrm{hom.}}}(X_0)
=
\delta(\tau_0\pm_0{| \xi_0 |})
\frac{g_j^{\pm_0}(\xi_0)}{{| \xi_0 |}{\langle \xi_0 \rangle}^{s-1/2}}
\qquad (X_0 = (\tau_0,\xi_0)),$$ and $g_j^+,g_j^- \in L^2({\mathbb{R}}^3)$ are defined by $$\label{S:4}
g_j^{\pm_0}(\xi_0)
=
{| \xi_0 |}{\langle \xi_0 \rangle}^{s-1/2}
\left( \frac{\widehat{a_j}(\xi_0)}{2} \pm_0 \frac{\widehat{\dot a_j}(\xi_0)}{2i{| \xi_0 |}} \right),$$ hence $$\label{S:6}
{\left\Vert g^\pm \right\Vert} \lesssim \mathcal I_0,$$ where $\mathcal I_0$ is as in . By –, $\nabla \cdot \mathbf a = 0$ and $\nabla \cdot \dot{\mathbf a} = - {| \psi_0 |}^2$, hence $$\label{S:10}
\xi_0^j g_j^{\pm_0}(\xi_0) \simeq {\langle \xi_0 \rangle}^{s-1/2} \widehat{{| \psi_0 |}^2}(\xi_0),$$ where we implicitly sum over $j=1,2,3$ on the left hand side.
Now write $\boldsymbol\Sigma = (\pm_0,\pm_1,\pm_2)$, and let $I^{\boldsymbol\Sigma}$ be defined like $I^{\pm_1,\pm_2}$ in , except that $A_j^{\mathrm{hom.}}$ is replaced by $A_{j,\pm_0}^{\mathrm{hom.}}$. By Plancherel’s formula, $$I^{\boldsymbol\Sigma}
\simeq
\iint
\widehat\rho(\tau_0\pm_0{| \xi_0 |})
\frac{\sigma^j(X_1,X_2) g_j^{\pm_0}(\xi_0)F_1(X_1)
F_2(X_2)}
{{| \xi_0 |}{\langle \xi_0 \rangle}^{s-1/2}\prod_{k=1}^2{\langle \xi_k \rangle}^{s_k} {\langle \tau_k\pm_k{| \xi_k |} \rangle}^{b_k}}
\, d\mu^{21}_{X_0}
\, dX_0,$$ where $$\begin{gathered}
s_1=-s_2=s,
\qquad
b_1 = 1/2+\varepsilon,
\qquad
b_2 = 1/2-2\varepsilon,
\\
\widetilde{\psi_k}
= z_k {\bigl\vert \widetilde{\psi_k} \bigr\vert},
\qquad
{\bigl\vert \widetilde{\psi_k}(X_k) \bigr\vert}
= \frac{F_k(X_k)}{{\langle \xi_k \rangle}^{s_k} {\langle \tau_k\pm_k{| \xi_k |} \rangle}^{b_k}},
\qquad X_k = (\tau_k,\xi_k),
\\
\sigma^j(X_1,X_2) = {\langle \, \boldsymbol\alpha^j\mathbf\Pi(e_1)z_1(X_1) , \mathbf\Pi(e_2)z_2(X_2) \, \rangle},
\qquad
e_k = \pm_k \frac{\xi_k}{{| \xi_k |}} \in \mathbb S^2.\end{gathered}$$ Here $z_k : {\mathbb{R}}^{1+3} \to {\mathbb{C}}^4$ is measurable, ${| z_k |}=1$, $F_k \in L^2({\mathbb{R}}^{1+3)}$ and $F_k \ge 0$, for $k=1,2$. The convolution measure $ d\mu^{21}_{X_0}$ is given by the rule in , hence $$X_0 = X_2 - X_1 \qquad \left( \iff \tau_0=\tau_2-\tau_1, \quad \xi_0=\xi_2-\xi_1 \right)$$ in the above integral. We also define the angles $\theta_{01}$, $\theta_{02}$ and $\theta_{12}$ by .
We want to prove the estimate $$\label{S:28}
{\left\vert I^{\boldsymbol\Sigma}_{{| \xi_0 |} \ge 1} \right\vert} \lesssim C(\mathcal I_0) {\left\Vert F_1 \right\Vert}{\left\Vert F_2 \right\Vert}.$$ Split $$I^{\boldsymbol\Sigma} = I^{\boldsymbol\Sigma}_{{| \xi_0 |} \ge 1} + I^{\boldsymbol\Sigma}_{{| \xi_0 |} < 1}$$ corresponding to the regions ${| \xi_0 |} \ge 1$ and ${| \xi_0 |} < 1$.
Estimate for $I^{\pm_1,\pm_2}_{{| \xi_0 |} \ge 1}$
--------------------------------------------------
Let $N_0,N_1,N_2,L_0,L_1,L_2 \ge 1$ be dyadic numbers representing the sizes of the weights, as in section \[C\]. Taking the absolute value and using the fact that $\widehat\rho$ is a Schwartz function (hence we can get as many powers as we like of $L_0$ in the denominator), we get $$\label{S:20}
{\left\vert I^{\boldsymbol\Sigma}_{{| \xi_0 |} \ge 1} \right\vert}
\lesssim
\sum_{{\boldsymbol{N}},{\boldsymbol{L}}} \frac{N_2^s I^{\boldsymbol\Sigma}_{{\boldsymbol{N}},{\boldsymbol{L}}}}{N_0^{s+1/2} N_1^s L_0 L_1^{1/2+\varepsilon} L_2^{1/2-2\varepsilon}},$$ where ${\boldsymbol{N}}= (N_0,N_1,N_2)$, ${\boldsymbol{L}}= (L_0,L_1,L_2)$ and, with notation as in , $$\begin{gathered}
I^{\boldsymbol\Sigma}_{{\boldsymbol{N}},{\boldsymbol{L}}}
=
\iint
{| \sigma_{12}^j(X_1,X_2) |}
\widetilde{u_{0,j}}(X_0)
\widetilde{u_1}(X_1)
\widetilde{u_2}(X_2)
\, d\mu^{21}_{X_0}
\, dX_0,
\\
\widetilde{u_{0,j}} = \chi_{K^{\pm_0}_{N_0,L_0}} F_{0,j},
\qquad
F_{0,j}(X_0)
=
\frac{{| g_j^{\pm_0}(\xi_0) |}}{{\langle \tau_0\pm_0{| \xi_0 |} \rangle}}
\qquad (j=1,2,3).\end{gathered}$$ The sum over ${\boldsymbol{N}}$ is restricted by .
By the same type of summation argument that was used in section \[Y\] (we omit the details), is easily deduced from if we can prove the following: $$\begin{aligned}
\label{S:30}
I^{\boldsymbol\Sigma}_{{\boldsymbol{N}},{\boldsymbol{L}}}
&\lesssim \left(N_0L_0^2L_1L_2\right)^{1/2}
C(\mathcal I_0) {\left\Vert u_1 \right\Vert}{\left\Vert u_2 \right\Vert},
\\
\label{S:32}
I^{\boldsymbol\Sigma}_{{\boldsymbol{N}},{\boldsymbol{L}}}
&\lesssim \left(
\left({N_{\mathrm{min}}}^{012}\right)^3 {L_{\mathrm{min}}}^{012}\right)^{1/2}
C(\mathcal I_0) {\left\Vert u_1 \right\Vert}{\left\Vert u_2 \right\Vert}.\end{aligned}$$
First, follows from , if we estimate ${| \sigma^j(X_1,X_2) |} \lesssim 1$ and use the fact that ${\left\Vert u_0 \right\Vert} \lesssim \mathcal I_0$, by .
To prove , on the other hand, we need to use the structure of the symbol $\sigma_{12}^j$, encoded in the identity . We claim that $$\begin{gathered}
\label{S:40}
{\left\vert \sigma^j(X_1,X_2) g_j^{\pm_0}(\xi_0) \right\vert}
\lesssim \theta_{12} {| g^{\pm_0}(\xi_0) |}
+ \min(\theta_{01},\theta_{02}) {| g^{\pm_0}(\xi_0) |}
\\
+
{\langle \xi_0 \rangle}^{s-3/2} {\left\vert \widehat{{| \psi_0 |}^2}(\xi_0) \right\vert},\end{gathered}$$ where again we sum over $j=1,2,3$, and we assume ${| \xi_0 |} \ge 1$.
To prove , we use the identity to write $$\begin{aligned}
\sigma^j(X_1,X_2) g_j^{\pm_0}(\xi_0)
&=
{\langle \, \boldsymbol\alpha^j\mathbf\Pi(e_1)z_1 , \mathbf\Pi(e_2)z_2 \, \rangle}
g_j^{\pm_0}(\xi_0)
\\
&=
{\langle \, \mathbf\Pi(e_2)\mathbf\Pi(-e_1)\boldsymbol\alpha^j z_1 , z_2 \, \rangle}
g_j^{\pm_0}(\xi_0)
+ {\langle \, z_1 , \mathbf\Pi(e_2)z_2 \, \rangle} e_1^j g_j^{\pm_0}(\xi_0)
\\
&=
{\langle \, \mathbf\Pi(e_2)\mathbf\Pi(-e_1)\boldsymbol\alpha^j z_1 , z_2 \, \rangle}
g_j^{\pm_0}(\xi_0)
\\
& \;\; + {\langle \, z_1 , \mathbf\Pi(e_2)z_2 \, \rangle} (e_1^j-e_0^j) g_j^{\pm_0}(\xi_0)
+ {\langle \, z_1 , \mathbf\Pi(e_2)z_2 \, \rangle} e_0^j g_j^{\pm_0}(\xi_0),\end{aligned}$$ and by and this implies with $\min(\theta_{01},\theta_{02})$ replaced by $\theta_{01}$. But since $\boldsymbol\alpha^j$ is self-adjoint, we can also move it onto the second factor in the inner product defining $\sigma^j$, and then we get instead the angle $\theta_{02}$. This proves .
Corresponding to the first and second terms in the right side of , we need to prove for the integrals $$\begin{aligned}
I^{\boldsymbol\Sigma,1}_{{\boldsymbol{N}},{\boldsymbol{L}}}
&=
\iint
\theta_{12}
\widetilde{u_0}(X_0)
\widetilde{u_1}(X_1)
\widetilde{u_2}(X_2)
\, d\mu^{21}_{X_0}
\, dX_0,
\\
I^{\boldsymbol\Sigma,2}_{{\boldsymbol{N}},{\boldsymbol{L}}}
&=
\iint
\min(\theta_{01},\theta_{02})
\widetilde{u_0}(X_0)
\widetilde{u_1}(X_1)
\widetilde{u_2}(X_2)
\, d\mu^{21}_{X_0}
\, dX_0,\end{aligned}$$ where now $$F_0(X_0) = \frac{{| g^{\pm_0}(\xi_0) |}}
{{\langle \tau_0\pm_0{| \xi_0 |} \rangle}},$$ hence ${\left\Vert u_0 \right\Vert} \simeq {\left\Vert F_0 \right\Vert} \lesssim \mathcal I_0$, by .
For $I^{\boldsymbol\Sigma,1}_{{\boldsymbol{N}},{\boldsymbol{L}}}$, we get (with only one power of $L_0$ inside the parentheses) from the null form estimate in Theorem \[D:Thm\].
Now consider $I^{\boldsymbol\Sigma,2}_{{\boldsymbol{N}},{\boldsymbol{L}}}$. By Lemma \[F:Lemma4\], $$\min(\theta_{01},\theta_{02})
\lesssim \left( \frac{{L_{\mathrm{max}}}^{012}}{N_0} \right)^{1/2},$$ so by the Cauchy-Schwarz inequality, $$\begin{aligned}
I^{\boldsymbol\Sigma,2}_{{\boldsymbol{N}},{\boldsymbol{L}}}
&\lesssim
\left( \frac{{L_{\mathrm{max}}}^{012}}{N_0} \right)^{1/2}
{\left\Vert u_0 \right\Vert} {\bigl\Vert {\mathbf{P}}_{K^{\pm_0}_{N_0,L_0}} \left( u_1 \overline{u_2} \right) \bigr\Vert}
\\
&\lesssim
\left( \frac{{L_{\mathrm{max}}}^{012}}{N_0} N_0^2 L_0 {L_{\mathrm{min}}}^{12} \right)^{1/2}
\mathcal I_0 {\left\Vert u_1 \right\Vert}{\left\Vert u_2 \right\Vert},\end{aligned}$$ where we used and ${\left\Vert u_0 \right\Vert} \lesssim \mathcal I_0$. This proves for $I^{\boldsymbol\Sigma,2}_{{\boldsymbol{N}},{\boldsymbol{L}}}$. Note that here we may actually pick up two powers of $L_0$ inside the parentheses (recall that this is allowed because $\widehat \rho$ is rapidly decreasing).
Now consider the case where the third term in the right side of dominates. We may assume $\theta_{12} \ll 1$, since otherwise we can reduce to $I^{\boldsymbol\Sigma,1}_{{\boldsymbol{N}},{\boldsymbol{L}}}$ by estimating ${| \sigma^j(X_1,X_2) |} \lesssim 1$. So by Lemma \[D:Lemma1\], $$\theta_{12} \lesssim \gamma \equiv \left( \frac{N_0{L_{\mathrm{max}}}^{012}}{N_1N_2} \right)^{1/2},$$ hence we need to prove for $$I^{\boldsymbol\Sigma,3}_{{\boldsymbol{N}},{\boldsymbol{L}}}
= N_0^{s-3/2}
\iint
\chi_{\theta_{12} \lesssim \gamma}
\widetilde{u_0}(X_0)
\widetilde{u_1}(X_1)
\widetilde{u_2}(X_2)
\, d\mu^{21}_{X_0}
\, dX_0,$$ where now $$F_0(\tau_0,\xi_0)
=
\frac{{\left\vert \widehat{{| \psi_0 |}^2}(\xi_0) \right\vert}}{{\langle \tau_0\pm_0{| \xi_0 |} \rangle}}.$$ By Lemma \[B:Lemma5\] applied to the pair $(\pm_1\xi_1,\pm_2\xi_2)$, $$\label{S:60}
I^{\boldsymbol\Sigma,3}_{{\boldsymbol{N}},{\boldsymbol{L}}}
\lesssim N_0^{s-3/2}
\sum_{\omega_1,\omega_2}
\iint
\widetilde{u_0}(X_0)
\widetilde{u_1^{\gamma,\omega_1}}(X_1)
\widetilde{u_2^{\gamma,\omega_2}}(X_2)
\, d\mu^{21}_{X_0}
\, dX_0,$$ where the sum is over $\omega_1,\omega_2 \in \Omega(\gamma)$ with $\theta(\omega_1,\omega_2) \lesssim \gamma$. Thus, $\xi_1,\xi_2$ are both restricted to a tube of radius $$r \sim {N_{\mathrm{max}}}^{12}\gamma \sim \left( \frac{{N_{\mathrm{max}}}^{12}N_0{L_{\mathrm{max}}}^{012}}{{N_{\mathrm{min}}}^{12}} \right)^{1/2}$$ around ${\mathbb{R}}\omega_1$, hence the same is true of $\xi_0=\xi_2-\xi_1$, so we get $$\begin{aligned}
I^{\boldsymbol\Sigma,3}_{{\boldsymbol{N}},{\boldsymbol{L}}}
&\lesssim N_0^{s-3/2}
\sum_{\omega_1,\omega_2}
{\left\Vert {\mathbf{P}}_{{\mathbb{R}}\times T_r(\omega_1)} u_0 \right\Vert}
{\left\Vert {\mathbf{P}}_{K^{\pm_0}_{N_0,L_0}} \left( u_1^{\gamma,\omega_1}
\overline{u_2^{\gamma,\omega_2}} \right) \right\Vert}
\\
&\lesssim
N_0^{s-3/2}
\sum_{\omega_1,\omega_2}
{\left\Vert {\mathbf{P}}_{T_r(\omega_1)}
{\mathbf{P}}_{{\langle \xi_0 \rangle} \sim N_0} {| \psi_0 |}^2 \right\Vert}
\left( N_0^2 L_0 {L_{\mathrm{min}}}^{12} \right)^{1/2} {\left\Vert u_1^{\gamma,\omega_1} \right\Vert}
{\left\Vert u_2^{\gamma,\omega_2} \right\Vert}.\end{aligned}$$ where we used . Applying the estimate (proved below) $$\label{S:56}
\sup_{\omega \in \mathbb S^2} {\left\Vert {\mathbf{P}}_{T_r(\omega)}
{\mathbf{P}}_{{\langle \xi_0 \rangle} \sim N_0} {| \psi_0 |}^2 \right\Vert}
\lesssim
\left( r^2 N_0 \right)^{1/2} N_0^{-s} {\left\Vert \psi_0 \right\Vert}_{H^s}^2,$$ and summing $\omega_1,\omega_2$ as in , we then obtain $$\begin{aligned}
I^{\boldsymbol\Sigma,3}_{{\boldsymbol{N}},{\boldsymbol{L}}}
&\lesssim N_0^{s-3/2}
\left( r^2 N_0 \right)^{1/2} N_0^{-s} {\left\Vert \psi_0 \right\Vert}_{H^s}^2
\left( N_0 {N_{\mathrm{min}}}^{012} L_0 {L_{\mathrm{min}}}^{12} \right)^{1/2}
{\left\Vert u_1 \right\Vert}{\left\Vert u_2 \right\Vert}
\\
&\sim
N_0^{-3/2}
\left( \frac{{N_{\mathrm{max}}}^{12}N_0{L_{\mathrm{max}}}^{012}}{{N_{\mathrm{min}}}^{12}} N_0 \right)^{1/2} {\left\Vert \psi_0 \right\Vert}_{H^s}^2
\left( N_0 {N_{\mathrm{min}}}^{012} L_0 {L_{\mathrm{min}}}^{12} \right)^{1/2}
{\left\Vert u_1 \right\Vert}{\left\Vert u_2 \right\Vert}
\\
&\sim
\left( \frac{{N_{\mathrm{max}}}^{12}{N_{\mathrm{min}}}^{012}}{{N_{\mathrm{min}}}^{12}} L_0 {L_{\mathrm{min}}}^{12} {L_{\mathrm{max}}}^{012} \right)^{1/2} {\left\Vert \psi_0 \right\Vert}_{H^s}^2
{\left\Vert u_1 \right\Vert}{\left\Vert u_2 \right\Vert}
\\
&\lesssim
N_0^{-3/2}
\left( N_0 L_0 {L_{\mathrm{min}}}^{12} {L_{\mathrm{max}}}^{012} \right)^{1/2} {\left\Vert \psi_0 \right\Vert}_{H^s}^2
{\left\Vert u_1 \right\Vert}{\left\Vert u_2 \right\Vert},\end{aligned}$$ where we used in the last step. This proves for $I^{\boldsymbol\Sigma,3}_{{\boldsymbol{N}},{\boldsymbol{L}}}$, under the assumption that holds.
In fact, is an easy consequence of the estimate $$\label{S:66}
\sup_{\omega \in \mathbb S^2} {\left\Vert {\mathbf{P}}_{T_r(\omega)}
{\mathbf{P}}_{{\langle \xi_0 \rangle} \sim N_0}(fg) \right\Vert}
\lesssim
\left( r^2 N_0 \right)^{1/2} {\left\Vert f \right\Vert} {\left\Vert g \right\Vert}
\qquad \left(\forall f,g \in \mathcal S({\mathbb{R}}^3)\right)$$ which reduces, by an argument based on the Cauchy-Schwarz inequality (see [@Tao:2001] or [@Selberg:2008a]), to the fact that the volume of the $\xi_0$-support is $O(r^2N_0)$.
This completes the proof of .
Estimate for $I^{\pm_1,\pm_2}_{{| \xi_0 |} < 1}$
------------------------------------------------
Since ${\langle \xi_2 \rangle} \lesssim {\langle \xi_0 \rangle} + {\langle \xi_1 \rangle} \lesssim {\langle \xi_1 \rangle}$, and since $\widehat \rho$ is rapidly decreasing, $$\begin{aligned}
{\left\vert I^{\boldsymbol\Sigma}_{{| \xi_0 |} < 1} \right\vert}
&\lesssim
\iint
\frac{\chi_{{| \xi_0 |} < 1} {| g^{\pm_0}(\xi_0) |}}{{| \xi_0 |}{\langle \tau_0\pm_0{| \xi_0 |} \rangle}^2}
F_1(X_1) F_2(X_2)
\, d\mu^{21}_{X_0}
\, dX_0
\\
&=
\iint
\frac{\chi_{{| \xi_0 |} < 1} {| g^{\pm_0}(\xi_0) |}}{{| \xi_0 |}{\langle \tau_0\pm_0{| \xi_0 |} \rangle}^2}
F_1(X_1) F_2(X_0+X_1)
\, dX_1
\, dX_0
\\
&\le
\int
\frac{\chi_{{| \xi_0 |} < 1} {| g^{\pm_0}(\xi_0) |}}{{| \xi_0 |}{\langle \tau_0\pm_0{| \xi_0 |} \rangle}^2}
\, dX_0
{\left\Vert F_1 \right\Vert}{\left\Vert F_2 \right\Vert}
\\
&\lesssim
\int
\frac{\chi_{{| \xi_0 |} < 1} {| g^{\pm_0}(\xi_0) |}}{{| \xi_0 |}}
\, d\xi_0
{\left\Vert F_1 \right\Vert}{\left\Vert F_2 \right\Vert}
\\
&\le
\left(\int
\frac{\chi_{{| \xi_0 |} < 1}}{{| \xi_0 |}^2}
\, d\xi_0 \right)^{1/2}
{\left\Vert g^{\pm_0} \right\Vert}
{\left\Vert F_1 \right\Vert}{\left\Vert F_2 \right\Vert},\end{aligned}$$ proving .
This completes the proof of .
Estimates for the electromagnetic field {#N}
=======================================
Here we prove Theorem \[A:Thm2\]. Since Maxwell’s equations are linear, uniqueness is trivial, so we only need to construct the solution. Let us define $(\mathbf E,\mathbf B)$ by . Since we know that $\square A_\mu = - J_\mu$ and that $A_\mu$ satisfies the Lorenz gauge condition, a direct calculation shows that is satisfied, so it only remains to prove $${\left\Vert \mathbf E(t) \right\Vert}_{H^{s-1/2}}
+
{\left\Vert \mathbf B(t) \right\Vert}_{H^{s-1/2}}
\le C
\qquad (\forall t \in [-T,T]),$$ where $C$ depends on the data norm $\mathcal I_0$ defined by . But the components of $\mathbf E, \mathbf B$ are just the nonzero components of the electromagnetic tensor $$F_{\kappa\lambda} = \partial_\kappa A_\lambda - \partial_\lambda A_\kappa,$$ so we need to prove $$\label{N:10}
{\left\Vert F_{\kappa\lambda}(t) \right\Vert}_{H^{s-1/2}}
\le C \qquad (\forall t \in [-T,T]).$$ Of course, it suffices to consider indices $(\kappa,\lambda) = (k,l), (k,0)$, where $k,l=1,2,3$.
Since $\square A_\kappa = - J_\kappa$, $$\square F_{\kappa\lambda} = -\partial_\kappa J_\lambda + \partial_\lambda J_\kappa.$$ Split $$F_{\kappa\lambda} = F_{\kappa\lambda}^{\text{hom.}} + F_{\kappa\lambda}^{\text{inh.}},$$ where $\square F_{\kappa\lambda}^{\text{hom.}} = 0$ with the initial data determined by $(\mathbf E_0,\mathbf B_0)$, and $$\label{N:10:2}
F_{\kappa\lambda}^{\text{inh.}}
=
\square^{-1}
\left( -\partial_\kappa J_\lambda + \partial_\lambda J_\kappa \right).$$
For the homogeneous part $F_{\kappa\lambda}^{\text{hom.}}$, holds by the energy inequality for the wave equation and the assumption $\mathbf E_0,\mathbf B_0 \in H^{s-1/2}$.
It remains to prove for $F_{\kappa\lambda}^{\text{inh.}}$. Splitting $\psi = \psi_+ + \psi_-$ we see from and that $${\left\vert \widetilde{J_\kappa}(X_0) \right\vert}
\le
\sum_{\pm_1,\pm_2}
\int
\frac{{\left\langle \, \boldsymbol\alpha_\kappa\boldsymbol\Pi(e_1)z_1 , \boldsymbol\Pi(e_2)z_2 \, \right\rangle} G_1(X_1) G_2(X_2)}
{{\langle \xi_1 \rangle}^s {\langle \xi_2 \rangle}^s {\langle \tau_1\pm_1{| \xi_1 |} \rangle}^{1/2+\varepsilon} {\langle \tau_2\pm_2{| \xi_2 |} \rangle}^{1/2+\varepsilon}}
\, d\mu^{12}_{X_0},$$ where $e_j = \pm_j \xi_j/{| \xi_j |}$, $z_j : {\mathbb{R}}^{1+3} \to {\mathbb{C}}^4$ is measurable, ${| z_j |}=1$, $G_j \in L^2({\mathbb{R}}^{1+3)}$ and $G_j \ge 0$, for $j=1,2$. Now observe that the symbol of $(1/i)\partial_\kappa$ is $$X_0^\kappa = \begin{cases} \tau_0 \quad &\text{for $\kappa=0$},
\\
\xi_0^\kappa \quad &\text{for $\kappa=1,2,3$},
\end{cases}$$ recalling that $\xi_0=(\xi_0^1,\xi_0^2,\xi_0^3)$. Thus, applying Lemma \[C:Lemma1\] to , and writing $\boldsymbol\Sigma = (\pm_0,\pm_1,\pm_2)$, we have $$\label{N:10:4}
{\left\Vert F_{\kappa\lambda}^{\text{inh.}}(t) \right\Vert}_{H^{s-1/2}}
\lesssim
\sum_{\boldsymbol\Sigma} I^{\boldsymbol\Sigma}
\qquad (\forall t \in [-T,T]),$$ where $$\begin{gathered}
I^{\boldsymbol\Sigma}
=
{\left\Vert \iint \frac{\sigma_{\kappa\lambda}(X_1,X_2) G_1(X_1) G_2(X_2)}
{{| \xi_0 |}{\langle \xi_0 \rangle}^{1/2-s} {\langle \tau_0\pm_0{| \xi_0 |} \rangle}
\prod_{j=1}^2{\langle \xi_j \rangle}^s{\langle \tau_j\pm_j{| \xi_j |} \rangle}^{1/2+\varepsilon}}
\, d\mu^{12}_{X_0} \, d\tau_0 \right\Vert}_{L^2_{\xi_0}},
\\
\sigma_{\kappa\lambda}(X_1,X_2)
= X_0^\kappa {\left\langle \, \boldsymbol\alpha_\lambda\boldsymbol\Pi(e_1)z_1 , \boldsymbol\Pi(e_2)z_2 \, \right\rangle}
-
X_0^\lambda {\left\langle \, \boldsymbol\alpha_\kappa\boldsymbol\Pi(e_1)z_1 , \boldsymbol\Pi(e_2)z_2 \, \right\rangle}.\end{gathered}$$
Define $\theta_{12}, \theta_{01}, \theta_{02}$ as in . We have the following null structure:
\[N:Lemma\] With notation as above, $$\begin{aligned}
\label{N:12}
\frac{{\left\vert \sigma_{kl}(X_1,X_2) \right\vert}}{{| \xi_0 |}}
&\lesssim
\theta_{12} + \min(\theta_{01},\theta_{02}),
\\
\label{N:14}
\frac{{\left\vert \sigma_{k0}(X_1,X_2) \right\vert}}{{| \xi_0 |}}
&\lesssim
\theta_{12} + \min(\theta_{01},\theta_{02})
+ \frac{{\bigl\vert \tau_0\pm_0{| \xi_0 |} \bigr\vert}}{{| \xi_0 |}},\end{aligned}$$ for $k,l=1,2,3$.
By the rule for raising or lowering indices, $\boldsymbol\alpha_k=\boldsymbol\alpha^k$ for $k=1,2,3$, whereas $\boldsymbol\alpha_0=-\boldsymbol\alpha^0 = \mathbf I_{4 \times 4}$. Thus, $$\begin{aligned}
\frac{\sigma_{kl}(X_1,X_2)}{{| \xi_0 |}} &= \frac{\xi_0^k}{{| \xi_0 |}} {\left\langle \, \boldsymbol\alpha^l\boldsymbol\Pi(e_1)z_1 , \boldsymbol\Pi(e_2)z_2 \, \right\rangle}
-
\frac{\xi_0^l}{{| \xi_0 |}} {\left\langle \, \boldsymbol\alpha^k\boldsymbol\Pi(e_1)z_1 , \boldsymbol\Pi(e_2)z_2 \, \right\rangle},
\\
- \frac{\sigma_{k0}(X_1,X_2)}{{| \xi_0 |}} &= \frac{\xi_0^k}{{| \xi_0 |}} {\left\langle \, \boldsymbol\Pi(e_1)z_1 , \boldsymbol\Pi(e_2)z_2 \, \right\rangle}
+
\frac{\tau_0}{{| \xi_0 |}} {\left\langle \, \boldsymbol\alpha^k\boldsymbol\Pi(e_1)z_1 , \boldsymbol\Pi(e_2)z_2 \, \right\rangle}.\end{aligned}$$
Then by the commutation identity we see that $$\frac{{\left\vert \sigma_{kl}(X_1,X_2) \right\vert}}{{| \xi_0 |}}
\lesssim \theta_{12}
+ {| e_0^k e_1^l - e_0^l e_1^k |}
\lesssim \theta_{12} + \theta_{01},$$ but since the $\boldsymbol\alpha^k$ are self-adjoint, we get also the same estimate with $\theta_{02}$ instead of $\theta_{01}$, proving .
Similarly we find that $$\frac{{\left\vert \sigma_{k0}(X_1,X_2) \right\vert}}{{| \xi_0 |}}
\lesssim \theta_{12}
+ {| e_0^k - e_1^k |}
+ \frac{{\bigl\vert \tau_0\pm_0{| \xi_0 |} \bigr\vert}}{{| \xi_0 |}}
\lesssim \theta_{12} + \theta_{01} + \frac{{\bigl\vert \tau_0\pm_0{| \xi_0 |} \bigr\vert}}{{| \xi_0 |}},$$ but again, by the self-adjointness of the $\boldsymbol\alpha^k$, we can also get $\theta_{02}$ instead of $\theta_{01}$, proving .
In view of and Lemma \[N:Lemma\], we reduce to proving (dropping the superscript $\boldsymbol\Sigma$ for simplicity) $$\label{N:20}
I_1, I_2, I_3 \lesssim {\left\Vert F_1 \right\Vert}{\left\Vert F_2 \right\Vert},$$ where $$\begin{aligned}
I_1 &=
{\left\Vert \iint \frac{\theta_{12} G_1(X_1) G_2(X_2)}{{\langle \xi_0 \rangle}^{1/2-s} {\langle \tau_0\pm_0{| \xi_0 |} \rangle}
\prod_{j=1}^2{\langle \xi_j \rangle}^s{\langle \tau_j\pm_j{| \xi_j |} \rangle}^{1/2+\varepsilon}}
\, d\mu^{12}_{X_0} \, d\tau_0 \right\Vert}_{L^2_{\xi_0}},
\\
I_2 &=
{\left\Vert \iint \frac{\min(\theta_{01},\theta_{02}) G_1(X_1) G_2(X_2)}{{\langle \xi_0 \rangle}^{1/2-s} {\langle \tau_0\pm_0{| \xi_0 |} \rangle}
\prod_{j=1}^2{\langle \xi_j \rangle}^s{\langle \tau_j\pm_j{| \xi_j |} \rangle}^{1/2+\varepsilon}}
\, d\mu^{12}_{X_0} \, d\tau_0 \right\Vert}_{L^2_{\xi_0}},
\\
I_3 &=
{\left\Vert \iint \frac{G_1(X_1) G_2(X_2)}{{\langle \xi_0 \rangle}^{1/2-s}{| \xi_0 |}
\prod_{j=1}^2{\langle \xi_j \rangle}^s{\langle \tau_j\pm_j{| \xi_j |} \rangle}^{1/2+\varepsilon}}
\, d\mu^{12}_{X_0} \, d\tau_0 \right\Vert}_{L^2_{\xi_0}}.\end{aligned}$$
Estimate for $I_1$
------------------
By dyadic decomposition as in section \[C\], $$\label{N:30}
I_1 \lesssim
\sum_{{\boldsymbol{N}},{\boldsymbol{L}}}
\frac{J_{{\boldsymbol{N}},{\boldsymbol{L}}}}{N_0^{1/2-s}(N_1N_2)^sL_0(L_1L_2)^{1/2+\varepsilon}},$$ where ${\boldsymbol{N}}= (N_0,N_1,N_2)$, ${\boldsymbol{L}}= (L_0,L_1,L_2)$ and $$\label{N:31}
J_{{\boldsymbol{N}},{\boldsymbol{L}}}
=
{\left\Vert \iint \chi_{K^{\pm_0}_{N_0,L_0}}\!(X_0)
\, \theta_{12} \,
\widetilde{u_1}(X_1) \widetilde{u_2}(X_2)
\, d\mu^{12}_{X_0} \, d\tau_0 \right\Vert}_{L^2_{\xi_0}},$$ where we use the notation , but with $G$’s instead of $F$’s.
Applying the Cauchy-Schwarz inequality with respect to $\tau_0$, followed by either Theorem \[D:Thm\] or the trivial estimate (which holds without the angle), $$\begin{aligned}
\label{N:32}
J_{{\boldsymbol{N}},{\boldsymbol{L}}}
&\lesssim L_0^{1/2} \left(N_0 L_0 L_1 L_2\right)^{1/2}
{\left\Vert u_1 \right\Vert}{\left\Vert u_2 \right\Vert},
\\
\label{N:34}
J_{{\boldsymbol{N}},{\boldsymbol{L}}}
&\lesssim L_0^{1/2} \left(\left({N_{\mathrm{min}}}^{012}\right)^3 {L_{\mathrm{min}}}^{012}\right)^{1/2}
{\left\Vert u_1 \right\Vert}{\left\Vert u_2 \right\Vert},\end{aligned}$$ Take the former to the power $1-2\varepsilon$ and the latter to the power $2\varepsilon$. Then plugging the interpolated estimate into and summing by the same type of argument that was used in section \[Y\], we get for $I_1$ (for $\varepsilon > 0$ sufficiently small depending on $s > 0$).
Estimate for $I_2$
------------------
Again we dominate by a sum like , but now $\theta_{12}$ in is replaced by $\min(\theta_{01},\theta_{02})$. We need to prove . By Lemma \[F:Lemma4\], $$\label{N:40}
\min(\theta_{01},\theta_{02})
\lesssim \left( \frac{{L_{\mathrm{max}}}^{012}}{N_0} \right)^{1/2}.$$ By comparison, in the estimate for $I_1$ we used $\theta_{12} \lesssim ({L_{\mathrm{max}}}^{012}/{N_{\mathrm{min}}}^{12})^{1/2}$ to get (implicitly, since we used Theorem \[D:Thm\]). Thus, the analysis for $I_1$ applies also here if $N_0 \sim {N_{\mathrm{min}}}^{12}$, so we may assume $$\label{N:41}
N_0 \ll N_1 \sim N_2.$$ We distinguish the cases $L_0 \lesssim {L_{\mathrm{max}}}^{12}$ and $L_1,L_2 \ll L_0$.
### The case $L_0 \lesssim {L_{\mathrm{max}}}^{12}$
We apply Cauchy-Schwarz as we did for $I_1$ (hence we pick up a factor $L_0^{1/2}$), and use and the bilinear estimate , obtaining $$\begin{aligned}
J_{{\boldsymbol{N}},{\boldsymbol{L}}}
&\lesssim L_0^{1/2} \left( \frac{{L_{\mathrm{max}}}^{12}}{N_0} \right)^{1/2}
\left(N_0^2 L_0 {L_{\mathrm{min}}}^{12}\right)^{1/2}
{\left\Vert u_1 \right\Vert}{\left\Vert u_2 \right\Vert},\end{aligned}$$ proving .
### The case $L_1,L_2 \ll L_0$
Now we get into trouble if we try to follow the same approach as above, since we would need Theorem \[M:Thm\] to hold with $C^2 \sim N_0^2 L_1 L_2$, but this is not true in general. The problem is that we pick up too many powers of $L_0$. But instead of using $\tau_0\pm_0{| \xi_0 |}=O(L_0)$ when we apply the Cauchy-Schwarz inequality with respect to $\tau_0$, we can find another restriction on $\tau_0$ by decomposing into angular sectors based on the maximal size of $\theta_{12}$, as we now show.
We may assume $\theta_{12} \ll 1$, since otherwise we reduce to $I_1$. Therefore $$\label{N:42}
\theta_{12} \lesssim \gamma \equiv \biggl(\frac{N_0L_0}{N_1N_2}\biggr)^{1/2},$$ and now we apply Lemma \[D:Lemma5\] and , thus dominating $J_{{\boldsymbol{N}},{\boldsymbol{L}}}$ by $$\label{N:43}
\sum_{\omega_1,\omega_2}
\left( \frac{L_0}{N_0} \right)^{1/2}
{\left\Vert \iint \chi_{K^{\pm_0}_{N_0,L_0}}\!(X_0)
\,
\widetilde{u_1^{\gamma,\omega_1}}(X_1)
\widetilde{u_2^{\gamma,\omega_2}}(X_2)
\, d\mu^{12}_{X_0} \, d\tau_0 \right\Vert}_{L^2_{\xi_0}},$$ where the sum is over $\omega_1,\omega_2 \in \Omega(\gamma)$ with $\theta(\omega_1,\omega_2) \lesssim \gamma$. Then by , $$\label{N:44}
\tau_0 + \xi_0\cdot\omega_1 = O(d'),$$ where $$d' = \max\left( {L_{\mathrm{max}}}^{12}, {N_{\mathrm{max}}}^{12} \gamma^2 \right)
\sim \max\left( {L_{\mathrm{max}}}^{12}, \frac{N_0L_0}{N_1} \right),$$ recalling . So now if we apply the Cauchy-Schwarz inequality with respect to $\tau_0$ in using , we get, defining $u_{012}^{\gamma,\omega_1,\omega_2}$ as in , $$\begin{aligned}
\eqref{N:43}
&\lesssim
\sum_{\omega_1,\omega_2}
\left(d'\right)^{1/2}
\left( \frac{L_0}{N_0} \right)^{1/2}
{\left\Vert u_{012}^{\gamma,\omega_1,\omega_2} \right\Vert}
\\
&\lesssim
\left(\max\left( {L_{\mathrm{max}}}^{12}, \frac{N_0L_0}{N_1} \right)\right)^{1/2}
\left( \frac{L_0}{N_0} \right)^{1/2}
\\
&\quad
\times\min\left( N_0^2L_0{L_{\mathrm{min}}}^{12}, N_0N_1L_1L_2 \right)^{1/2}
\sum_{\omega_1,\omega_2}
{\left\Vert u_1^{\gamma,\omega_1} \right\Vert} {\left\Vert u_2^{\gamma,\omega_2} \right\Vert}
\\
&\lesssim
\left(N_0L_0^2L_1L_2\right)^{1/2}
{\left\Vert u_1 \right\Vert}{\left\Vert u_2 \right\Vert},\end{aligned}$$ where we used Theorem \[M:Thm\], and we summed $\omega_1,\omega_2$ as in .
This concludes the proof of for $I_2$.
Estimate for $I_3$
------------------
Here the $\tau$-integrations decouple, so $I_3$ can be written $$I_3 =
{\left\Vert \iint \frac{g_1(\xi_1) g_2(\xi_2)}{{\langle \xi_0 \rangle}^{1/2-s}{| \xi_0 |}
{\langle \xi_1 \rangle}^s{\langle \xi_2 \rangle}^s}
\, \delta(\xi_0-\xi_1+\xi_2) \,d\xi_1 \, d\xi_2 \right\Vert}_{L^2_{\xi_0}},$$ where $$g_j(\xi_j) = \int \frac{G(X_j)}{{\langle \tau_j\pm_j{| \xi_j |} \rangle}^{1/2+\varepsilon}} \, d\tau_j \qquad (j=1,2),$$ hence $${\left\Vert g_j \right\Vert} \le C_\varepsilon {\left\Vert G_j \right\Vert}.$$ Thus, it suffices to prove $$I_3 \lesssim {\left\Vert g_1 \right\Vert}{\left\Vert g_2 \right\Vert},$$ but this follows from the Sobolev product estimate (in physical space) $$\label{N:50}
{\left\Vert {| D |}^{-1}(f_1f_2) \right\Vert}_{H^{s-1/2}} \lesssim {\left\Vert f_1 \right\Vert} {\left\Vert f_2 \right\Vert}_{H^s}
\qquad (\forall f_1,f_2 \in \mathcal S({\mathbb{R}}^3)),$$ which we claim holds for any $s > 0$.
In fact, by Sobolev embedding and Hölder’s inequality, $${\left\Vert {| D |}^{-1}(f_1f_2) \right\Vert}_{H^{s-1/2}} \lesssim
{\left\Vert f_1f_2 \right\Vert}_{L^p}
\lesssim {\left\Vert f_1 \right\Vert} {\left\Vert f_2 \right\Vert}_{L^{\frac{3}{3/2-s}}}
\lesssim {\left\Vert f_1 \right\Vert} {\left\Vert f_2 \right\Vert}_{H^s},$$ where $$\frac{1}{p} - \frac12 = \frac{3/2-s}{3},$$ and this proves .
This concludes the proof of the estimate , hence .
Proof of
---------
By the energy inequality for the wave equation, it is enough to prove $$\int_{-T}^T {\left\Vert \square A_\mu^{\text{inh.}}(t) \right\Vert}_{H^{s-3/2}} \, dt
< \infty,$$ which in view of reduces to $$\int_{-T}^T {\left\Vert {| \psi(t) |}^2 \right\Vert}_{H^{s-3/2}} \, dt
< \infty.$$ But the latter follows immediately from (which holds with ${| D |}^{-1}$ replaced by ${\langle D \rangle}^{-1}$), since $\psi \in C([-T,T];H^s)$.
This concludes the proof of Theorem \[A:Thm2\].
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'Rotation and reflection symmetries impose that out-of-plane (flexural) phonons of free-standing graphene membranes have a quadratic dispersion at long wavelength and can be excited by charge carriers in pairs only. As a result, we find that flexural phonons dominate the phonon contribution to the resistivity $\rho$ below a crossover temperature $T_x$ where we obtain an anomalous temperature dependence $\rho\propto T^{5/2}_{}\ln T$. The logarithmic factor arises from renormalizations of the flexural phonon dispersion due to coupling between bending and stretching degrees of freedom of the membrane.'
author:
- Eros Mariani
- Felix von Oppen
title: 'Flexural phonons in free-standing graphene'
---
[*Introduction*]{}.—The experimental realization of monolayers of graphite, termed graphene, has opened new horizons in the physics of two-dimensional electron systems (2DES) [@Geim; @Kim]. Unlike conventional 2DES, the low-energy electronic bandstructure of graphene [@Wallace; @Gonzales; @Dresselhaus] is described by a massless Dirac equation with velocity $v$. The $4\times 4$ matrix structure of the Dirac equation reflects the two sublattices of the graphene honeycomb lattice in combination with a valley degeneracy due to the presence of two Dirac cones within the Brillouin zone. Pioneering experiments on this novel two-dimensional electron system have shown that the Dirac nature of carriers induces an anomalous integer quantum Hall effect as well as a finite conductivity at vanishing carrier density [@Geim; @Kim].
Recently, it has become possible to experiment on free-standing graphene sheets [@GeimRipples; @McEuen] which provide a realization of a two-dimensional (2d) solid. Studies of the stability of 2d solids against thermal fluctuations date back to early work by Peierls and Landau [@Peierls; @Landau] who pointed out the absence of true long-range translational order. Much later, it was understood [@Kosterlitz] that nevertheless, quasi-long-range translational order can persist up to a finite-temperature Kosterlitz-Thouless transition. While these works focus on in-plane distortions of the lattice, free-standing membranes also support out-of-plane distortions. It is believed that there is a low-temperature flat phase even in the presence of out-of-plane distortions [@Nelson], which gives way to a crumpled phase at high temperatures [@NelsonCrumpling].
Within the low-temperature flat phase, long-wavelength elastic distortions, both in-plane and out-of-plane, can be described by the appropriate elastic Lagrangian density [@Nelson; @LandauBook], $${\cal L} = \frac{\rho_0}{2} (\dot{\bf u}^2+\dot h^2) -\frac{1}{2}\kappa (\nabla^2 h)^2 - \mu
u^2_{ij}-\frac{1}{2} \lambda u^2_{kk}
\label{elastic}$$ in terms of the mass density $\rho_0$, the out-of-plane distortions $h({\bf r})$, and the strain tensor $u_{ij}=\frac{1}{2}[\partial_i u_j + \partial_j u_i + (\partial_i h) (\partial_j h)]$. Here, ${\bf u}({\bf r})$ denotes the in-plane distortions. The elastic constants $\lambda$ and $\mu$ characterize the in-plane rigidity of the lattice, $\kappa$ the bending rigidity. Both the absence of a $(\nabla h)^2$ term in the elastic Lagrangian and the appearance of the out-of-plane distortions $h({\bf r})$ in the strain tensor are direct consequences of the rotational symmetry of the membrane in the embedding space.
The elastic Lagrangian in Eq. (\[elastic\]) encapsulates a distinct difference between in-plane and out-of-plane (flexural) phonons. Indeed, to quadratic order in the displacement fields $h({\bf r})$ and ${\bf u}({\bf r})$, both longitudinal and transverse in-plane phonons have a [*linear*]{} dispersion $\omega^{(l)}_{\mathbf{q}}=v^{(l)}q$ and $\omega^{(t)}_{\mathbf{q}}=v^{(t)}q$ with group velocities $v^{(l)}=\left[\left(2\mu +\lambda\right)/\rho_0\right]^{1/2}$ and $v^{(t)}=\left[\mu/\rho_0\right]^{1/2}$. In contrast, flexural phonons obey a [*quadratic*]{} dispersion $\omega_{\mathbf{q}}^{(h)} = \alpha q^2$ with $\alpha =\left[\kappa/\rho_0\right]^{1/2}$, which is a consequence of rotational symmetry. In-plane and flexural phonons also differ in their coupling to the charge carriers of graphene membranes. While the coupling is conventional for in-plane phonons, the reflection symmetry $h\to -h$ demands that out-of-plane displacements enter only [*quadratically*]{} into the Dirac Hamiltonian. Consequently, charge carriers can excite flexural phonons only [*in pairs*]{}.
Due to these differences, we find that flexural phonons dominate the phonon contribution to the resistivity of free-standing graphene membranes below a crossover temperature $T_x$. Indeed, the transport scattering rate of Dirac fermions [*diverges*]{} logarithmically for a strictly quadratic dispersion of flexural phonons. This divergence is cut off by the coupling between bending and stretching degrees of freedom of the membrane, as captured by the elastic Lagrangian Eq. (\[elastic\]). At finite temperature, this coupling renormalizes the bending rigidity of the membrane, inducing a stiffening of the flexural-mode dispersion at long wavelengths. Including this physics within a simple one-loop RG, we find that the contribution of flexural phonons to the resistivity of graphene membranes scales as $T^{5/2}_{}\ln T$.
[*Graphene*]{}.—The bandstructure of graphene is well approximated by the tight-binding Hamiltonian $$\label{H}
H=-t \sum_{\langle ij\rangle} [c^{\dagger}_{i}c_{j}+c^{\dagger}_{j}c_{i}]$$ on a honeycomb lattice. Here, $t$ is the hopping matrix element, $c_{i}$ annihilates an electron on lattice site $i$, and only nearest-neighbor hopping has been included. The 2d hexagonal lattice consists of two identical sublattices A and B, and thus two sites per unit cell. We denote the vectors connecting a B site with the neighboring A sites as $\mathbf{e}^{}_{1}=a\, (-1,0)$, $\mathbf{e}^{}_{2}=a\, (1/2,\sqrt{3}/2)$ and $\mathbf{e}^{}_{3}=a\, (1/2,-\sqrt{3}/2)$ where $a$ is the bond length. The band structure of the Hamiltonian in Eq. (\[H\]) has zero energy (corresponding to the Fermi energy at half filling) at two inequivalent points in the Brillouin zone, which we choose to be at $\mathbf{k}^{}_{\pm}=\pm\mathbf{k}_{D}^{}$, with $\mathbf{k}_{D}^{}=2\pi/(3\sqrt{3}a)\, (\sqrt{3},1)$. In the vicinity of these Dirac points, the spectrum is described by the $4\times 4$ Dirac Hamiltonian $$\label{HDirac}
H=\hbar v\,\mathbf{\Sigma}\cdot\mathbf{k}$$ with velocity $v=3ta/2$. The 2d wavenumber $\mathbf{k}$ is measured from the Dirac point. The Hamiltonian in Eq. (\[HDirac\]) acts on four-component spinors $(u_{A,\mathbf{k}}^{+},u_{B,\mathbf{k}}^{+},u_{B,\mathbf{k}}^{-},u_{A,\mathbf{k}}^{-})$ of Bloch amplitudes in the space spanned by the sublattices ($A/B$) and Dirac points ($+/-$). The matrices $\Sigma_{x,y}^{}=\Pi_{z}^{}\otimes\mathbf{\sigma}_{x,y}^{}$ denote components of a vector $\mathbf{\Sigma}$. ($\Pi_{i}^{}$ and $\sigma_{j}^{}$ are Pauli matrices acting in the spaces of the Dirac points and the sublattices, respectively). It is also useful to introduce a corresponding vector ${\bf \Lambda}$ by $\Lambda_{x,y}^{}=\Pi_{x,y}^{}\otimes\sigma_{z}^{}$ and $\Lambda_{z}^{}=\Pi_{z}^{}\otimes\sigma_{0}^{}$ [@FalkoWL].
[*Electron-phonon coupling*]{}.—The dominant electron-phonon coupling arises from distortion-induced modifications of the bond lengths and hence the hopping amplitude [@Mahan; @Manes; @Arpes]. The electron-phonon coupling can then be described in terms of a fictitious gauge field ${\bf A}({\bf r})$ entering into the Dirac Hamiltonian, $H=v{\bf \Sigma}\cdot ({\bf p}+e{\bf A}({\bf r},t))$, where $\mathbf{p}$ denotes the momentum. Expressed in terms of the strain tensor $u_{ij}$, the gauge field ${\bf A}({\bf r})$ takes the form [@Katsnelson] $$\begin{aligned}
e{\bf A}({\bf r},t)&=& \Pi_z \otimes {\bf 1}\, \frac{\hbar}{t}\frac{\partial t}{\partial a}\left[\begin{array}{c} u_{xy} \\
\frac{1}{2}(u_{xx}-u_{yy})
\end{array}\right] \; .
\label{fictitious}\end{aligned}$$ The combination of components of the strain tensor entering into ${\bf A}({\bf r})$ reflects the symmetry of the underlying honeycomb lattice. The factor of $\Pi_z$ implies that the associated fictitious magnetic field points in opposite directions at the two Dirac points, as required by time-reversal symmetry. From the definition of the strain tensor, we directly conclude that electrons couple linearly to the in-plane phonons and quadratically to flexural ones. The corresponding electron-phonon vertices are depicted in Fig. \[Vertices\].
![Electron-phonon vertices. a) Coupling of electrons to in-plane phonons. b) Coupling of electrons to flexural phonons. Straight lines correspond to electrons, wavy (dashed) lines to in-plane (flexural) phonons. \[Vertices\]](Vertices2.eps){width="0.8\columnwidth"}
At sufficiently low temperatures, only long-wavelength phonons contribute to the resistivity and scattering between different Dirac cones can be neglected. Thus, we can restrict attention to the vicinity of, say, the Dirac point $\mathbf{k}_{+}^{}$ where the Hamiltonian reduces to a $2\times 2$ Hamiltonian in the $A-B$ sublattice space.
The electron-phonon coupling $H_{\mathrm{ep}}^{}$ corresponding to Eq. (\[fictitious\]) can be expressed in second quantization after expanding the in-plane and out-of-plane distortions into Fourier series as $\mathbf{u}(\mathbf{r}) = \sum_{\mathbf{q}}^{}\mathbf{u}_{\mathbf{q}}^{}\, e^{i\mathbf{q}\cdot\mathbf{r}}_{}$ and $h(\mathbf{r}) = \sum_{\mathbf{q}}^{}h_{\mathbf{q}}^{}\, e^{i\mathbf{q}\cdot\mathbf{r}}_{}$, decomposing $\mathbf{u}_{\mathbf{q}}^{}$ into longitudinal and transverse components, $$\label{ulongtrans}
\mathbf{u}_{\mathbf{q}}^{}=u_{\mathbf{q}}^{(l)}\hat{\mathbf{q}}+u_{\mathbf{q}}^{(t)}\hat{\mathbf{z}}\times\hat{\mathbf{q}}\; ,$$ and quantizing the amplitude of the distortions as $$u_{\mathbf{q}}^{(\nu )} = \sqrt{\frac{\hbar}{2M\omega_{\mathbf{q}}^{(\nu )}}}\,\left(a^{(\nu )}_{\mathbf{q}}+a^{(\nu )\dagger}_{-\mathbf{q}}\right)\; .$$ Here $M$ is the atomic mass and $a^{(\nu )}_{\mathbf{q}}$ the annihilation operator of a phonon of type $\nu$ ($\nu=l,t,h$) with wavenumber $\mathbf{q}$. This yields $$\begin{aligned}
H_{\mathrm{ep}}^{}=\sum_{\mathbf{k},\mathbf{k}^{\prime},\mathbf{q}}^{}\sum_{\nu =l,t}^{}\, V^{(\nu )}_{\mathbf{q}}c^{\dagger}_{B,\mathbf{k}}c^{}_{A,\mathbf{k}^{\prime}}\left(a^{(\nu )}_{\mathbf{q}}+a^{(\nu )\dagger}_{-\mathbf{q}}\right)\delta_{\mathbf{k}^{\prime},\mathbf{k}^{}-\mathbf{q}^{}}^{} && \nonumber \\
+\sum_{\mathbf{k},\mathbf{k}^{\prime},\mathbf{q}^{}_{},\mathbf{q}^{\prime}}\, V^{(h)}_{\mathbf{q},\mathbf{q}^{\prime}}c^{\dagger}_{B,\mathbf{k}}c^{}_{A,\mathbf{k}^{\prime}}\left( a^{(h)}_{\mathbf{q}}+a^{(h)\dagger}_{-\mathbf{q}}\right)\quad &&\\
\quad\quad\times\left( a^{(h)}_{\mathbf{q}^{\prime}}+a^{(h)\dagger}_{-\mathbf{q}^{\prime}}\right)\delta_{\mathbf{k}^{\prime},\mathbf{k}^{}-\mathbf{q}^{}-\mathbf{q}^{\prime}}^{} +h.c. \quad &&\nonumber
\label{Helph}\end{aligned}$$ in terms of the coupling terms $$\begin{aligned}
V^{(l)}_{\mathbf{q}}=&&-\epsilon \, q\, e^{-2i\phi}_{}\sqrt{\frac{\hbar}{2M\omega_{\mathbf{q}}^{(l)}}}\nonumber \\
V^{(t)}_{\mathbf{q}}=&&-i\epsilon \, q\, e^{-2i\phi}_{}\sqrt{\frac{\hbar}{2M\omega_{\mathbf{q}}^{(t)}}}\\
V^{(h)}_{\mathbf{q},\mathbf{q}^{\prime}}=&&-\frac{1}{2}\,i\epsilon \, q\, q^{\prime}_{}\, e^{-i(\phi +\phi^{\prime}_{})}_{}\frac{\hbar}{2M\sqrt{\omega_{\mathbf{q}}^{(h)}\omega_{\mathbf{q}^{\prime}}^{(h)}}}\nonumber
\label{Elements}\end{aligned}$$ where $\epsilon =(3a/4)\partial t/\partial a$, $c^{\dagger}_{B,\mathbf{k}}$ denotes the creation operator of an electron in a Bloch state in sublattice $B$ and wave-vector $\mathbf{k}$, and $\phi$ ($\phi^{\prime}_{}$) is the angle of $\mathbf{q}$ ($\mathbf{q}^{\prime}_{}$) with respect to the $x$-axis.
These considerations give rise to an interesting competition: In-plane phonons are strongly coupled to electrons (the vertex is first order in the phonons) but they have a linear dispersion and hence a linearly vanishing density of states at small energy. In contrast, flexural phonons are weakly coupled to electrons (their vertex is second order in the phonons) but their dispersion is quadratic with constant density of states. In the following, we quantitatively analyze the consequences of this competition for the temperature dependence of the resistivity. The latter is determined by $$\label{resistivity}
\rho =\frac{2}{e^{2}_{}v_{}^{2}\nu_{F}^{}}\,\frac{1}{\tau_{\mathrm{tr}}^{}}$$ ($\nu_{F}^{}$ is the electronic density of states at the Fermi level) in terms of the transport scattering rate $$\frac{1}{\tau_{\rm tr}}=\frac{2\pi}{\hbar} \sum_{f}^{} \left|M_{fi}^{}\right|^{2}_{}\left(1-\cos \theta\right)\delta(E_{f}-E_{i}\pm\hbar\omega)\; .
\label{Fermi}$$ of Dirac fermions due to absorption (or emission) of phonons. Here, $\theta$ is the scattering angle of the Dirac fermions, $|i \rangle$ and $|f \rangle$ the initial and final scattering states, $E_{i}$ and $E_{f}$ the initial and final electronic energies, $\hbar\omega$ the energy of the absorbed (emitted) phonons, and $M_{fi}^{}=\langle f|H_{\rm ep}^{}|i \rangle$.
We consider the doped regime with $E_{F}^{}\gg k_{\mathrm{B}}^{}T$ and employ the conventional quasielastic approximation of neglecting the phononic contribution to energy conservation, as the typical phonon energies are small compared to $E_{F}^{}$. Moreover, we focus on sufficiently low temperatures so that we can restrict attention to the quadratic (linear) region of the dispersion of flexural (in-plane) phonons.
[*Flexural phonons*]{}.—Scattering of Dirac fermions by flexural phonons requires absorption (or emission) of two phonons, say with wavenumbers $\mathbf{q}$ and $\mathbf{q}^{\prime}_{}$. Thus, the corresponding initial and final scattering states are $|i\rangle =|\mathbf{k},\sigma^{}_{}\rangle \otimes |n^{(h)}_{\mathbf{q}}\rangle$ and $|f\rangle =|\mathbf{k}^{}_{}+\mathbf{q}^{}_{}+\mathbf{q}^{\prime}_{},\sigma^{\prime}_{}\rangle \otimes |n^{(h)}_{\mathbf{q}}-1,n^{(h)}_{\mathbf{q}^{\prime}_{}}-1\rangle$, where $n^{(h)}_{\mathbf{q}}=\left[ \exp (\hbar\omega^{(h)}_{\mathbf{q}}/k_{\mathrm{B}}^{}T)-1\right]^{-1}_{}$ denotes the Bose distribution function of the flexural phonons and $|\mathbf{k}^{}_{},\sigma^{}_{}\rangle=1/\sqrt{2}\left(\sigma e^{-i\xi}_{}c^{\dagger}_{A,\mathbf{k}}+c^{\dagger}_{B,\mathbf{k}}\right)|\mathrm{vac}\rangle$ is a Dirac fermion state state with momentum $\mathbf{k}$ and chirality $\sigma$. (Here, $\xi$ is the angle between $\mathbf{k}$ and the $x$-axis and $|\mathrm{vac}\rangle$ is the electronic vacuum). Thus, $|M_{fi}^{}|^{2}_{}= |V^{(h)}_{\mathbf{q},\mathbf{q}_{}^{\prime}}|^{2}_{}\left[1- \sigma\sigma^{\prime}_{}\cos (\xi +\xi^{\prime}_{}+2\phi+2\phi^{\prime}_{})\right]/2$. The summation over final states requires integration over $\mathbf{q}$ and $\mathbf{q}_{}^{\prime}$, while averaging over the direction of the incoming electron will suppress the oscillatory term in $|M_{fi}^{}|^{2}_{}$. Using $a^{(\nu )}_{\mathbf{q}}|n^{(\nu )}_{\mathbf{q}}\rangle =(n^{(\nu )}_{\mathbf{q}})^{1/2}_{}|n^{(\nu )}_{\mathbf{q}}-1\rangle$, $E_{i}^{}=\hbar vk=E_{F}^{}$, and $E_{f}^{}=\hbar v|\mathbf{k}+\mathbf{q}+\mathbf{q}^{\prime}_{}|$, we obtain $$\frac{1}{\tau_{\mathrm{tr}}^{(h)}}=\sum_{\mathbf{q},\mathbf{q}^{\prime}_{}}^{} \frac{\pi|V^{(h)}_{\mathbf{q},\mathbf{q}^{\prime}}|^{2}_{}}{\hbar}\left(1-\cos \theta\right)n^{(h)}_{q}n^{(h)}_{q^{\prime}_{}}\delta(\Phi_{\mathbf{q}+\mathbf{q}^{\prime}_{}}) \; ,
\label{Rateflex}$$ where $\Phi_{\mathbf{q}+\mathbf{q}^{\prime}_{}}=\hbar v|\mathbf{k}+\mathbf{q}+\mathbf{q}^{\prime}_{}|-E_{F}^{}$ and $|V^{(h)}_{\mathbf{q},\mathbf{q}^{\prime}}|^{2}_{}=\hbar^{2}_{}\epsilon^{2}_{}/(16 M^{2}_{}\alpha^{2}_{})$ is independent of wavenumbers due to the quadratic dispersion of flexural phonons. The rate in Eq. (\[Rateflex\]) is formally singular at small $q,q^{\prime}_{}$ since both Bose distributions diverge as $T/q^{2}_{}$. Rescaling momenta by $\sqrt{T}$ and introducing a cutoff $q_{c}$ at small wavenumbers (to be specified below), we obtain the scattering rate $$\label{RateflexT}
\frac{1}{\tau^{(h)}_{\rm tr}}\simeq\frac{C^{(h)}_{}}{32\, \pi^{2}}\frac{\hbar\omega^{(h)}_{k_{F}^{}}}{\kappa}\left(\frac{k_{\mathrm{B}}^{}T}{\hbar\omega_{k_{F}^{}}^{(h)}}\right)^{5/2}_{}\ln \left(\frac{k_{\mathrm{B}}^{}T}{\hbar \omega^{(h)}_{q_{c}^{}}}\right) \; ,$$ where $C^{(\nu )}_{}=\epsilon^{2}_{}\hbar k^{4}_{F}/(4\rho_0 E_{F}^{}\hbar \omega^{(\nu )}_{k_{F}^{}})$. [*The unusual $T^{5/2}_{}$ scaling implies that scattering from flexural phonons dominates the phonon contribution to the resistivity at low temperatures.*]{} Indeed, the conventional phonon contribution to the resistivity due to in-plane phonons scales as $T^4$, which is the direct two-dimensional analog of the $T^5$ law in bulk solids.
Eq. (\[RateflexT\]) shows that the scattering rate from flexural phonons diverges logarithmically for a strictly quadratic phonon dispersion. For clean elastic membranes, a low-momentum cutoff arises from the coupling terms between bending and stretching degrees of freedom of the membrane which are included in the Lagrangian Eq. (\[elastic\]). These lead to long-wavelength corrections to the elastic constants and hence the phonon dispersions [@Nelson]. Indeed, it is these renormalizations which are responsible for the stability of the flat phase at low temperatures.
In order to study the renormalization of the bending rigidity by the coupling of bending and stretching modes, we integrate out the in-plane distortions in Eq. (\[elastic\]) and obtain an effective energy functional for the flexural modes alone [@Nelson], $$F=\frac{1}{2}\int \mathrm{d}\mathbf{r} \left[\kappa \left(\nabla^{2}_{}h\right)^{2}_{}+\frac{K_{0}^{}}{4}\left[ P^{\perp}_{\alpha\beta}\left(\partial_{\alpha}^{}h\right)\left(\partial_{\beta}^{}h\right)\right]^{2}_{}\right]\, ,
\label{Flexaction}$$ where $K_{0}^{}=4\mu (\mu+\lambda)/(2\mu +\lambda)$ and $P^{\perp}_{\alpha\beta}=(\epsilon_{\alpha l}^{}\,\epsilon_{\beta k}^{}\, \partial_{l}^{}\, \partial_{k}^{})/\nabla^{2}_{}$ is a transverse projector. This effective energy functional contains a four-leg interaction between flexural distortions. Treating this quartic term to one-loop order (depicted in Fig. \[Loop\]) in a conventional momentum-shell RG, one obtains the flow equation $$\label{RG}
\frac{d\kappa}{dq}=-\frac{3}{16\pi}\,\frac{K_{0}^{}k_{\mathrm{B}}^{}T}{\kappa q^{3}_{}}\, ,$$ with $q$ denoting the running shell wavevector.
![One-loop correction to the bending rigidity due to the effective interaction between flexural modes. \[Loop\]](Loop.eps){width="0.5\columnwidth"}
The flow equation Eq. (\[RG\]) is readily solved, and the resulting scaling of $\kappa$ with $q$ is $$\label{kappalambda}
\kappa(q)=\kappa\sqrt{1+q_{c}^{2}/q^{2}_{}}$$ in terms of the temperature-dependent momentum scale $q_{c}^{}=[3K_{0}^{}k_{\mathrm{B}}^{}T/ (8\pi\kappa^{2}_{})]^{1/2}_{}$. Thus, the flexural phonon dispersion is quadratic for $q\gg q_c$. In contrast, for $q\ll q_c$, thermal fluctuations effectively stiffen the membrane and, within the simple one-loop analysis presented here, we find a renormalized dispersion $\omega_{q\ll q_{c}^{}}^{(h)}=(\kappa q_{c}^{})^{1/2}_{} q^{3/2}_{}$. It is this renormalization of the flexural-phonon dispersion at long wavelength which removes the singularity in the phonon scattering rate Eq. (\[RateflexT\]). We therefore identify the low-momentum cutoff $q_c$ entering into Eq. (\[RateflexT\]) with this momentum scale.
As a result, we find that the temperature cancels from the argument of the logarithm in Eq. (\[RateflexT\]). This analysis is approximate in that it is restricted to one-loop order and that it neglects the flow of the elastic constants for stretching deformations. While the stiffening against bending deformation would survive inclusion in a more refined treatment, the precise power-law dependence of $q_c$ on temperature would change. As a result, the argument of the logarithm would become temperature dependent, resulting in an overall $T^{5/2}\ln T$ scaling of the flexural phonon contribution to the resistivity.
[*Crossover temperature.*]{}—In order to estimate the crossover temperature $T_x$ below which phonon scattering of Dirac fermions is dominated by flexural modes, we note that an analogous calculation yields $$\label{RateinT}
\frac{1}{\tau_{\rm tr}^{(\nu )}}\simeq C^{(\nu )}_{}\,\left(\frac{k_{\mathrm{B}}^{}T}{\hbar\omega^{(\nu )}_{k_{F}^{}}}\right)^{4}$$ for the transport scattering rate from in-plane phonons ($\nu = l,t$). Thus, the crossover temperature obtained from a comparison of Eqs. (\[RateflexT\]) and (\[RateinT\]) becomes $$\label{Tx}
T_{x}^{} = \frac{1}{k_{\mathrm{B}}}\left(\frac{\ln \left( k_{\mathrm{B}}^{}T/\hbar \omega^{(h)}_{q_{c}^{}}\right)}{8\, (2\pi )^{2}_{}}\, \frac{\hbar\omega^{(t)\, 5}_{k_{F}^{}}}{\kappa \left(\hbar\omega^{(h)}_{k_{F}^{}}\right)^{5/2}}\right)^{2/3}_{}.$$ This result is independent of the Fermi energy (i.e. the doping level) and with typical parameters for graphene [@Parameters], we obtain $T_{x}^{}\simeq 70\, \mathrm{K}$. There is therefore a significant temperature range accessible in experiment over which our predictions for the resistivity can be tested.
[*Conclusions*]{}.—For clean graphene membranes, long-wavelength renormalizations of their elastic properties due to thermal fluctuations are crucial in order to obtain a finite transport scattering rate and hence resistivity $\rho\propto T^{5/2}_{}\ln T$ from scattering by flexural phonons. In the presence of disorder, the elastic properties of the membrane are renormalized even at zero temperature [@Radzihovsky; @Lubensky]. Indeed, the recent experimental observations of rippling [@GeimRipples] suggest that disorder, inside or close to the 2d membrane, exists in present graphene membranes. Such disorder-induced renormalizations of the elastic moduli may compete with the renormalizations by thermal fluctuations and lead to a temperature-independent saturation of the cutoff $q_c$ at low temperatures. A detailed study of the effects of disorder in this context remains an important topic for future research.
[*Acknowledgments*]{}—We are grateful to F. Guinea, T. Nattermann, L. Peliti, and Ady Stern for instructive discussions. This work was supported in part by DIP. FvO enjoyed the hospitality of the KITP Santa Barbara (supported in part by NSF grant PHY99-07949) and EM acknowledges the hospitality of the Weizmann Institute of Science (supported by grant RITA-CT-2003-506095) .
[99]{} K.S. Novoselov et al., Science **306**, 666 (2004); K.S. Novoselov et al., Nature **438**, 197 (2005).
Y. Zhang et al., Nature **438**, 201 (2005); Y. Zhang et al., Phys. Rev. Lett. **94**, 176803 (2005).
P.R. Wallace, Phys. Rev. [**71**]{}, 622 (1947).
J. Gonzales, F. Guinea, and M.A.H. Vozmediano, Nucl. Phys. B [**406**]{}, 771 (1993).
R. Saito, G. Dresselhaus and M.S. Dresselhaus, *Physical Properties of Carbon Nanotubes*, Imperial College Press, London 1998.
J.C. Meyer et al. , cond-mat/0701379 (2007).
J. Scott Bunch et al. , Science [**315**]{}, 490 (2007).
R.E. Peierls, Ann. I. H. Poincare [**5**]{}, 177 (1935).
L.D. Landau, Phys. Z. Sowjetunion [**11**]{}, 26 (1937).
V.L. Berezinskii, Sov. Phys. JETP [**32**]{}, 493 (1971); J.M. Kosterlitz and D.J. Thouless, J. Phys. C [**6**]{}, 1181 (1973).
D.R. Nelson and L. Peliti, J. Phys. (Paris) [**48**]{}, 1085 (1987).
M. Paczuski, M. Kardar, and D.R. Nelson, Phys. Rev. Lett. [**60**]{}, 2638 (1988).
L.D. Landau and E.M. Lifshitz, [*Theory of Elasticity*]{}, (Pergamon, New York, 1986).
E. McCann [*et al.*]{}, Phys. Rev. Lett. [**97**]{}, 146805 (2006).
L.M. Woods and G.D. Mahan, Phys. Rev. B [**61**]{}, 10651 (2000).
J.L. Manes, cond-mat/0702465 (2007).
For the effects of in-plane phonons on the graphene bandstructure, see M. Calandra and F. Mauri cond-mat/0707.1467 (2007), C.H. Park [*et al.*]{}, cond-mat/0707.1666 (2007), W.-K. Tse and S. Das Sarma, cond-mat/0707.3651 (2007).
See also A.K. Geim and M.I. Katsnelson, cond-mat/0706.2490 (2007).
L. Radzihovsky and D.R. Nelson, Phys. Rev. A [**44**]{}, 3525 (1991).
D.C. Morse, T.C. Lubensky, and G.S. Grest, Phys. Rev. A [**45**]{}, R2151 (1992); D.C. Morse and T.C. Lubensky, Phys. Rev. A [**46**]{}, 1751 (1992).
Typical parameters for graphene sheets are: $\kappa =1\,\mathrm{eV}$, $\mu \simeq 4\lambda =9\,\mathrm{eV}\AA^{-2}_{}$, $\hbar \alpha =35\,\mathrm{meV}\AA^{2}_{}$.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: |
In this paper, we establish a connection between the $1/k$-Eulerian polynomials introduced by Savage and Viswanathan ([Electron. J. Combin. 19 (2012), \#P9]{}) and $k$-Stirling permutations. We also introduce the dual set of Stirling permutations.
[*Keywords*]{}: $k$-Stirling permutations; $1/k$-Eulerian polynomials; Ascent-plateau
address:
- 'School of Mathematics and Statistics, Northeastern University at Qinhuangdao, Hebei 066004, P. R. China'
- 'Department of Mathematics, University of Haifa, 3498838 Haifa, Israel'
author:
- 'Shi-Mei Ma'
- Toufik Mansour
title: 'The $1/k$-Eulerian polynomials and $k$-Stirling permutations'
---
Introduction
============
For $k\geq 1$, the [*$1/k$-Eulerian polynomials*]{} $A_n^{(k)}(x)$ are defined by $$\label{Ankx-def01}
\sum_{n\geq 0}A_n^{(k)}(x)\frac{z^n}{n!}=\left(\frac{1-x}{e^{kz(x-1)}-x} \right)^{\frac{1}{k}}.$$
Let $e=(e_1,e_2,\ldots,e_n)\in{ \mathbb{Z}}^n$. Let $I_{n,k}=\left\{ e|0\leq e_i\leq (i-1)k\right\}$, which known as the set of $n$-dimensional [*$k$-inversion sequences*]{}. The number of [*ascents*]{} of $e$ is defined by $${{\rm asc\,}}(e)=\#\left\{i:1\leq i\leq n-1\mid\frac{e_i}{(i-1)k+1}<\frac{e_{i+1}}{ik+1}\right\}.$$ Savage and Viswanathan [@Savage12] showed that $$\label{Ankx-def02}
A_n^{(k)}(x)=\sum_{e\in I_{n,k}}x^{{{\rm asc\,}}(e)}.$$
Let ${\mathfrak{S}_n}$ be the symmetric group on the set $[n]=\{1,2,\ldots,n\}$ and $\pi=\pi_1\pi_2\cdots\pi_n\in{\mathfrak{S}_n}$. The number of [*excedances*]{} of $\pi$ is ${{\rm exc\,}}(\pi):=\#\{i:1\leq i\leq n-1|\pi_i>i\}$. Let ${{\rm cyc\,}}(\pi)$ be the number of [*cycles*]{} in the disjoint cycle representation of $\pi$. In [@FS70], Foata and Schützenberger introduced a $q$-analog of the classical Eulerian polynomials defined by $$\label{anxq-def}
A_n(x;q)=\sum_{\pi\in{\mathfrak{S}_n}}x^{{{\rm exc\,}}(\pi)}q^{{{\rm cyc\,}}(\pi)}.$$ The polynomials $A_n(x;q)$ satisfy the recurrence relation $$\label{anxq-rr}
A_{n+1}(x;q)=(nx+q)A_{n}(x;q)+x(1-x)\frac{d}{d x}A_{n}(x;q),$$ with the initial conditions $A_{1}(x;q)=1$ and $A_{2}(x;q)=q$ (see [@Brenti00 Proposition 7.2]). Savage and Viswanathan [@Savage12 Section 1.5] discovered that $$\label{Ankx-def03}
A_n^{(k)}(x)=k^nA_n(x;1/k)=\sum_{\pi\in{\mathfrak{S}_n}}x^{{{\rm exc\,}}(\pi)}k^{n-{{\rm cyc\,}}(\pi)}.$$ Let $A_n^{(k)}(x)=\sum_{j=0}^{n-1}a_{n,j}^{(k)}x^j$. It follows from and that $$\label{anj-rr}
a_{n+1,j}^{(k)}=(1+kj)a_{n,j}^{(k)}+k(n-j+1)a_{n,j-1}^{(k)},$$ with the initial condition $a_{1,0}^{(k)}=1$.
Let ${\genfrac{[}{]}{0pt}{}{n}{k}}$ be the Stirling number of the first kind, i.e., the number of permutations in ${\mathfrak{S}_n}$ with precisely $k$ cycles. It is well known that $$\label{Stirlingnumbers}
\sum_{k=0}^n{\genfrac{[}{]}{0pt}{}{n}{k}}x^k=\prod_{i=0}^{n-1}(x+i).$$ Thus it follows from that $$\label{Ank1}
A_n^{(k)}(1)=\prod_{i=1}^{n-1}(ik+1)\quad\textrm{for $n\ge 1$}.$$ Since $\prod_{i=1}^{n-1}(ik+1)$ also count $k$-Stirling permutations of order $n$ (see [@Janson11; @Kuba12]), it is natural to consider the following question: Is there existing a connection between $A_n^{(k)}(x)$ and $k$-Stirling permutations? The main object of this paper is to provide a solution to this problem.
$k$-Stirling permutations and their longest ascent-plateau
==========================================================
In the following discussion, we always let $j^i=\underbrace{j,j,\ldots,j}_i$ for $i,j\geq 1$. Stirling permutations were defined by Gessel and Stanley [@Gessel78]. A [*Stirling permutation*]{} of order $n$ is a permutation of the multiset $\{1^2,2^2,\ldots,n^2\}$ such that for each $i$, $1\leq i\leq n$, all entries between the two occurrences of $i$ are larger than $i$. We call a permutation of the multiset $\{1^k,2^k,\ldots,n^k\}$ a $k$-[*Stirling permutation*]{} of order $n$ if for each $i$, $1\leq i\leq n$, all entries between the two occurrences of $i$ are at least $i$. Denote by ${\mathcal{Q}_n}(k)$ the set of $k$-[*Stirling permutation*]{} of order $n$. Clearly, ${\mathcal{Q}_n}(1)={\mathfrak{S}_n}$ and ${\mathcal{Q}_n}(2)$ is the set of ordinary Stirling permutations of order $n$.
For $\sigma=\sigma_1\sigma_2\cdots\sigma_{2n}\in{\mathcal{Q}_n}(2)$, an occurrence of an [*ascent*]{} (resp. [*plateau*]{}) is an index $i$ such that $\sigma_i<\sigma_{i+1}$ (resp. $\sigma_i=\sigma_{i+1}$). The reader is referred to [@Bona08; @Janson11; @Kuba12; @Remmel14] for recent progress on the study of patterns in Stirling permutations.
Let $\sigma=\sigma_1\sigma_2\cdots\sigma_{kn}\in {\mathcal{Q}_n}(k)$. We say that an index $i\in \{2,3,\ldots,nk-k+1\}$ is a longest ascent-plateau if $$\sigma_{i-1}<\sigma_i=\sigma_{i+1}=\sigma_{i+2}=\cdots=\sigma_{i+k-1}.$$
Let ${{\rm ap\,}}(\sigma)$ be the number of the longest ascent-plateau of $\sigma$. For example, ${{\rm ap\,}}(1122\textbf{3}3321)=1$.
Now we present the main results of this paper.
\[thm:01\] For $n\geq 1$ and $k\geq 2$, we have $$\label{Ankx-stirling}
A_n^{(k)}(x)=\sum_{\sigma\in {\mathcal{Q}_n}(k)}x^{{{\rm ap\,}}(\sigma)}.$$
Let $$T(n,j;k)=\#\{\sigma\in {\mathcal{Q}_n}(k):{{\rm ap\,}}(\sigma)=j\}.$$ There are two ways in which a permutation $\widetilde{\sigma}\in\mathcal{Q}_{n+1}(k)$ with the number of the longest ascent-plateau equals $j$ can be obtained from a permutation $\sigma\in\mathcal{Q}_{n}(k)$.
1. If the number of the longest ascent-plateau of $\sigma$ equals $j$, then we can insert $k$ copies of $(n+1)$ into $\sigma$ without increasing the number of the longest ascent-plateau. Let $i$ be one of the longest ascent-plateau of $\sigma$. Then we can insert $k$ copies of $(n+1)$ before $\sigma_i$ or after $\sigma_{t}$, where $i\leq t\leq i+k-2$. Moreover, the $k$ copies of $(n+1)$ can also be inserted into the front of $\sigma$. This accounts for $(1+kj)T(n,j;k)$ possibilities.
2. If the number of the longest ascent-plateau of $\sigma$ equals $j-1$, then we insert $k$ copies of $(n+1)$ into the remaining $1+kn-(1+k(j-1))=k(n-j+1)$ positions. This gives $k(n-j+1)T(n,j-1;k)$ possibilities.
Hence $$T(n+1,j;k)=(1+kj)T(n,j;k)+k(n-j+1)T(n,j-1;k).$$ Clearly, $T(n,0;k)=1$, corresponding to the permutation $n^k(n-1)^k\cdots1^k$. Therefore, the numbers $T(n,j;k)$ satisfy the same recurrence relation and initial conditions as $a_{n,j}^{(k)}$, so they agree.
Define $${\mathcal{Q}}^0_{n}(k)=\{0\sigma: \sigma\in{\mathcal{Q}_n}(k)\}.$$ Therefore, for $\sigma\in{\mathcal{Q}}^0_{n}(k)$, we let $\sigma_0=0$ and the indices of the longest ascent-plateau belong to $\{1,2,3,\ldots,nk-k+1\}$. For example, ${{\rm ap\,}}(0\textbf{1}12\textbf{3}32)=2$.
Define $$x^nA_n^{(k)}\left(\frac{1}{x}\right)=\sum_{j=1}^nb_{n,j}^{(k)}x^j.$$ Then $b_{n,j}^{(k)}=a_{n,n-j}^{(k)}$. It follows from that $$b_{n+1,j}^{(k)}=kjb_{n,j}^{(k)}+(kn-kj+k+1)b_{n,j-1}^{(k)}.$$ Along the same lines of the proof of Theorem \[thm:01\], we get the following result.
\[thm:02\] For $n\geq 1$ and $k\geq 2$, we have $$\label{Ankn-j}
x^nA_n^{(k)}\left(\frac{1}{x}\right)=\sum_{\sigma\in {\mathcal{Q}}^0_{n}(k)}x^{{{\rm ap\,}}(\sigma)}.$$
The dual set of Stirling permutations
=====================================
For convenience, we let ${\mathcal{Q}_n}={\mathcal{Q}_n}(2)$. Let $\sigma=\sigma_1\sigma_2\cdots\sigma_{2n}\in{\mathcal{Q}_n}$. Let $\Phi$ be the bijection which map each first occurrence of letter $j$ in $\sigma$ to $2j$ and the second occurrence of letter $j$ in $\sigma$ to $2j-1$, where $j\in [n]$. For example, $\Phi(221331)=432651$. The [*dual set*]{} $\Phi({\mathcal{Q}_n})$ of ${\mathcal{Q}_n}$ is defined by $$\Phi({\mathcal{Q}_n})=\{\pi: \sigma\in{\mathcal{Q}_n}, \Phi(\sigma)=\pi\}.$$ Clearly, $\Phi({\mathcal{Q}_n})$ is a subset of ${\mathfrak{S}_{2n}}$. Let $ab$ be an ascent in $\sigma$, so $a<b$. Using $\Phi$, we see that $ab$ is maps into $(2a-1)(2b-1)$, $(2a-1)(2b)$, $(2a)(2b-1)$ or $(2a)(2b)$, and vice versa. Let ${{\rm as\,}}(\sigma)$ (resp. ${{\rm as\,}}(\pi)$) be the number of ascents of $\sigma$ (resp. $\pi$). Then $\Phi$ preserving ascents, i.e., ${{\rm as\,}}(\sigma)={{\rm as\,}}(\Phi(\sigma))={{\rm as\,}}(\pi)$. Hence the well known [*Eulerian polynomial of second kind*]{} $P_n(x)$ (see [@Sloane A008517]) has the expression $$P_n(x)=\sum_{\pi\in\Phi({\mathcal{Q}_n})}x^{{{\rm as\,}}(\pi)}.$$
Perhaps one of the most important permutation statistics is the peaks statistic; see, e.g., [@Dilks09; @Ma12] and the references contained therein. Let $\pi=\pi_1\pi_2\cdots\pi_n\in{\mathfrak{S}_n}$. An [*interior peak*]{} in $\pi$ is an index $i\in\{2,3,\ldots,n-1\}$ such that $\pi_{i-1}<\pi_i>\pi_{i+1}$. Let ${{\rm ipk\,}}(\pi)$ denote the number of interior peaks in $\pi$. A [*left peak*]{} in $\pi$ is an index $i\in[n-1]$ such that $\pi_{i-1}<\pi_i>\pi_{i+1}$, where we take $\pi_0=0$. Denote by ${{\rm lpk\,}}(\pi)$ the number of left peaks in $\pi$. For example, ${{\rm ipk\,}}(21435)=1$ and ${{\rm lpk\,}}(21435)=2$.
As pointed out by Savage and Viswanathan [@Savage12 Section 4] that the numbers $a_{n,j}^{(2)}$ appear as A185410 in [@Sloane], and the numbers $a_{n,n-j}^{(2)}$ appear as A156919 in [@Sloane]. We can now present a unified characterization of these numbers.
For $n\geq 1$, we have $$\label{An2x}
A_n^{(2)}(x)=\sum_{\pi\in\Phi({\mathcal{Q}_n})}x^{{{\rm ipk\,}}(\pi)},$$ $$\label{An2x-reverse}
x^nA_n^{(2)}\left(\frac{1}{x}\right)=\sum_{\pi\in\Phi({\mathcal{Q}_n})}x^{{{\rm lpk\,}}(\pi)}.$$
Recall that an occurrence of a [*pattern*]{} $\tau$ in a sequence $\pi$ is defined as a subword in $\pi$ whose letters are in the same relative order as those in $\tau$.
Let $\sigma\in{\mathcal{Q}_n}$ and let $\Phi(\sigma)=\pi$. Let $$C=\{112,211,122,221,213,312,123,321\}.$$ For all $\sigma\in{\mathcal{Q}_n}$, we see that all patterns of length three of $\sigma$ are belong to $C$. Let $abb$ be an occurrence of the pattern $122$ in $\sigma$, so $a<b$. Using $\Phi$, we see that $abb$ is maps to either $(2a-1)(2b)(2b-1)$ or $(2a)(2b)(2b-1)$, which is an interior peak of the pattern $132$. Moreover, one can easily verify that interior peaks can not be generated by the other patterns. Recall that an occurrence of the longest ascent-plateau in Stirling permutations is an occurrence of the pattern $122$. Then we get by using Theorem \[thm:01\]. Similarly, from Theorem \[thm:02\], we get .
For $n\geq 1$, we define $C_n(x)$ by $$\label{def:Cnx}
(1+x)C_n(x)=xA_n^{(2)}(x^2)+x^{2n}A_n^{(2)}\left(\frac{1}{x^2}\right).$$ Set $C_0(x)=1$. It follows from that $$C(x,z)=\sum_{n\geq0} C_n(x)\frac{z^n}{n!}=\frac{{e^{z \left( x-1 \right) \left( 1+x \right) }}+x}{1+x}\sqrt {{\frac {1-x^2}{{e^{2\,z \left( x-1
\right) \left( 1+x \right) }}-x^2}}}.$$
The first few $C_n(x)$ are given as follows: $$\begin{aligned}
C_1(x)&=x, \\
C_2(x)&=x+x^2+x^3, \\
C_3(x)&=x+3x^2+7x^3+3x^4+x^5,\\
C_4(x)&=x+7x^2+29x^3+31x^4+29x^5+7x^6+x^7,\\
C_5(x)&=x+15x^2+101x^3+195x^4+321x^5+195x^6+101x^7+15x^8+x^9.\end{aligned}$$
Let $\pi=\pi_1\pi_2\cdots\pi_n\in{\mathfrak{S}_n}$. We say that $\pi$ changes direction at position $i$ if either $\pi_{i-1}<\pi_i>\pi_{i+1}$, or $\pi_{i-1}>\pi_i<\pi_{i+1}$, where $i\in\{2,3,\ldots,n-1\}$. We say that $\pi$ has $k$ [*alternating runs*]{} if there are $k-1$ indices $i$ where $\pi$ changes direction (see [@Sloane A059427]). Let ${{\rm run\,}}(\pi)$ denote the number of alternating runs of $\pi$. For example, ${{\rm run\,}}(214653)=3$. There is a large literature devoted to the distribution of alternating runs. The reader is referred to [@CW08; @Ma122] for recent results on this subject.
We can now conclude the following result.
For $n\geq 1$, we have $$\label{Cnx:run}
C_n(x)=\sum_{\pi\in\Phi({\mathcal{Q}_n})}x^{{{\rm run\,}}(\pi)}.$$
Define $$\begin{aligned}
S_1& =\{\pi\in \Phi({\mathcal{Q}_n}): {{\rm lpk\,}}(\pi)={{\rm ipk\,}}(\pi)\}, \\
S_2& =\{\pi\in \Phi({\mathcal{Q}_n}): {{\rm lpk\,}}(\pi)={{\rm ipk\,}}(\pi)+1\}.\end{aligned}$$ Then $\Phi({\mathcal{Q}_n})$ can be partitioned into subsets $S_1$ and $S_2$.
From , we have $$\begin{aligned}
(1+x)C_n(x)& =\sum_{\pi\in\Phi({\mathcal{Q}_n})}x^{2{{\rm ipk\,}}(\pi)+1}+\sum_{\pi\in\Phi({\mathcal{Q}_n})}x^{2{{\rm lpk\,}}(\pi)}\\
& =x\sum_{\pi\in S_1}x^{{{\rm ipk\,}}(\pi)+{{\rm lpk\,}}(\pi)}+\sum_{\pi\in S_2}x^{{{\rm ipk\,}}(\pi)+{{\rm lpk\,}}(\pi)}+\sum_{\pi\in S_1}x^{{{\rm ipk\,}}(\pi)+{{\rm lpk\,}}(\pi)}+\\
&x\sum_{\pi\in S_2}x^{{{\rm ipk\,}}(\pi)+{{\rm lpk\,}}(\pi)}\\
&=(1+x)\sum_{\pi\in S_1}x^{{{\rm ipk\,}}(\pi)+{{\rm lpk\,}}(\pi)}+(1+x)\sum_{\pi\in S_2}x^{{{\rm ipk\,}}(\pi)+{{\rm lpk\,}}(\pi)}.\end{aligned}$$ Thus $$C_n(x)=\sum_{\pi\in\Phi({\mathcal{Q}_n})}x^{{{\rm ipk\,}}(\pi)+{{\rm lpk\,}}(\pi)}.$$ Note that all $\pi\in\Phi({\mathcal{Q}_n})$ ends with a descent, i.e., $\pi_{2n-1}>\pi_{2n}$. Hence follows from the fact that ${{\rm run\,}}(\pi)={{\rm ipk\,}}(\pi)+{{\rm lpk\,}}(\pi)$.
Concluding remarks
==================
It follows from and Theorem \[thm:01\], we have $$\label{problem:01}
\sum_{e\in I_{n,k}}x^{{{\rm asc\,}}(e)}=\sum_{\sigma\in {\mathcal{Q}_n}(k)}x^{{{\rm ap\,}}(\sigma)}.$$ Combining and Theorem \[thm:01\], we have $$\label{problem:02}
\sum_{\pi\in{\mathfrak{S}_n}}x^{{{\rm exc\,}}(\pi)}k^{n-{{\rm cyc\,}}(\pi)}=\sum_{\sigma\in {\mathcal{Q}_n}(k)}x^{{{\rm ap\,}}(\sigma)}.$$ It would be interesting to present a combinatorial proof of or .
[14]{} M. Bóna,Real zeros and normal distribution for statistics on Stirling permutations defined by Gessel and Stanley, 23 (2008) 401–406.
F. Brenti,A class of $q$-symmetric functions arising from plethysm, 91 (2000) 137–170.
E.R. Canfield, H. Wilf, Counting permutations by their alternating runs, 115 (2008), 213–225.
K. Dilks, T.K. Petersen, J.R. Stembridge, Affine descents and the Steinberg torus, 42 (2009), 423–444.
D. Foata, M. Schützenberger, Théorie Géométrique des Polynômes Euleriens, Lecture Notes in Mathematics, vol. 138, Springer-Verlag, Berlin-New York, 1970.
I. Gessel and R.P. Stanley, Stirling polynomials, 24 (1978) 25–33.
S. Janson, M. Kuba and A. Panholzer, Generalized Stirling permutations, families of increasing trees and urn models, 118 (2011) 94–114.
M. Kuba, A. Panholzer, Enumeration formulae for pattern restricted Stirling permutations, 312 (2012) 3179–3194.
S.-M. Ma, Derivative polynomials and enumeration of permutations by number of interior and left peaks, 312 (2012), 405–412.
S.-M. Ma, Enumeration of permutations by number of alternating runs, 313 (2013), 1816–1822.
J.B. Remmel, A.T. Wilson, Block patterns in Stirling permutations, http://arxiv.org/abs/arXiv:1402.3358. C.D. Savage and G. Viswanathan, The $1/k$-Eulerian polynomials, 19 (2012) \#P9.
N.J.A. Sloane, The On-Line Encyclopedia of Integer Sequences, published electronically at http://oeis.org, 2010.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
address: 'Hanoi, Vietnam.'
author:
- '§ç Minh Chİ'
date: 1979
title: 'Ph¬ng tr×nh Nh©n qu¶'
---
25.5truecm
pnumwidth[2em]{}
specialpagefalse oddhead[§ç Minh ChİPh¬ng tr×nh Nh©n qu¶]{} evenheadoddhead oddfoot[@font@page=1 ]{} evenfootoddfoot
[2]{} NÕu thÕ giíi lµ [*thèng nhÊt*]{} th× nã thèng nhÊt trong mèi quan hÖ [*nh©n qu¶*]{} vµ sù thèng nhÊt chØ cã thÓ ®îc biÓu hiÖn trong ı nghÜa Êy mµ th«i.
Theo tinh thÇn ®ã, sù ngÉu nhiªn, nÕu nh thËt sù cã c¸i g× ®ã lµ ngÉu nhiªn, còng chØ lµ s¶n phÈm cña sù tÊt yÕu.
Bëi thÕ giíi lµ thèng nhÊt trong mèi quan hÖ nh©n qu¶, kh«ng mét c¸i g× cña thÕ giíi n»m ngoµi mèi quan hÖ Êy, nªn ta cã thÓ chia toµn bé thÕ giíi thµnh hai tËp hîp: tËp hîp $A$ bao gåm tÊt c¶ nh÷ng g× ®îc coi lµ nguyªn nh©n, tËp hîp $B$ bao gåm tÊt c¶ nh÷ng g× lµ hÖ qu¶.
Chóng ta h·y lo¹i bá khái c¶ hai tËp hîp tÊt c¶ c¸c phÇn tö gièng nhau. Nh thÕ sÏ cã 4 kh¶ n¨ng sau:
1. C¶ hai tËp hîp $A$, $B$ ®Òu trë thµnh c¸c tËp hîp rçng, tøc lµ kh«ng cã nguyªn nh©n thuÇn tuı còng nh hÖ qu¶ thuÇn tuı. Nãi c¸ch kh¸c, thÕ giíi kh«ng cã më ®Çu, còng kh«ng cã kÕt côc cuèi cïng.
2. $A$ kh«ng rçng, $B$ rçng. Nh vËy tån t¹i nguyªn nh©n thuÇn tuı - thÕ giíi cã b¾t ®Çu nhng kh«ng cã kÕt côc cuèi cïng.
3. $A$ rçng, $B$ kh«ng rçng. Kh«ng cã nguyªn nh©n thuÇn tuı nhng cã hÖ qu¶ thuÇn tuı. ThÕ giíi kh«ng cã më ®Çu nhng cã kÕt côc cuèi cïng.
4. C¶ $A$, c¶ $B$ ®Òu kh«ng rçng. ThÕ giíi cã më ®Çu vµ cã kÕt côc cuèi cïng.
ChØ mét trong 4 kh¶ n¨ng trªn lµ ®óng víi hiÖn thùc. §ã lµ kh¶ n¨ng nµo vµ ®iÒu ®ã phô thuéc vµo c¸i g×?
Mét ®éng lùc ®Çy bİ Èn lu«n lu«n th«i thóc con ngêi t×m kiÕm nguyªn nh©n cho mäi hiÖn tîng vµ sù vËt. Dßng t tëng duy t©m (kh¸ch quan vµ chñ quan) cho r»ng ı niÖm tuyÖt ®èi, tinh thÇn tèi cao hay ®Êng s¸ng t¹o, thîng ®Õ,... lµ nguyªn nh©n tèi cao, lµ nguyªn nh©n cña tÊt c¶. Dßng t tëng duy vËt l¹i cho r»ng vËt chÊt míi lµ nguån gèc cña tÊt c¶, lµ c¸i ph¶i cã tríc nhÊt.
HiÖn tr¹ng ®ã lµ [*m©u thuÉn*]{}!
NÕu qu¶ thùc tån t¹i mét nguyªn nh©n tèi cao, th× ®ã lµ [*sù kh¸c biÖt*]{}!
Qu¶ vËy, nÕu kh«ng tån t¹i sù kh¸c biÖt, th× sÏ kh«ng tån t¹i bÊt cø c¸i g×, kÓ c¶ dßng t tëng duy t©m víi nh÷ng ı niÖm vµ tinh thÇn cña nã, kÓ c¶ dßng t tëng duy vËt víi c¸c c¬ së vËt chÊt cña nã. Tãm l¹i, nÕu kh«ng cã sù kh¸c biÖt th× kh«ng cã thÕ giíi nµy.
ThÕ nhng, nÕu kh¸c biÖt lµ nguyªn nh©n tèi cao, tøc lµ nguyªn nh©n cña mäi nguyªn nh©n th× nã ph¶i lµ nguyªn nh©n cña chİnh nã n÷a, hay nãi kh¸c h¬n, nã lµ hÖ qu¶ cña chİnh nã.
Chóng ta ®· thõa nhËn sù tån t¹i cña kh¸c biÖt, ®iÒu ®ã cã nghÜa lµ chóng ta ngÇm thõa nhËn [*tİnh b¶o toµn t¬ng ®èi*]{} cña nã: Qu¶ vËy, nÕu lóc nµy b¹n lµ nhµ duy vËt th× b¹n kh«ng thÓ còng lµ kÎ duy t©m ®îc n÷a. Qu¶ thŞ, khi chøa c« TÊm ë bªn trong, kh«ng ph¶i lµ qu¶ thŞ theo ®óng nghÜa cña nã. C¸i bµn kia, chõng nµo nã cßn lµ c¸i bµn, th× nã kh«ng thÓ lµ trang giÊy mµ b¹n ®ang ®äc ®îc!.
[**C©u chuyÖn x¶y ra trong h×nh häc**]{}
Chóng ta trë l¹i mét c©u chuyÖn ®· cò: cuéc tranh c·i xung quanh hÖ tiªn ®Ò h×nh häc Euclid.
VÉn do chİnh c¸i ®éng lùc bİ Èn duy nhÊt Êy chi phèi mµ ngêi ta lu«n khao kh¸t t×m kiÕm c¸i “nguyªn nh©n tèi cao”. [ë]{} ®©y môc ®İch ®ã khiªm tèn h¬n, chØ giíi h¹n trong ph¹m vi h×nh häc, vµ ngêi ®Çu tiªn thùc hiÖn ®îc ®iÒu ®ã lµ Euclid.
Bèn tiªn ®Ò ®Çu cña «ng ®îc dÔ dµng chÊp nhËn v× chóng râ rµng vµ hiÓn nhiªn, nhng tiªn ®Ò thø n¨m, cßn gäi lµ ®Şnh ®Ò Euclid, ®· khiÕn ngêi ta nghi ngê vÒ b¶n chÊt tiªn ®Ò cña nã: “®Şnh ®Ò nµy phøc t¹p vµ kĞm hiÓn nhiªn”.
Trong suèt h¬n 20 thÕ kû c¸c nhµ to¸n häc mäi thêi ®¹i ®· g¾ng søc chøng minh nã chØ lµ mét ®Şnh lı. Nhng mäi cè g¾ng ®Òu v« İch. Vµ kÕt qu¶ lµ dÉn tíi sù ra ®êi mét h×nh häc míi: h×nh häc phi Euclid.
B»ng c¸ch ®ã, ngêi ta ®· kh¼ng ®Şnh r»ng, vÊn ®Ò nh thÕ lµ ®· ®îc gi¶i quyÕt: ®Şnh ®Ò Euclid ®óng lµ mét tiªn ®Ò, v× r»ng ®iÒu gi¶ thiÕt ngîc l¹i ®· dÉn tíi h×nh häc phi Euclid kh«ng cã m©u thuÉn néi t¹i.
Nhng... kÕt luËn nh vËy cã dÔ d·i qu¸ kh«ng?
Khi ngêi ta hoan hØ v× h×nh nh mäi viÖc ®· ®©u vµo ®Êy vµ môc ®İch ®Ò ra ®· ®îc thùc hiÖn: gi¶m ®Õn møc tèi thiÓu sè lîng tiªn ®Ò cña h×nh häc vµ lµm trong s¸ng chóng, th× trí trªu thay, mét tiªn ®Ò míi l¹i ®îc lĞn ®a thªm vµo mét c¸ch ®êng hoµng: tiªn ®Ò Lobachevski. Tiªn ®Ò nµy vµ tiªn ®Ò thø n¨m Euclid lo¹i trõ lÉn nhau!
Kh«ng mét ai nhËn thøc râ rµng vµ s©u s¾c m©u thuÉn nµy nghÜa lµ thÕ nµo. Nhng m©u thuÉn vÉn lµ m©u thuÉn: nã g©y ra bao cuéc tranh c·i vµ ph¶n b¸c kŞch liÖt, thËm chİ c¶ sù h»n häc.
Sau nµy, khi Bentrami ®· chøng minh ®îc sù ®óng ®¾n cña h×nh häc Lobachevski trªn mÆt gi¶ cÇu, sù ph¶n b¸c cã dŞu xuèng.
NÕu ngay tõ ®Çu, c¸c nhµ h×nh häc phi Euclid khi b¾t tay vµo x©y dùng h×nh häc cña m×nh, tuyªn bè râ víi ®éc gi¶ r»ng: ®èi tîng cña h×nh häc míi kh«ng ph¶i lµ mÆt ph¼ng Euclid mµ lµ mÆt gi¶ cÇu, kh«ng ph¶i lµ ®êng th¼ng Euclid mµ lµ ®êng th¼ng cña mÆt gi¶ cÇu th× cã lÏ ®· ch¼ng cã ai th¾c m¾c vµ ph¶n ®èi g× c¶!
ThËt ®¸ng tiÕc! hay kh«ng ®¸ng tiÕc lµ ®· kh«ng x¶y ra nh vËy?
Nhng ®iÒu ®¸ng tiÕc thËt sù lµ: toµn bé vÊn ®Ò l¹i kh«ng n»m trong c¸i ®· ®îc ®a ra vµ gi¶i quyÕt ngoµi s©n khÊu mµ ë hËu qu¶ cña nã trong hËu trêng. Bëi v×, cho dï h×nh häc phi Euclid cã tuyÖt ®èi ®óng ë ®©u ch¨ng n÷a, th× ®iÒu ®ã cã nghÜa lµ: còng cïng nh÷ng ®èi tîng Êy cña h×nh häc - mÆt ph¼ng vµ ®êng th¼ng Euclid - vËy mµ gi÷a chóng cã thÓ tån t¹i hai kiÓu quan hÖ lo¹i trõ nhau! ®îc diÔn ®¹t trong c¸c tiªn ®Ò Euclide vµ tiªn ®Ò Lobachevski.
Cã thÓ viÖn ®Õn c¸c lı lÏ nµy hoÆc kh¸c ®Ó Ğp m×nh chÊp nhËn ®iÒu khã chŞu nµy, nhng ®ã sÏ kh«ng ph¶i lµ sù trung thùc. [ë]{} ®©y, tİnh ®¬n trŞ nh©n qu¶ ®· bŞ ph¸ vì; ë ®©y tİnh b¶o toµn t¬ng ®èi cña sù kh¸c biÖt ®· bŞ lÉn lén tr¾ng ®en, cã nguy c¬ trang giÊy còng lµ c¸i bµn, vµ c¸i bµn lµ trang giÊy.
Ph¬ng ph¸p tiªn ®Ò ®îc sö dông réng r·i trong to¸n häc râ rµng ®· mang l¹i nhiÒu tiÖn lîi, nhng ph¬ng ph¸p ®ã chØ tèt khi tİnh ®¬n trŞ nh©n qu¶ ®îc b¶o ®¶m, khi chóng ta lu«n lu«n chó ı ®Ó kh«ng tíc ®o¹t mÊt ë c¸c ®èi tîng kh¶o s¸t ı nghÜa thùc, vËt lı cña chóng. NÕu ®a ra c¸c kiÓu quan hÖ cña c¸c ®èi tîng díi d¹ng c¸c tiªn ®Ò mµ bÊt kÓ ®Õn c¸c ®èi tîng - chñ nh©n thùc sù cña c¸c mèi quan hÖ ®ã th× rÊt cã thÓ, ë mét chç kh«ng ngê nhÊt, tİnh ®¬n trŞ nh©n qu¶ bŞ ph¸ vì vµ m©u thuÉn n¶y sinh.
Bëi v× c¸i mµ chóng ta thèng nhÊt víi nhau lµ: c¸c ®èi tîng lµ c¸i cã tríc, c¸c mèi quan hÖ cña c¸c ®èi tîng lµ hÖ qu¶ tÊt yÕu do sù tån t¹i ®ång thêi cña c¸c ®èi tîng g©y ra, chø kh«ng ph¶i ngîc l¹i.
NÕu ta cã mét tËp hîp c¸c ®èi tîng vµ ta muèn t×m kiÕm tÊt c¶ c¸c mèi quan hÖ cã thÓ gi÷a chóng b»ng ph¬ng ph¸p lËp luËn logic th× cã lÏ tríc hÕt vµ İt ra ta ph¶i biÕt mèi quan hÖ néi t¹i cña tõng lo¹i ®èi tîng.
C¸c mèi quan hÖ néi t¹i quyÕt ®Şnh b¶n chÊt cña ®èi tîng, ®Õn lît m×nh b¶n chÊt cña c¸c ®èi tîng quyÕt ®Şnh c¸c mèi quan hÖ cã thÓ cã gi÷a chóng, vµ tÊt nhiªn gi÷a chóng kh«ng thÓ ®ång thêi tån t¹i c¸c mèi quan hÖ lo¹i trõ nhau.
Mèi quan hÖ néi t¹i, theo c¸ch nãi cña c¸c triÕt gia, lµ tİnh tù th©n cña sù vËt. Khoa häc ngµy nay chİnh lµ ®ang t×m kiÕm c¸i tİnh tù th©n ®ã cña sù vËt trªn c¶ hai híng, réng h¬n vµ c¬ b¶n h¬n.
Trë l¹i m¹ch cña c©u chuyÖn, nh chóng ta ®· thÊy, còng chİnh nh÷ng ®èi tîng Êy cña h×nh häc Euclid mµ l¹i cã hai kiÓu quan hÖ lo¹i trõ nhau, ®iÒu ®ã ph¶i ®îc hiÓu nh thÕ nµo?
ChØ cã thÓ lµ hÖ tiªn ®Ò Euclid lµ cha ®Çy ®ñ theo nghÜa: sù hiÓu vÒ c¸c ®èi tîng cña h×nh häc nµy lµ cha hoµn h¶o. B¶n th©n Euclid còng ®· tõng ®a ra c¸c ®Şnh nghÜa vÒ c¸c ®èi tîng h×nh häc cña «ng, nhng nh c¸c nhµ to¸n häc hiÖn ®¹i ®· phª ph¸n lµ: “lóng tóng”, “nÆng tİnh trùc gi¸c”. Theo hä, c¸c ®èi tîng xuÊt ph¸t cña h×nh häc lµ kh«ng ®Şnh nghÜa ®îc vµ chØ ®¬n thuÇn gäi chóng lµ ®iÓm, ®êng, mÆt v.v. v× lı do t«n träng lŞch sö mµ th«i.
Nhng, c¸c ®èi tîng h×nh häc cßn cã c¸c tªn gäi kh¸c n÷a: c¸c kh«ng gian “kh«ng chiÒu”, “mét chiÒu”, “hai chiÒu”, vµ “ba chiÒu”. (Kh«ng gian “kh«ng chiÒu” tøc ®iÓm, lµ do t¸c gi¶ ®a vµo cho “trän bé”.)
Ta cã thÓ hái r»ng, c¸c ®èi tîng nµy cã thÓ tù tån t¹i ®éc lËp víi nhau ®îc kh«ng, nÕu cã th× v× sao chóng cã thÓ quan hÖ víi nhau ®îc?
LÇn theo m¹ch logic cña sù viÖc, ta thÊy r»ng c¸c kh¸i niÖm vÒ nh÷ng ®èi tîng nµy n¶y sinh ra tõ kinh nghiÖm thu ®îc qua ho¹t ®éng thùc tiÔn cña con ngêi trong Tù nhiªn chø kh«ng ph¶i do bÈm sinh, tù chóng cã s½n trong ®Çu cña chóng ta (Bëi vËy, ta kh«ng nªn t¸ch rêi chóng khái trùc gi¸c, kh«ng nªn tíc ®o¹t kh¶ n¨ng h×nh dung ra chóng; ®iÒu ®ã phi lı biÕt bao!)
Nh×n nhËn ë møc ®é s©u h¬n, cã thÓ thÊy r»ng, kh«ng ph¶i tÊt c¶ trong sè c¸c ®èi tîng h×nh häc ®Òu cã thÓ tån t¹i ®éc lËp, mµ bÊt kú mét kh«ng gian $n$ chiÒu nµo còng lµ giao cña hai kh«ng gian kh¸c cã chiÒu lín h¬n mét ®¬n vŞ $n+1$.
VËy lµ, h×nh nh chóng ta cã ®Şnh nghÜa: ®iÓm lµ giao cña hai ®êng, ®êng lµ giao cña hai mÆt, mÆt lµ giao cña hai khèi, cßn khèi ... lµ giao cña nh÷ng kh«ng gian nµo n÷a?
Tuy nhiªn, trong h×nh häc, nhê kh¶ n¨ng h×nh dung cña n·o, chóng mÆc nhiªn trë thµnh c¸c ®èi tîng ®éc lËp, vµ ®Ó gi¶n tiÖn, ta sÏ gäi chóng lµ c¸c thùc thÓ kh«ng gian.
C¸c ®èi tîng ®¬n gi¶n nhÊt cña h×nh häc lµ c¸c thùc thÓ ®ång nhÊt. §ã lµ c¸c thùc thÓ, mµ nãi mét c¸ch gi¶n ®¬n, khi ta dŞch chuyÓn trªn chóng theo tÊt c¶ c¸c bËc tù do cã thÓ cã cña chóng, ta kh«ng thÓ ph¸t hiÖn ®îc bÊt kú sù kh¸c biÖt néi t¹i nµo trong chóng.
C¸c ®èi tîng cña h×nh häc Euclid lµ mét bé phËn cña tËp hîp c¸c thùc thÓ ®ång nhÊt. NÕu ta x©y dùng hÖ tiªn ®Ò chØ riªng cho bé phËn nµy th× râ rµng hÖ ®ã sÏ kh«ng tæng qu¸t.
HÖ tiªn ®Ò dïng cho c¸c ®èi tîng kh«ng gian ®ång nhÊt chİnh lµ hÖ tiªn ®Ò dïng cho mÆt cÇu. H×nh häc Euclid chØ lµ trêng hîp giíi h¹n cña h×nh häc tæng qu¸t nµy.
§èi víi mÆt cÇu, tøc lµ mÆt ®ång nhÊt ë d¹ng tæng qu¸t, tån t¹i ®Şnh ®Ò sau: hai ®êng th¼ng bÊt kú kh«ng trïng nhau (®êng th¼ng lµ ®êng ®ång nhÊt chia mÆt chøa nã thµnh hai phÇn b»ng nhau) bao giê còng c¾t nhau t¹i hai ®iÓm vµ hai ®iÓm ®ã chia ®«i mçi ®êng.
Cã thÓ ph¸t biÓu kh¸c h¬n: hai ®iÓm bÊt kú trªn mét mÆt ®ång nhÊt chØ thuéc vÒ mét ®êng th¼ng duy nhÊt còng trªn mÆt ®ã nÕu chóng kh«ng chia ®êng nµy thµnh hai phÇn b»ng nhau.
[¸]{}p dông ®Şnh ®Ò nµy cho mÆt ph¼ng Euclid, lµ trêng hîp giíi h¹n, ta thÊy ngay r»ng ®©y chİnh lµ néi dung cña tiªn ®Ò thø nhÊt Euclid: qua hai ®iÓm chØ cã thÓ kÎ ®îc mét ®êng th¼ng duy nhÊt mµ th«i. Qu¶ vËy, hai ®iÓm bÊt kú trong ph¹m vi kh¶o s¸t ®îc cña mÆt ph¼ng Euclid chØ thuéc vÒ mét ®êng th¼ng duy nhÊt v× chóng kh«ng chia ®êng ®ã thµnh hai phÇn b»ng nhau.
Nh vËy, cã thÓ nãi r»ng, c¸ch ph¸t biÓu ®Şnh ®Ò 5 Euclid lµ kh«ng chİnh x¸c ngay tõ ®Çu, bëi v× bÊt kú hai ®êng th¼ng nµo cña mÆt ®ång nhÊt ®· cho lu«n lu«n c¾t nhau t¹i hai ®iÓm vµ chia ®«i nhau. Trªn mÆt ph¼ng Euclid ta chØ thÊy hoÆc mét giao ®iÓm cña chóng cßn ®iÓm kia ë ngoµi v« cïng; hoÆc lµ ta kh«ng thÊy ®iÓm nµo c¶ - chóng ®Òu n»m ë ngoµi v« cùc. Trong trêng hîp ®ã hai ®êng th¼ng ®îc gäi lµ song song (biÓu kiÕn) víi nhau.
C¸c c¸ch ph¸t biÓu t¬ng ®¬ng cña ®Şnh ®Ò 5, sau khi ®îc chİnh x¸c ho¸ theo tinh thÇn cña nhËn xĞt trªn, ®Òu cã thÓ ®îc chøng minh nh mét ®Şnh lı.
Mét tİnh chÊt rÊt quan träng cña c¸c thÓ kh«ng gian lµ: mét thùc thÓ kh«ng gian bÊt kú chØ cã thÓ bŞ chøa trong mét thùc thÓ kh«ng gian kh¸c ®ång chiÒu, ®ång ®é cong hoÆc cã chiÒu lín h¬n nhng ®é cong kh«ng lín h¬n.
§iÒu nµy h×nh nh ®· qu¸ hiÓn nhiªn: hai ®êng trßn cã ®é cong kh¸c nhau th× kh«ng thÓ chøa trong nhau ®îc; mét mÆt trßn cã ®é cong lín h¬n kh«ng thÓ chøa mét ®êng trßn cã ®é cong nhá h¬n...
T¬ng tù nh vËy, hai kh«ng gian (khèi) cã ®é cong kh¸c nhau th× kh«ng thÓ chøa trong nhau ®îc. (Ta h·y trë l¹i vİ dô vÒ qu¶ thŞ vµ c« TÊm, vÒ c¸i bµn vµ trang giÊy...). §é cong ë ®©y t¬ng øng víi mét ®¹i lîng nµo ®ã ®Æc trng cho mèi quan hÖ néi t¹i cña ®èi tîng kh¶o s¸t.
[**M©u thuÉn, sinh ra trªn c¬ së cña sù kh¸c biÖt,**]{}\
[**lµ nguån ®éng lùc cña tÊt c¶**]{}
VÒ thùc chÊt, Tù nhiªn lµ mét tËp hîp cña c¸c [*kh¼ng ®Şnh*]{} vµ [*phñ ®Şnh*]{}.
VËy Tù nhiªn kh¼ng ®Şnh nh÷ng g× vµ phñ ®Şnh nh÷ng g×?
Nh÷ng bİ mËt ®ã ngµy cµng ®îc khoa häc kh¸m ph¸, ph¸t hiÖn vµ trong cuéc t×m kiÕm ®ã, nÕu kh«ng kÓ ®Õn nguån gèc ®éng lùc cña nã, lËp luËn logic ®ãng vai trß to lín.
Nhng c¸i mµ ta gäi lµ logic ph¶i ch¨ng kh«ng ph¶i lµ mét chuçi nh÷ng kh¼ng ®Şnh vµ phñ ®Şnh ë c¸c cÊp bËc vµ tæ hîp kh¸c nhau?
VËy th× khi nµo c¸i nµy ®îc kh¼ng ®Şnh, cßn c¸i kia th× kh«ng?
Bëi v× t duy còng chØ lµ mét hiÖn tîng cña Tù nhiªn, nªn quy luËt kh¼ng ®Şnh vµ phñ ®Şnh cña t duy còng lµ quy luËt kh¼ng ®Şnh vµ phñ ®Şnh cña Tù nhiªn. Nãi c¸ch kh¸c, quy luËt kh¼ng ®Şnh vµ phñ ®Şnh cña Tù nhiªn ®· ph¶n ¸nh vµ thÓ hiÖn qua chİnh quy luËt kh¼ng ®Şnh vµ phñ ®Şnh cña t duy.
Quy luËt ®ã lµ: [*c¸i g× kh«ng cã m©u thuÉn néi t¹i th× ®îc kh¼ng ®Şnh, c¸i g× chøa m©u thuÉn néi t¹i th× bŞ phñ ®Şnh*]{}.
Sù kh¼ng ®Şnh (nÕu nh×n vÒ phİa tríc cña qu¸ tr×nh) hay sù phñ ®Şnh (nÕu nh×n vÒ phİa sau cña tiÕn tr×nh) ®Òu cã ®İch cuèi cïng lµ ®¹t tíi vµ kÕt thóc ë mét kh¼ng ®Şnh míi.
Chóng ta h·y lÊy mét líp nh÷ng kh¸i niÖm rÊt gÇn nhau lµ: cã, tån t¹i, b¶o toµn, kh¼ng ®Şnh.
§èi lËp l¹i víi chóng lµ líp kh¸i niÖm phñ ®Şnh cña chóng: kh«ng, kh«ng tån t¹i, kh«ng b¶o toµn, phñ ®Şnh.
Chóng n»m trong sè nh÷ng kh¸i niÖm tæng qu¸t nhÊt, c¬ b¶n nhÊt, bëi trong bÊt kú hiÖn tîng nµo cña Tù nhiªn: c¶m gi¸c, suy nghÜ, vËn ®éng, biÕn ®æi v.v. ®Òu lu«n lu«n cã sù biÓu hiÖn cña chóng.
Nhng ho¸ ra lµ søc m¹nh cña hai líp kh¸i niÖm nµy kh«ng t¬ng ®¬ng nhau (vµ ®ã lµ mét ®iÒu thËt may m¾n).
Chóng ta h·y thiÕt lËp kh¼ng ®Şnh sau gäi lµ [*kh¼ng ®Şnh*]{} $A$:
“Tån t¹i tÊt c¶, cã tÊt c¶, b¶o toµn tÊt c¶, kh¼ng ®Şnh tÊt c¶”.
Cßn [*kh¼ng ®Şnh*]{} $B$ cã néi dung ngîc l¹i:
“Kh«ng cã c¸i g× hÕt, kh«ng tån t¹i bÊt cø c¸i g×, kh«ng b¶o toµn c¸i g× c¶, phñ ®Şnh tÊt c¶”.
Kh¼ng ®Şnh [*kh¼ng ®Şnh*]{} $A$, tøc lµ phñ ®Şnh [*kh¼ng ®Şnh*]{} $B$. Vµ ngîc l¹i.
[*Kh¼ng ®Şnh*]{} $B$ nãi r»ng:
- Kh«ng cã c¸i g× hÕt, tøc lµ kh«ng cã chİnh [*kh¼ng ®Şnh*]{} $B$.
- Kh«ng tån t¹i c¸i g×, tøc lµ kh«ng tån t¹i b¶n th©n [*kh¼ng ®Şnh*]{} $B$.
- Kh«ng b¶o toµn c¸i g×, vËy lµ chİnh [*kh¼ng ®Şnh*]{} $B$ còng kh«ng ®îc b¶o toµn.
- Phñ ®Şnh tÊt c¶, vËy lµ phñ ®Şnh chİnh [*kh¼ng ®Şnh*]{} $B$.
Tãm l¹i, [*kh¼ng ®Şnh*]{} $B$ chøa m©u thuÉn néi t¹i. Nã tù phñ ®Şnh chİnh m×nh. Tù phñ ®Şnh m×nh, [*kh¼ng ®Şnh*]{} $B$ mÆc nhiªn kh¼ng ®Şnh [*kh¼ng ®Şnh*]{} $A$, vµ ®iÒu ®ã cã nghÜa lµ: kh«ng tån t¹i h v« hay sù trèng rçng tuyÖt ®èi, vµ chİnh bëi lÏ ®ã mµ thÕ giíi ®· ®îc sinh ra!
Cßn [*kh¼ng ®Şnh*]{} $A$ kh¼ng ®Şnh tÊt c¶, kÓ c¶ chİnh nã lÉn [*kh¼ng ®Şnh*]{} $B$, nhng [*kh¼ng ®Şnh*]{} $B$ tù phñ ®Şnh chİnh m×nh, nªn [*kh¼ng ®Şnh*]{} $A$ kh«ng chøa m©u thuÉn néi t¹i.
Nh vËy, trong khu«n khæ cña [*kh¼ng ®Şnh*]{} $A$ nh÷ng g× kh«ng tù phñ ®Şnh th× ®îc kh¼ng ®Şnh.
[**C¸i g× c¬ b¶n h¬n**]{}
Bèn kh¸i niÖm rÊt quan träng cña tri thøc lµ: [*thêi gian*]{}, [*kh«ng gian*]{}, [*vËt chÊt*]{} vµ [*vËn ®éng*]{}. Chóng kh¸c biÖt nhau, nhng ph¶i ch¨ng chóng b×nh ®¼ng víi nhau vµ cã thÓ tån t¹i ®éc lËp víi nhau?
Ta h·y b¾t ®Çu tõ thêi gian. Nã lµ mét thùc thÓ ch¨ng? nã cã thÓ tån t¹i ®éc lËp, t¸ch rêi khái kh«ng gian, vËt chÊt vµ vËn ®éng ch¨ng?. HiÓn nhiªn lµ kh«ng. ChØ cÇn c¸ch ly thêi gian khái vËn ®éng lµ kh¸i niÖm vÒ nã sÏ mÊt mäi ı nghÜa, thêi gian sÏ chÕt. Kh¸i niÖm vËn ®éng cã tİnh ®éc lËp cao h¬n so víi kh¸i niÖm thêi gian.
VËy th× thêi gian cha ph¶i lµ c¸i ®Çu tiªn. Nã kh«ng tù tån t¹i ®îc nªn nã chØ cã thÓ lµ hÖ qu¶ cña nh÷ng c¸i cßn l¹i.
VËn ®éng còng kh«ng ph¶i lµ c¸i c¬ b¶n ®Çu tiªn. Nã kh«ng thÓ tù tån t¹i t¸ch rêi khái vËt chÊt vµ kh«ng gian. Thùc chÊt, vËn ®éng chØ lµ mét biÓu hiÖn cña mèi quan hÖ gi÷a vËt chÊt vµ kh«ng gian.
VËy th× gi÷a hai c¸i cßn l¹i, vËt chÊt vµ kh«ng gian, c¸i nµo c¬ b¶n h¬n, c¸i nµo cã tríc, hay chóng b×nh ®¼ng víi nhau vµ cïng do mét c¸i g× ®ã c¬ b¶n h¬n sinh ra? Cã lÏ ®Æt vÊn ®Ò nh vËy lµ thõa, bëi v× còng nh thêi gian vµ vËn ®éng, vËt chÊt kh«ng thÓ t¸ch rêi vµ tån t¹i ngoµi kh«ng gian. H·y lÊy mét biÓu hiÖn cô thÓ cña vËt chÊt, ch¼ng h¹n, trang giÊy nµy. Nã tån t¹i kh«ng ph¶i chØ nhê b¶n th©n nã, mµ cßn do sù tån t¹i ®ång thêi cña kh«ng gian bao quanh nã (hay chøa nã), lµm cho trang giÊy vÉn cßn lµ trang giÊy.
Râ rµng lµ vËt chÊt còng thuéc vÒ ph¹m trï kh«ng gian vµ nã cã thÓ lµ c¸i g× kh¸c h¬n n÷a nÕu kh«ng ph¶i lµ chİnh kh«ng gian cã mèi quan hÖ néi t¹i kh¸c víi c¸i kh«ng gian th«ng thêng mµ ta vÉn hiÓu?!
Nhng khi ®ã, theo tİnh chÊt cña c¸c thÓ kh«ng gian ®· nãi ë trªn, ®iÒu nµy lµ m©u thuÉn: hai kh«ng gian ®ång chiÒu cã ®é cong (tøc mèi quan hÖ néi t¹i) kh¸c nhau th× kh«ng thÓ chøa trong nhau ®îc!
VËy th× hoÆc lµ chóng ta ®· sai: hiÓn nhiªn lµ cã thÓ ®Æt trïng khİt lªn nhau hai vßng trßn cã b¸n kİnh kh¸c nhau. HoÆc lµ Tù nhiªn ®· sai: ®Æt c¸c kh«ng gian kh¸c nhau vµo trong nhau, bÊt chÊp m©u thuÉn.
Vµ m©u thuÉn sinh ra do ®iÒu ®ã lµ ®éng lùc cña vËn ®éng, vËn ®éng ®Ó tho¸t ra khái m©u thuÉn.
Nh vËy, cã thÓ nãi vËt chÊt lµ tÊt c¶ c¸c thÓ kh«ng gian cã ®é cong (tøc mèi quan hÖ néi t¹i) nµo ®ã.
Nhng tõ ®©u mµ sinh ra c¸c thÓ kh«ng gian nµy vµ lµm sao chóng cã thÓ tån t¹i ®îc?
Chóng ta h·y tëng tîng r»ng tÊt c¶ biÕn mÊt hÕt: vËt chÊt, kh«ng gian,... vµ nãi chung biÕn mÊt hÕt mäi sù kh¸c biÖt cã thÓ cã.
Khi ®ã cßn l¹i c¸i g×?
Ch¼ng cßn c¸i g× c¶!
Nhng ®ã chİnh lµ c¸i duy nhÊt cßn l¹i!
Râ rµng c¸i duy nhÊt nµy lµ v« h¹n vµ ®ång nhÊt ë “kh¾p mäi n¬i”. NÕu kh«ng thÕ, sÏ vi ph¹m ®ßi hái cña chóng ta.
B©y giê chóng ta ®ßi hái mét ®iÒu tiÕp theo: ngay c¶ c¸i duy nhÊt nµy còng biÕn mÊt nèt! SÏ cßn l¹i c¸i g× sau nã?
Kh«ng ph¶i vÊt v¶ l¾m, ta thÊy ngay r»ng c¸i ®Õn thay thÕ cho nã l¹i lµ chİnh nã! V× vËy ta h·y gäi c¸i ®ã lµ [*kh«ng gian tuyÖt ®èi*]{}.
Kh«ng gian tuyÖt ®èi cã thÓ biÕn mÊt vµo chİnh nã, nãi kh¸c h¬n, sù phñ nhËn nã dÉn tíi sù kh¼ng ®Şnh chİnh nã. §iÒu ®ã cã nghÜa lµ, c¸i kh«ng gian tuyÖt ®èi cña chóng ta cã thÓ tù tån t¹i mµ kh«ng cÇn nhê ®Õn ai c¶. Nã lµ c¸i c¬ b¶n ®Çu tiªn.
Nã còng lµ “nguyªn nh©n tèi cao” n÷a. V× tr¸i víi mäi ı muèn cña ai ®ã, nã vÉn chøa ®ùng sù kh¸c biÖt.
Qu¶ vËy, trong c¸i duy nhÊt ®ã kh«ng chøa c¸i g× c¶, vËy mµ vÉn cã: c¸i [*Kh«ng*]{}! c¸i [*Kh«ng*]{} chøa trong c¸i [*Cã*]{}, c¸i [*Kh«ng*]{} t¹o nªn c¸i [*Cã*]{}. Cã, nhng kh«ng lµ g× c¶!
[ë]{} ®©y, sù phñ ®Şnh còng lµ sù kh¼ng ®Şnh, c¸i [*Kh«ng*]{} còng lµ c¸i [*Cã*]{}, vµ ngîc l¹i. M©u thuÉn néi t¹i cña tr¹ng th¸i nµy lµ lín v« h¹n.
DiÔn ®¹t mét c¸ch to¸n häc h¬n ®iÒu ®ã: kh«ng gian tuyÖt ®èi cña chóng ta cã ®é cong b»ng kh«ng. Trong kh«ng gian nµy tån t¹i c¸c ®iÓm mµ ®é cong cña chóng lµ v« cïng. Sù kh¸c biÖt nµy lín v« h¹n vµ do ®ã m©u thuÉn sinh ra còng lín v« h¹n.
Tù nhiªn kh«ng muèn tån t¹i trong tr¹ng th¸i m©u thuÉn nh vËy. Nã tù t×m c¸ch gi¶i quyÕt, vµ kÕt qu¶ lµ, thÕ giíi v× thÕ mµ ®îc sinh ra.
Nh vËy, l¹i mét lÇn n÷a, c¸i ch©n lı m¬ hå mµ ai còng quen thuéc nÕu ®· qua c¸c trêng phæ th«ng lµ: “VËt chÊt kh«ng tù nhiªn sinh ra (tøc lµ kh«ng sinh ra tõ h v«), kh«ng tù nhiªn mÊt ®i, lu«n lu«n vËn ®éng vµ chuyÓn ho¸ tõ d¹ng nµy sang d¹ng kh¸c”, nay cÇn ®îc kh¼ng ®Şnh l¹i r»ng: “vËt chÊt” ®óng lµ sinh ra tõ “kh«ng cã g×”. Nhng kh«ng ph¶i v« cí mµ nh vËy. §éng lùc khiÕn nã sinh ra còng lµ ®éng lùc khiÕn nã tån t¹i, biÕn ®æi, vµ vËn ®éng.
[**BiÓu diÔn m©u thuÉn díi d¹ng ®Şnh lîng:**]{}\
[**[Ph¬ng tr×nh Nh©n qu¶]{}**]{}
Mäi m©u thuÉn ®Òu ph¸t sinh bëi sù tån t¹i ®ång thêi cña hai kh¼ng ®Şnh lo¹i trõ nhau.
§iÒu ®ã ®îc biÓu diÔn nh sau: $$M = \left\{ \begin{array}{ll}
A \neq B & \mbox{\it - Kh¼ng ®Şnh } K_1 \\
A = B & \mbox{\it - Kh¼ng ®Şnh } K_2
\end{array}\right.$$ hoÆc $$M = \left\{ \begin{array}{ll}
A = A & \mbox{\it - Kh¼ng ®Şnh } K_1 \\
A \neq A & \mbox{\it - Kh¼ng ®Şnh } K_2
\end{array}\right.$$
Râ rµng lµ m©u thuÉn sÏ cµng gay g¾t nÕu nh møc ®é phñ nhËn lÉn nhau cña hai kh¼ng ®Şnh cµng lín. Nhng møc ®é phñ nhËn lÉn nhau cña hai kh¼ng ®Şnh chØ cã thÓ ®îc ®¸nh gi¸ b»ng møc ®é kh¸c biÖt gi÷a hai kh¼ng ®Şnh ®ã. Bëi vËy ta cã thÓ biÓu diÔn m©u thuÉn mét c¸ch tîng trng nh sau: $M = [K_1 - K_2]$.
M©u thuÉn ®îc gi¶i quyÕt tøc lµ hiÖu $[K_1 - K_2]$ sÏ gi¶m tíi kh«ng. §iÒu ®ã cã nghÜa lµ c¶ hai kh¼ng ®Şnh $K_1$ vµ $K_2$ ®Òu ph¶i biÕn ®æi nh thÕ nµo ®ã ®Ó ®¹t tíi vµ kÕt thóc ë mét kh¼ng ®Şnh míi $K_3$.
Nh vËy, c¸c hiÖu $[K_1 - K_3]$ vµ $[K_2 - K_3]$ phô thuéc vµo nh÷ng g×? Râ rµng c¸c hiÖu ®ã phô thuéc vµo søc b¶o toµn cña c¸c kh¼ng ®Şnh $K_1$ vµ $K_2$. Søc b¶o toµn cña kh¼ng ®Şnh nµo cµng lín th× sù kh¸c biÖt gi÷a nã vµ kh¼ng ®Şnh cuèi cïng cµng nhá.
VËy ®Õn lît m×nh, søc b¶o toµn cña mét kh¼ng ®Şnh nµo ®ã phô thuéc vµo g×?
Cã hai yÕu tè. 1) Phô thuéc vµo m©u thuÉn néi t¹i cña kh¼ng ®Şnh ®ã, m©u thuÉn néi t¹i cµng lín søc b¶o toµn cña kh¼ng ®Şnh cµng nhá. 2) Phô thuéc vµo m©u thuÉn míi, sinh ra do sù biÕn ®æi cña kh¼ng ®Şnh. M©u thuÉn nµy cµng lín, sù biÕn ®æi cña kh¼ng ®Şnh cµng bŞ c¶n trë vµ do ®ã søc b¶o toµn cña kh¼ng ®Şnh cµng lín.
Sù biÕn ®æi, mµ vËn ®éng lµ mét d¹ng cña nã, sinh ra do m©u thuÉn. Nãi ®óng h¬n, vËn ®éng lµ biÓu hiÖn cña sù gi¶i quyÕt m©u thuÉn.
[ë]{} trªn ta ®· nãi r»ng, m©u thuÉn cµng gay g¾t nÕu nh møc ®é kh¸c biÖt gi÷a hai kh¼ng ®Şnh lo¹i trõ nhau cµng lín. Chóng ta sÏ tiÕp tôc lµm ®Çy ®ñ h¬n kÕt luËn ®Şnh lîng nµy: m©u thuÉn cµng gay g¾t th× nhu cÇu gi¶i tho¸t khái nã cµng bøc thiÕt vµ do ®ã sù vËn ®éng, biÕn ®æi cña tr¹ng th¸i, tøc lµ cña m©u thuÉn, cµng quyÕt liÖt, nhanh chãng.
NÕu ta gäi sù quyÕt liÖt, hay ®é nhanh chãng biÕn ®æi cña m©u thuÉn lµ $Q$, m©u thuÉn tr¹ng th¸i lµ $M$ th× cã thÓ biÓu diÔn nguyªn lı trªn nh sau: $$Q \sim M \hspace{1cm} \mbox{vËy} \hspace{0.5cm} Q = K_{(M)}M.$$
Ta h·y gäi ®ã lµ [*ph¬ng tr×nh nh©n qu¶*]{}, trong ®ã $K_{(M)}$ lµ ph¬ng tiÖn ®Ó gi¶i quyÕt m©u thuÉn.
Do c¸ch ®Æt vÊn ®Ò ë møc ®¬n gi¶n nhÊt nªn $K_{(M)}$ chØ cã thÓ lµ mét hµm cña tr¹ng th¸i, tøc lµ hµm cña chİnh m©u thuÉn. Thùc chÊt, nã biÓu hiÖn møc ®é dÔ dµng cña sù tho¸t biÕn khái m©u thuÉn cña tr¹ng th¸i.
NÕu m©u thuÉn ®îc ®Æc trng bëi c¸c ®¹i lîng $..., x, y, z,...$ th× còng chİnh nh÷ng ®¹i lîng nµy sÏ lµ ph¬ng tiÖn chuyÓn t¶i m©u thuÉn, lµ c¸c bËc tù do mµ theo ®ã m©u thuÉn sÏ ®îc gi¶i quyÕt. Khi ®ã, møc ®é dÔ dµng cña sù tho¸t biÕn ph¶i ®îc ®¸nh gi¸ nh lµ ®¹o hµm cña m©u thuÉn theo c¸c bËc tù do cña nã.
Gi¸ trŞ ®¹o hµm cña m©u thuÉn theo mét bËc tù do nµo ®ã cña nã cµng lín th× kh¶ n¨ng “®¸nh h¬i” thÊy lèi tho¸t theo híng Êy cña tr¹ng th¸i cµng lín, “lîng m©u thuÉn” ®îc gi¶i tho¸t theo bËc tù do Êy cµng nhiÒu.
Nh vËy $$K_{(M)} \sim |M'(..., x, y, z, ...)|$$ vµ ta cã $$Q = a |M'(..., x, y, z, ...)| M(..., x, y, z, ...), \tag{I}
\label{NQ-pt}$$ hÖ sè $a$ chØ cã thÓ sinh ra do sù lùa chän hÖ ®¬n vŞ cña c¸c ®¹i lîng.
Chóng ta ®· nãi r»ng, sù kh¸c biÖt lµ nguån gèc cña tÊt c¶. Nhng sù kh¸c biÖt tù nã kh«ng cã nghÜa. C¸i gäi lµ “cã nghÜa” Êy chØ sinh ra trong mèi quan hÖ trùc tiÕp, trong sù so s¸nh trùc tiÕp. Tù nhiªn kh«ng thÓ c¶m biÕt ®îc sù kh¸c biÖt qua “kho¶ng c¸ch”. Chóng ta thõa nhËn cã hai lo¹i kh¸c biÖt: chÊt vµ lîng. §èi víi Tù nhiªn, lo¹i kh¸c biÖt nµo lµ thùc sù tån t¹i?
Mét tr¹ng th¸i nµo ®ã nÕu cã m©u thuÉn néi t¹i th× nã ph¶i biÕn ®æi ®Ó ®¹t tíi tr¹ng th¸i kh«ng cã m©u thuÉn néi t¹i, hay ®óng h¬n, ®¹t tíi tr¹ng th¸i cã m©u thuÉn néi t¹i nhá nhÊt cã thÓ ®îc.
Qu¸ tr×nh ®ã lµ mét chiÒu, tr¶i qua liªn tôc tÊt c¶ c¸c gi¸ trŞ cña m©u thuÉn, tõ gi¸ trŞ ban ®Çu ®Õn gi¸ trŞ cuèi cïng.
Nh vËy, chóng ta ®· cè g¾ng tù thuyÕt phôc r»ng vËn ®éng (biÕn ®æi) nhÊt thiÕt ph¶i cã nguyªn nh©n cña nã vµ tİnh chÊt cña sù biÕn ®æi tu©n theo ph¬ng tr×nh nh©n qu¶. VËy th× sù kh«ng biÕn ®æi, tøc sù b¶o toµn, ph¶i lµ ®iÒu mÆc nhiªn, kh«ng cÇn cã nguyªn nh©n ch¨ng? vµ cã thÓ nãi r»ng: bÊt kú tr¹ng th¸i nµo chØ cã hai kh¶ n¨ng: hoÆc ®îc b¶o toµn, hoÆc bŞ biÕn ®æi; hay ®óng h¬n, tÊt c¶ ®Òu ®îc b¶o toµn ([*kh¼ng ®Şnh*]{} $A$) nhng nÕu sù b¶o toµn ®ã g©y ra m©u thuÉn th× nã ph¶i nhêng chç cho sù biÕn ®æi ®Ó tho¸t khái m©u thuÉn vµ biÕn ®æi ®ã tu©n theo ph¬ng tr×nh nh©n qu¶.
NÕu luËn ®iÓm nµy ®óng th× c«ng viÖc cña chóng ta chØ lµ ë chç: häc c¸ch hiÓu, ®¸nh gi¸ ®óng vµ ®Çy ®ñ m©u thuÉn cña tr¹ng th¸i, m« t¶ nã theo ph¬ng tr×nh nh©n qu¶, khi ®ã ta sÏ cã ®îc quy luËt cña bÊt kú sù biÕn ®æi nµo.
Nhng chØ cã thÕ th× ®· ®ñ cha, cho sù nhËn thøc ®Õn cïng cña chóng ta vÒ Tù nhiªn, vÒ chİnh b¶n th©n con ngêi víi søc m¹nh t duy cña nã, ®Ó cã thÓ gi¶i thİch ®iÒu kú diÖu, m·i m·i lµm ng¹c nhiªn mäi thÕ hÖ: v× sao Tù nhiªn l¹i cã thÓ tù nhËn thøc ®îc chİnh m×nh, th«ng qua s¶n phÈm cña nã: con ngêi?!
\*
\* \*
[**[ø]{}ng dông nguyªn lı nh©n qu¶ vµo mét vµi trêng hîp**]{}\
[**cô thÓ, ®¬n gi¶n nhÊt**]{}
Chóng ta h·y xem xĞt mét vµi thİ dô ®¬n gi¶n nhÊt ®Ó minh ho¹ cho nguyªn lı nh©n qu¶.
[*1. Ph¬ng tr×nh truyÒn nhiÖt*]{}
Gi¶ sö trong mét kho¶ng nµo ®ã cña kh«ng gian mét chiÒu ta cã ph©n bè cña mét ®¹i lîng $L$ nµo ®ã.
NÕu ph©n bè cã sù kh¸c biÖt néi t¹i - tøc lµ cã chøa m©u thuÉn néi t¹i - th× ph©n bè sÏ biÕn ®æi ®Ó ®¹t tíi tr¹ng th¸i cã m©u thuÉn néi t¹i nhá nhÊt. Sù biÕn ®æi ®ã tu©n theo ph¬ng tr×nh nh©n qu¶ (\[NQ-pt\]).
Ta ®a vµo ®¹i lîng $T$, nghŞch ®¶o cña $Q$, gäi lµ ®é tr× trÖ cña sù gi¶i quyÕt m©u thuÉn. Nh vËy $$T = \frac 1{a |M'| M}.$$
Tæng sè ®é tr× trÖ t¹o ra trong qu¸ tr×nh gi¶i quyÕt m©u thuÉn tõ gi¸ trŞ ($M_0$) ®Õn gi¸ trŞ ($M_0 - \Delta M$) ta gäi lµ thêi gian, ®îc sinh ra do qu¸ tr×nh biÕn ®æi ®ã ($\Delta t$).
![\[fig:h1\]](NQ-pt_fig1.eps){width="1\columnwidth"}
Theo ®Şnh nghÜa vµ tõ H×nh \[fig:h1\], ta thÊy r»ng $$\Delta t \approx - \frac{2T + \Delta T}2 \ \Delta M,$$ nh vËy $$\frac{\Delta M}{\Delta t} \approx - \frac 2{2T + \Delta T}.$$ Ta cã $$\lim_{\Delta T \rightarrow 0, \Delta M \rightarrow 0, \Delta t
\rightarrow 0} \frac{\Delta M}{\Delta t} = \frac{dM}{dt} = - \frac
1 T = -a|M'|M.$$ Ta thu ®îc c¸ch diÔn ®¹t míi cña nguyªn lı nh©n qu¶: $$\frac{dM}{dt} = - a |M'(..., x, y, z, ...)| M(..., x, y, z, ...).
\tag{II} \label{NQ-pt2}$$
Nh vËy, nÕu chóng ta quy íc víi nhau ®Ó thêi gian trë thµnh mét ®¹i lîng ®éc lËp, cßn m©u thuÉn l¹i lµ ®¹i lîng phô thuéc nã th× tèc ®é tho¸t biÕn theo thêi gian cña m©u thuÉn tû lÖ víi ®é lín cña m©u thuÉn vµ ph¬ng tiÖn gi¶i tho¸t nã.
Trêng hîp m©u thuÉn ®îc ®Æc trng bëi chİnh nã, tøc lµ $M =
M_{(M)}$, ta cã $$M = M_0 e^{-a(t - t_0)}.$$
Trë l¹i víi ph©n bè cña chóng ta. §Ó thuËn tiÖn, ta tr¶i ph©n bè nµy theo trôc $x$ vµ lÊy mét ®iÓm nµo ®ã lµm gèc to¹ ®é.
![\[fig:h2\]](NQ-pt_fig2.eps){width="1\columnwidth"}
Bëi ph©n bè lµ cña mét ®¹i lîng $L$ nµo ®ã cho nªn mäi gi¸ trŞ cña nã t¹i c¸c ®iÓm cña kh«ng gian ph©n bè sÏ cã cïng thø nguyªn (®ång nhÊt).
Sù kh¸c biÖt néi t¹i cña ph©n bè sÏ lµ sù kh¸c biÖt vÒ lîng. T¹i hai ®iÓm $x_1$ vµ $x_2$ ®¹i lîng $L$ lÊy hai gi¸ trŞ $L_1$ vµ $L_2$ t¬ng øng. V× kh¸c biÖt vÒ lîng nªn chØ cã mét c¸ch ®¸nh gi¸ duy nhÊt: lÊy hiÖu $(L_2 - L_1)$.
Nhng hai ®iÓm $x_1$, $x_2$ chØ “c¶m thÊy sù kh¸c biÖt” cña nhau trong mèi quan hÖ trùc tiÕp, m©u thuÉn xuÊt hiÖn hay kh«ng chØ trong mèi quan hÖ trùc tiÕp ®ã: t¹i ranh giíi cña hai ®iÓm l©n cËn $x_1$, $x_2$ ®¹i lîng $L$ ®ång thêi lÊy hai gi¸ trŞ $L_1$, $L_2$, hai kh¼ng ®Şnh nµy phñ nhËn lÉn nhau vµ ®é lín cña m©u thuÉn phô thuéc vµo hiÖu $(L_2 - L_1)$. Bëi vËy, ®Ó cho hiÖu $(L_2 - L_1)$ lµ s¶n phÈm cña mèi quan hÖ trùc tiÕp gi÷a hai ®iÓm $x_1$, $x_2$ th× ta ph¶i cho, ch¼ng h¹n, ®iÓm $x_2$ tiÕn gÇn v« h¹n tíi ®iÓm $x_1$ (nhng kh«ng trïng víi nã).
Khi ®ã, m©u thuÉn néi t¹i t¹i khu vùc ®iÓm $x_1$ sÏ ®îc ®¸nh gi¸ nh lµ giíi h¹n cña tØ sè $\frac{L_2 - L_1}{x_2 - x_1}$ khi $x_2
\rightarrow x_1$, tøc lµ b»ng gi¸ trŞ ®¹o hµm cña ®¹i lîng $L$ theo kh«ng gian ph©n bè t¹i ®iÓm $x_1$.
Tõ nh÷ng ®iÒu ®· tr×nh bµy, ta cã $M = \frac{dL}{dx} =
\frac{\partial L}{\partial x}$. Thay gi¸ trŞ cña $M$ vµo ph¬ng tr×nh nh©n qu¶ (\[NQ-pt2\]): $$\frac{\partial}{\partial t}\frac{\partial L}{\partial x} = -a
\frac{\partial L}{\partial x}. \label{pt1}$$
M©u thuÉn néi t¹i t¹i mçi ®iÓm ®îc gi¶i quyÕt theo ph¬ng tr×nh (\[pt1\]). §iÒu ®ã lµm ph©n bè bŞ biÕn ®æi. Ta h·y t×m quy luËt cña sù biÕn ®æi nµy.
M©u thuÉn néi t¹i t¹i khu vùc ®iÓm $x$ lµ $M_{x, t} =
\left.\frac{\partial L}{\partial x}\right|_{x, t}$.
Sau kho¶ng thêi gian $\Delta t$ m©u thuÉn nµy gi¶m xuèng ®Õn gi¸ trŞ $M_{x, t+\Delta t} = \left.\frac{\partial L}{\partial
x}\right|_{x, t+\Delta t}$.
![\[fig:h3\]](NQ-pt_fig3.eps){width="1\columnwidth"}
Nh vËy, h×nh nh lµ sù biÕn ®æi nµy ®· dån Ğp mét lîng nµo ®ã gi¸ trŞ cña ®¹i lîng $L$ tõ c¸c ®iÓm cã gi¸ trŞ cao h¬n sang nh÷ng ®iÓm cã gi¸ trŞ thÊp h¬n, t¹o nªn mét “dßng ch¶y” gi¸ trŞ cña ®¹i lîng $L$ qua ®iÓm $x$. Trong vİ dô cña chóng ta “dßng ch¶y” ®ã ch¶y tõ nh÷ng ®iÓm bªn tr¸i h¬n ®iÓm $x$ sang nh÷ng ®iÓm bªn ph¶i h¬n nã.
Râ rµng, ®é lín cña “dßng ch¶y”, tøc lµ lîng c¸c gi¸ trŞ $L$ ch¶y qua ®iÓm $x$ trong kho¶ng thêi gian $\Delta t$ lµ $${\cal J}_x = \left.\frac{\partial}{\partial t}\frac{\partial
L}{\partial x}\right|_x \Delta t = - a \left.\frac{\partial
L}{\partial x}\right|_x \Delta t.$$ T¬ng tù nh vËy, t¹i ®iÓm $x + \Delta x$, ta cã $${\cal J}_{x+\Delta x} = - a \left.\frac{\partial L}{\partial
x}\right|_{x +\Delta x} \Delta t.$$
Trong vİ dô trªn, dßng ${\cal J}_x$ lµm gi¸ trŞ ®¹i lîng $L$ t¹i c¸c ®iÓm trong kho¶ng $\Delta x$ t¨ng lªn, cßn dßng ${\cal J}_{x
+\Delta x}$ th× lµm chóng gi¶m xuèng. KÕt qu¶ lµ sè gia $\Delta L$ mµ kho¶ng $\Delta x$ nhËn ®îc lµ $$\begin{aligned}
\left.\Delta L\right|_{\Delta t} &=& a \Delta t \left(
\left.\frac{\partial L}{\partial x}\right|_{x + \Delta x} -
\left.\frac{\partial L}{\partial x}\right|_x \right) \\ &=& a
\Delta t \left.\frac{\partial^2 L}{\partial x^2}\right|_{x\leq \xi
\leq x +\Delta x} \Delta x.\end{aligned}$$ MËt ®é trung b×nh gi¸ trŞ $\overline{\Delta L}$ t¹i mçi ®iÓm trong kho¶ng $\Delta x$ sÏ lµ $$\left.\overline{\Delta L}\right|_{\Delta t} \cong \frac{a \Delta t
\left.\frac{\partial^2 L}{\partial x^2}\right|_\xi \Delta
x}{\Delta x}.$$ Gi¸ trŞ chİnh x¸c ®¹t ®îc ë giíi h¹n $$\left. \Delta L \right|_{x, \Delta t} = \lim_{\Delta x \rightarrow
0} \left.\overline{\Delta L}\right|_{\Delta t} = a \Delta t
\left.\frac{\partial^2 L}{\partial x^2}\right|_x.$$ Nh vËy, $$\lim_{\Delta t \rightarrow 0} \left.\frac{\Delta L}{\Delta
t}\right|_x = a \frac{\partial^2 L}{\partial x^2},$$ hay $$\frac{\partial L}{\partial t} = a \frac{\partial^2 L}{\partial
x^2}. \label{pt2}$$
Tèc ®é biÕn ®æi theo thêi gian cña ®¹i lîng $L$ t¹i l©n cËn cña bÊt kú ®iÓm nµo cña ph©n bè tØ lÖ víi ®¹o hµm bËc hai theo kh«ng gian ph©n bè cña ®¹i lîng nµy t¹i chİnh ®iÓm Êy.
Nhng nh ta ®· biÕt, ph¬ng tr×nh (\[pt2\]) chİnh lµ ph¬ng tr×nh truyÒn nhiÖt mµ vËt lı häc ®· t×m ra.
MÆt kh¸c, hÖ qu¶ cña c¸ch lËp luËn trªn cho ta sù b¶o toµn gi¸ trŞ cña ®¹i lîng $L$ trong toµn bé ph©n bè, dï r»ng gi¸ trŞ cña ®¹i lîng nµy t¹i mçi ®iÓm riªng biÖt cã thÓ biÕn ®æi, hÔ gi¸ trŞ t¹i ®iÓm nµy gi¶m ®i mét lîng nµo ®ã th× gi¸ trŞ cña ®iÓm l©n cËn t¨ng lªn ®óng mét lîng nh vËy. NÕu kh«ng gian ph©n bè lµ v« h¹n, th× cïng víi sù t¨ng lªn cña thêi gian, gi¸ trŞ trung b×nh cña ph©n bè dÇn gi¶m tíi kh«ng.
[*2. Con quay håi chuyÓn*]{}
XĞt mét vİ dô kh¸c. Chóng ta cã hai con quay víi m«men ®éng lîng t¬ng øng lµ $k_1 \vec{\omega}_1$ vµ $k_2 \vec{\omega}_2$. Sù tån t¹i cña hai con quay víi sù b¶o toµn cña c¸c m«men ®éng lîng cña chóng, xĞt theo gãc ®é vÜ m«, lµ c¸c kh¼ng ®Şnh $K_1$ vµ $K_2$.
Sù b¶o toµn cña c¸c vector m«men ®éng lîng cã thÓ coi lµ sù b¶o toµn cña hai yÕu tè: b¶o toµn ph¬ng vµ b¶o toµn ®é lín. NÕu chóng ®îc g¾n víi nhau theo H×nh \[fig:h4\].a th× sù b¶o toµn ph¬ng cña chóng kh«ng bŞ x©m ph¹m, nhng sù b¶o toµn ®é lín cña mét vector sÏ bŞ sù b¶o toµn cña vector kia x©m ph¹m. M©u thuÉn sÏ cµng gay g¾t nÕu sù kh¸c biÖt gi÷a hai ®é lín cña c¸c vector nµy cµng lín. KÕt qu¶ lµ hÖ thèng nµy ph¶i biÕn ®æi thÕ nµo ®ã ®Ó c¶ hÖ sÏ cã mét vector m«men ®éng lîng duy nhÊt $k_3
\vec{\omega}_3$. Trong trêng hîp hai con quay ®îc g¾n víi nhau nh trong c¸c H×nh \[fig:h4\].b m©u thuÉn tr¹ng th¸i sÏ phøc t¹p h¬n. Kh«ng chØ sù b¶o toµn vÒ ®é lín mµ c¶ sù b¶o toµn vÒ ph¬ng ®Òu bŞ x©m ph¹m. C¸ch thøc gi¶i quyÕt m©u thuÉn cña tr¹ng th¸i phô thuéc vµo kÕt cÊu cña khíp nèi.
![\[fig:h4\]](NQ-pt_fig4.eps){width="1\columnwidth"}
Chóng ta xĞt trêng hîp thø ba, ë ®ã c¸c yÕu tè träng lùc vµ ly t©m (cã trong trêng hîp thø hai) cã thÓ coi nh kh«ng ®¸ng kÓ (H×nh \[fig:h5\]).
![\[fig:h5\]](NQ-pt_fig5.eps){width="1\columnwidth"}
§Ó ®¬n gi¶n, chóng ta cho r»ng ®éng c¬ duy tr× vËn tèc gãc $\omega$ cña hÖ kh«ng ®æi. Nh vËy chóng ta sÏ chØ quan t©m ®Õn m©u thuÉn sinh ra do sù b¶o toµn ph¬ng cña $k \vec{\omega}_0$ bŞ x©m ph¹m.
Kh¼ng ®Şnh $K_1 \equiv$ sù b¶o toµn cña $k \vec{\omega}_0$, nãi r»ng: tèc ®é biÕn thiªn cña ph¬ng vector $k \vec{\omega}_0$ b»ng kh«ng. Nhng kh¼ng ®Şnh $K_2 \equiv$ sù b¶o toµn cña $\omega$, nãi r»ng: kh«ng, ph¬ng cña $k \vec{\omega}_0$ ph¶i biÕn ®æi víi vËn tèc gãc $\omega \cos\alpha$.
Nh vËy, ë møc ®é vÜ m«, hiÖu $[K_1 - K_2] = \omega \cos\alpha$ lµ nguån gèc cña m©u thuÉn vµ m©u thuÉn ®ã tØ lÖ víi hiÖu nµy, $$M \sim \omega\cos\alpha; \hspace{1cm} M = k\omega_0
\omega\cos\alpha,$$ hÖ sè tØ lÖ $k\omega_0$ ®îc ®a vµo (vÉn trªn quan ®iÓm vÜ m«) dùa trªn lı lÏ: nÕu $\omega_0$ b»ng kh«ng th× ph¬ng cña vector $k
\vec{\omega}_0$ kh«ng tån t¹i mét c¸ch x¸c ®Şnh vµ do ®ã vÊn ®Ò m©u thuÉn ph¸t sinh do sù b¶o toµn ph¬ng cña nã kh«ng ®îc ®Æt ra.
Thay gi¸ trŞ cña $M$ vµo ph¬ng tr×nh nh©n qu¶ (\[NQ-pt2\]), ta ®îc $$\frac{\partial M}{\partial t} = -a k^2 \omega_0^2 \omega^2
\sin\alpha \cos\alpha.$$ Tõ ph¬ng tr×nh ta thÊy r»ng, nÕu $\alpha =0$ th× tèc ®é tho¸t biÕn cña m©u thuÉn tr¹ng th¸i b»ng kh«ng.
LÊy ®¹o hµm cña m©u thuÉn theo thêi gian, $$\frac{\partial \alpha}{\partial t} = a k \omega_0 \omega
\cos\alpha, \hspace{1cm} (\alpha \neq 0). \label{pt3}$$
Sù biÕn ®æi cña $\alpha$ g©y ra m©u thuÉn míi, m©u thuÉn nµy tû lÖ víi gi¸ trŞ cña $\frac{\partial \alpha}{\partial t}$, do ®ã sÏ kh«ng cã sù b¶o toµn vËn ®éng theo yÕu tè $\alpha$, vµ nh vËy tèc ®é tho¸t biÕn trong c«ng thøc (\[pt3\]) chİnh lµ vËn tèc tøc thêi cña trôc mÆt ph¼ng quay theo yÕu tè $\alpha$.
Thêi gian ®Ó gãc gi÷a trôc mÆt ph¼ng quay (tøc ph¬ng cña vector $k \vec{\omega}_0$) vµ ph¬ng n»m ngang biÕn ®æi tõ gi¸ trŞ ($+0$) ®Õn gi¸ trŞ ($\alpha$) sÏ lµ $$t = \frac 1{2ak\omega_0 \omega} \left.\ln \frac{1 + \sin\alpha}{1
- \sin\alpha}\right|_{+0}^\alpha.$$
[*3. VÒ b¶n chÊt cña c¸i gäi lµ “trêng”*]{}
VËt lı häc hiÖn nay cho r»ng thÕ giíi ®îc cÊu t¹o bëi c¸c h¹t c¬ b¶n. VËy th× mét vÊn ®Ò ®îc ®Æt ra lµ: c¸c h¹t c¬ b¶n ®ã ph¶i cã cÊu tróc néi t¹i thÕ nµo ®Ó cã thÓ b¶o toµn ®îc ®èi víi nguyªn lı nh©n qu¶?
Cã thÓ gi¶i quyÕt vÊn ®Ò nµy mét c¸ch thuÇn tuı lı thuyÕt ®îc kh«ng?
Cho r»ng cã mét kh«ng gian h÷u h¹n $[A]$ cã cÊu tróc néi t¹i tho¶ m·n sù bÊt biÕn ®èi víi nguyªn lı nh©n qu¶.
Kh«ng gian nµy n»m trong kh«ng gian tuyÖt ®èi \[O\] cña chóng ta. T¹i n¬i ranh giíi cña hai kh«ng gian xuÊt hiÖn m©u thuÉn do sù kh¸c biÖt gi÷a hai kh«ng gian ®ã g©y ra.
Bëi c¶ hai kh«ng gian ®Òu tù b¶o toµn nªn m©u thuÉn ®ã chØ cã thÓ ®îc gi¶i quyÕt b»ng c¸ch h×nh thµnh mét vïng ®Öm (tøc trêng), nhê ®ã sù kh¸c biÖt trë lªn dŞu h¬n, ®iÒu hoµ h¬n. CÊu tróc cña vïng ®Öm ph¶i thÕ nµo ®ã ®Ó møc ®é ®iÒu hoµ ®¹t tíi gi¸ trŞ lín nhÊt, tøc lµ m©u thuÉn néi t¹i t¹i mçi ®iÓm cña trêng cã gi¸ trŞ nhá nhÊt cã thÓ ®îc.
Mét ®iÒu râ rµng lµ cµng xa t©m cña kh«ng gian $[A]$, tİnh chÊt $[A]$ cµng gi¶m ®i. Nãi c¸ch kh¸c, vïng ®Öm (trêng) bao quanh kh«ng gian $[A]$ còng cã tİnh chÊt $[A]$ nhng tİnh chÊt nµy lµ hµm cña $r$, tøc kho¶ng c¸ch tõ ®iÓm ®ang xĞt cña trêng tíi t©m cña kh«ng gian $[A]$.
Tõ nh÷ng ®iÒu ®· tr×nh bµy vµ nÕu kı hiÖu vïng ®Öm lµ $T$, ta cã $$T_{[A]} = g(r) \frac{[A]}{r},$$ $g(r)$ lµ mét hµm cha biÕt, nã ®Æc trng cho sù hµi hoµ néi t¹i cña trêng.
![\[fig:h6\]](NQ-pt_fig6.eps){width="1\columnwidth"}
NÕu trong vïng trêng $T_{[A]}$ cã mét kh«ng gian $[B]$ vµ kh«ng gian nµy kh«ng lµm nhiÔu lo¹n ®¸ng kÓ trêng $T_{[A]}$, th× khi ®ã sù kh¸c biÖt gi÷a $[B]$ vµ $T_{[A]}$ sÏ buéc $[B]$ ph¶i vËn ®éng trong trêng ®Ó ®¹t tíi vŞ trİ mµ ë ®ã kh¸c biÖt gi÷a $[B]$ vµ $T_{[A]}$ cã trŞ sè nhá nhÊt (ë ®©y chóng ta ®· cho r»ng kh«ng gian $[B]$ còng cã kh¶ n¨ng tù b¶o toµn). M©u thuÉn tr¹ng th¸i ®ã tû lÖ víi hiÖu $\left[ [B] - T_{[A]}\right]$.
NÕu t×m ®îc hÖ sè $c$ dïng ®Ó “dŞch ng«n ng÷” cña tİnh chÊt $[B]$ sang “ng«n ng÷” cña tİnh chÊt $[A]$ th× m©u thuÉn cã thÓ ®îc diÔn ®¹t nh sau: $$M = f\left( c[B] - g(r) \frac{[A]}{r}\right), \hspace{1cm} f -
\mbox{hÖ sè tû lÖ.}$$
Quy luËt vËn ®éng cña kh«ng gian $[B]$ trong trêng $T_{[A]}$ sÏ ®îc t×m ra qua ph¬ng tr×nh nh©n qu¶ (\[NQ-pt2\]): $$\begin{aligned}
\frac{\partial M}{\partial t} &=& -a |M'|M \\ &=& -a[A]f^2 \left|
\frac{g(r) - rg'(r)}{r^2}\right| \left( c[B] - [A]
\frac{g(r)}r\right).\end{aligned}$$ [ë]{} ®©y, ®¹i lîng chuyÓn t¶i (bËc tù do) cña m©u thuÉn lµ $r$.
Bëi sù vËn ®éng cña kh«ng gian $[B]$ ph¶i x¶y ra ®ång thêi theo tÊt c¶ c¸c ph¬ng cã thµnh phÇn híng t©m, do ®ã vËn tèc tho¸t biÕn tæng hîp cña tr¹ng th¸i - tøc lµ vËn tèc tæng hîp cña kh«ng gian $[B]$ trong trêng $T_{[A]}$ sÏ ®îc ®¸nh gi¸ nh lµ tİch ph©n cña tèc ®é tho¸t biÕn theo mäi ph¬ng cã thµnh phÇn híng t©m. $$\begin{aligned}
\frac{\partial M}{\partial t} \!\! &=&\!\! -a[A] 4\pi f^2\!\!\!\!
\int_0^{\frac{\pi}{2}}\!\! \left| \frac{g(r) \!-\! r
g'\!(r)}{r^2}\right| \!\! \left( \! c[B] \!-\! [A] \frac{g(r)}r\!
\right)\! \cos^2\!\!\varphi d\varphi \\ &=& -a\pi^2 f^2 [A] \left|
\frac{g(r) \!-\! r g'\!(r)}{r^2}\right| \!\!\left( \! c[B] \!-\!
[A] \frac{g(r)}r\! \right).\end{aligned}$$ Khai triÓn vÕ tr¸i ta ®îc $$\begin{aligned}
f [A] \frac{g(r)}{r^2} \frac{\partial r}{\partial t} &=& -a\pi^2
f^2 [A] \left| \frac{g(r) \!-\! r g'\!(r)}{r^2}\right| \!\!\left(
\! c[B] \!-\! [A] \frac{g(r)}r \! \right) \\ \frac{\partial
r}{\partial t} &=& -a f \pi^2 \frac{\left| g(r) - r g'(r)
\right|}{g(r)} \left( c[B] - [A] \frac{g(r)}r\right).\end{aligned}$$ NÕu chøng tá ®îc r»ng sù biÕn ®æi cña $r$ còng nh sù b¶o toµn cña $\frac{\partial r}{\partial t}$ g©y ra m©u thuÉn míi tû lÖ víi chİnh $\frac{\partial r}{\partial t}$, th× tèc ®é tho¸t biÕn võa thu ®îc còng chİnh lµ vËn tèc tøc thêi cña $[B]$ trong trêng $T_{[A]}$.
------------------------------------------------------------------------
------------------------------------------------------------------------
------------------------------------------------------------------------
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
author:
- |
Yan Zhang\
University of Southampton\
Southampton, UK\
`[email protected]`\
Jonathon Hare\
University of Southampton\
Southampton, UK\
`[email protected]`\
Adam Prügel-Bennett\
University of Southampton\
Southampton, UK\
`[email protected]`\
title: 'FSPool: Learning Set Representations with Featurewise Sort Pooling'
---
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
author:
- Pranjal Awasthi
- 'Maria-Florina Balcan'
- Colin White
bibliography:
- 'distributed\_clustering.bib'
title: 'General and Robust Communication-Efficient Algorithms for Distributed Clustering [^1] '
---
Introduction {#sec:intro}
============
Related Work {#sec:related_work}
------------
Preliminaries {#sec:prelim}
=============
General Robust Distributed Clustering {#sec:distr}
=====================================
Communication Complexity Lower Bounds {#sec:lowerbounds}
=====================================
Distributed Clustering under Approximation Stability {#sec:as}
====================================================
Distributed Clustering under Spectral Stability {#sec:spectral}
===============================================
Conclusion {#sec:conclusion}
==========
Proof from Section \[sec:lowerbounds\] {#app:lowerbounds}
======================================
Proofs from Section \[sec:as\] {#app:as}
==============================
[^1]: Authors’ addresses: `[email protected], [email protected], [email protected]`. This work was supported in part by NSF grants CCF-1422910, CCF-1535967, IIS-1618714, a Sloan Research Fellowship, a Microsoft Research Faculty Fellowship, a Google Research Award, and a National Defense Science & Engineering Graduate (NDSEG) fellowship.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
author:
- '[^1]'
bibliography:
- '../master.bib'
title: 'Surface effects in solar-like oscillators'
---
Introduction {#s:intro}
============
The era of space-based asteroseismology, driven chiefly by COROT [@corot] and *Kepler* [@kepler], has provided observations of hundreds of cool main-sequence stars in which dozens of individual mode frequencies can be measured. To exploit this data, however, we need to correct for a systematic difference between observed and modelled mode frequencies caused by improper modelling of the near-surface layers of these stars: the so-called *surface term* or *surface effect*. Motivated by a newfound need to correct for the surface effect, significant progress has been achieved in the last few years and can be expected in the near future.
The purpose of this review is to first briefly recount our physical understanding of the surface effect (Sec. \[s:problem\]) and then review recent progress along two lines. First, several authors have proposed parametrizations of the surface effect (as a function of frequency) to suppress its influence when fitting stellar models to observed mode frequencies (Sec. \[s:param\]). Second, a few research groups have begun replacing the near-surface layers of stellar models with average structures taken from detailed three-dimensional radiation hydrodynamic simulations (3D RHD, Sec. \[s:3drhd\]). Finally, I close with a few thoughts on how we might progress further on the problem of surface effects in the near future (Sec. \[s:future\]).
I do not pretend that this review is exhaustive. Judging by the amount of material I excluded from my talk, it would be impossible to cover all the literature on the subject in 30 minutes. I apologize to anyone who feels their contribution has been omitted and seek to assure them that the cause is only brevity, not malice!
![The basic problem of surface effects in the Sun. The white points show differences between low-degree ($\ell\leq3$) mode frequencies observed by the Birmingham Solar Oscillation Network [BiSON; @broomhall2009; @davies2014a] and linear adiabatic mode frequencies predicted for a standard solar model [Model S; @modelS]. The blue annotations show the two main features that lead us to believe that the systematic difference is a surface phenomenon: the differences grow with frequency and are largely independent of the angular degree $\ell$ (see text). The shaded region shows where the modes have their greatest power and are most easily observed, demonstrating that for most asteroseismic targets, *all* the observed frequencies are probably affected by the surface effect.[]{data-label="f:problem"}](problem.pdf){width="\hsize"}
The problem {#s:problem}
===========
Phenomenology of the surface effects
------------------------------------
Suppose that one calibrates a solar model in the traditional sense by varying the mixing length parameter, initial helium abundance and initial metallicity to evolve a stellar model that matches the Sun’s current radius, luminosity and surface metallicity, with the mass fixed at $1{\hbox{$\rm\thinspace \text{M}_{\odot}$}}$ and the age at the meteoritic age of the solar system. We take Model S [@modelS] as an example. If we compute the linear adiabatic mode frequencies of this model and compare them with observed values [e.g. the low-degree data from the Birmingham Solar Oscillation Network, BiSON; @broomhall2009; @davies2014a] we might hope that the differences between observed and modelled frequencies are randomly scattered about zero.
Instead, one gets the values plotted in Fig. \[f:problem\]. The white points indicate the differences between the mode frequencies predicted for Model S and those observed by BiSON. The discrepancy is much larger than the quoted uncertainties but it is also not random, and its structure tells us something about where the problem arises. First, the frequency differences do not depend on the angular degree $\ell$, which suggests that the discrepancy lies well above the modes’ lower turning points. The low-degree data alone only tells us that the cause is not very deep in the Sun but the frequency differences of the higher-degree modes are also $\ell$-independent. Because they have shallower lower turning points, this suggests that the problem is quite close to the Sun’s surface. Second, the frequency differences are close to zero at frequencies below about $2200{\hbox{\rm\thinspace $\mu$Hz}}$. Higher-frequency modes have shallower upper turning points, which again implies that the problem is somewhere near the Sun’s surface. At $2200{\hbox{\rm\thinspace $\mu$Hz}}$, the modes’ upper turning points are around $1{\rm\thinspace Mm}$ below the surface, which implies that the effect really is confined to the near-surface layers.
The shaded region in Fig. \[f:problem\] indicates the range of frequency covering about five radial orders either side of the frequency of maximum oscillation power $\nu{_\text{max}}$. This is the range in which modes oscillate with the greatest power and are thus most easily observed. It shows that the surface effect probably affects nearly all the observed modes in distant Sun-like stars, unlike the Sun, in which we observe low-frequency modes that appear unaffected by the surface effect. The best targets from the nominal *Kepler* mission have lowest mode frequencies equivalent to about $2300{\hbox{\rm\thinspace $\mu$Hz}}$ in the Sun, which is within the range of affected modes. For this reason, the surface effect is unavoidable: when fitting stellar models to individual mode frequencies, something *must* be done about the surface effect. In the case of the Sun, even the large separations of the modelled and observed frequencies differ by about $1{\hbox{\rm\thinspace $\mu$Hz}}$. When applied to the standard scaling relations [@kjeldsen1995], this bias in the large separation corresponds to biases in mass and radius of about $3$ and $1.5$ per cent, respectively.[^2]
------------------------------------------------- --------------------------------------------------
{width="0.5\hsize"} {width="0.5\hsize"}
------------------------------------------------- --------------------------------------------------
The physical cause
------------------
It may come as a surprise that we have a fairly good idea about what causes the surface effect: improper modelling of near-surface convection. Most stellar models use some form of mixing-length theory (MLT), in which the convection zone is presumed to contain buoyantly-unstable rising and falling parcels of material (see Fig. \[f:diagram\], left). These parcels rise or fall by one mixing length, typically parametrized in terms of the local pressure scale height, $H_P=-\mathrm{d}r/\mathrm{d}\ln P$, after which they disperse, mixing the heat and composition of their origin into their new surroundings.
In reality, the flows are much more complicated, as is now understood from detailed 3D RHD simulations that accurately reproduce many observable features of convection at the Sun’s surface [see @nordlund2009 for an excellent review of the Sun’s surface convection]. Let us start with one of the slow upflows. As it rises and the density decreases, so the flow expands horizontally and, to conserve mass, part of it must turn over and join whatever downflows exist (see Fig. \[f:diagram\], right). The rising plume ultimately appears as a granule at the surface, where the flows are chiefly horizontal. They then radiate heat to the vacuum of space before plummeting downward in narrow, turbulent intragranular lanes. Along the way back down, these downflows will draw material turning over from the widening upflows.
This is a very different picture from the calm rise and fall of MLT’s parcels and it leads to a number of effects that affect the mode frequencies. Following the thorough discussion by Rosenthal [@rosenthal1997], we can broadly divide these into two types of effect. *Model* physics includes everything that is wrong with the background model that we perturb. This includes, but is not limited to, MLT’s incorrect temperature gradient, the incorrect atmospheric structure and the absence of turbulent pressure. *Modal* physics includes everything that is wrong with the calculation of the mode frequencies, which are affected by the perturbation to the turbulent pressure [e.g. @rosenthal1999], the modification of wave speeds when travelling with or against the flows [e.g. @brown1984] and various effects of non-adiabaticity [e.g. @houdek1996]. All of these effects are most pronounced near the surface where convection becomes inefficient and the temperature gradient deviates furthest from the adiabatic value.
That so many physical effects contribute to the surface effect makes it a difficult problem to tackle piece by piece. In working on one component, one might think the problem is solved, only to find that another component returns you to square one. But it is not hopeless! We can learn how much each component might contribute and gradually add them up, bearing in mind that as our models improve, we might sometimes veer further from the observations before once again closing the gap.
![Frequency differences, as a function of frequency, for the observed surface effect in the Sun (white points) and several parametrizations (see Sec. \[s:param\]). The shaded region shows the frequency range over which the modes have their greatest power, as in Fig. \[f:problem\]. The parametrizations shown are a power-law [solid blue, @kbcd2008], the cubic and combined formulae [dashed orange and dash-dotted red, @ball2014] and a modified Lorentzian [dotted green, @sonoi2015].[]{data-label="f:param"}](modelS_corr.pdf){width="\hsize"}
Parametrizations {#s:param}
================
The surface effect in Fig. \[f:problem\] appears to be a relatively simple function of mode frequency only. Thus, several groups have proposed parametric forms for this function whose parameters can be fit when comparing stellar models to observations. Here I shall review the best known and compare them for the Sun.
First, Kjeldsen et al.[@kbcd2008] proposed that the surface effect can be described as a power law with an index fixed to a solar-calibrated value (usually around $5$ but slightly dependent on the precise physics of the stellar model). They also proposed that the magnitude of the power law be fit after rescaling the frequencies so that the stellar model being compared has the same mean density as the observed star. To rescale the frequencies so, they propose using the ratio of the large separations. This simple parametrization has been widely used since its publication [e.g. @silva2013].
More recently, Ball & Gizon[@ball2014] proposed parametrizations based on surface perturbations and the asymptotic behaviour of the eigenmodes. Roughly speaking, the displacement eigenfunctions are exponentially decaying functions near the photosphere and, combining them with the variational principle for the linear, adiabatic oscillation equations [@ledoux1958], one finds that, for a sound speed perturbation or pressure scale height perturbation near the surface, the frequency shifts go either like $\nu^3/\mathcal{I}$ or $\nu^{-1}/\mathcal{I}$, where $\nu$ is the mode frequency and $\mathcal{I}$ the normalized mode inertia. These parametric forms, which Ball & Gizon[@ball2014] refer to as the *cubic* and *inverse* terms, respectively, were originally derived by Gough[@gough1990][^3] in a discussion of the Sun’s frequency shifts over the magnetic activity cycle. The inverse term alone does not fit the data well so Ball & Gizon[@ball2014] proposed to combine it with the cubic term, giving what they call the *combined* surface correction.
Most recently, Sonoi et al.[@sonoi2015] proposed to describe the surface effect as a modified Lorentzian function and calibrated its parameters to frequency shifts induced by replacing the near-surface layers of stellar models with averaged data from hydrodynamics simulations (see Sec. \[s:3drhd\], below). This parametrization is very new and has not yet been tested on observed data.
Fig. \[f:param\] shows the same data as Fig. \[f:problem\] (BiSON against Model S), along with the above-mentioned parametrizations. The power law fit performs reasonably well in the shaded range around $\nu{_\text{max}}$ but overestimates the surface effect both where it begins to rise and at higher frequencies. The cubic and the combined terms fare better. Though the improvement by using the combined term (rather than just the cubic term) is significant for the Sun, this was not the case for the COROT target HD 52265 studied by Ball & Gizon[@ball2014]. Finally, the modified Lorentzian captures most of the low-frequency behaviour but underestimates the difference at high frequencies.
Though different in principle, it is worth mentioning several methods proposed by Roxburgh (and Vorontsov in earlier work) [@roxburgh2003; @roxburgh2015; @roxburgh2016]. These are all based on representing the oscillation modes as simple oscillations with phase shifts at the inner and outer boundaries. The outer phase shift contains the undesired and presumably $\ell$-independent surface term whereas the inner phase shift is related to the structure of the stellar core. One can combine the frequencies into ratios of differences or so-called *separation ratios* that are nearly independent of the near-surface layers [@roxburgh2003]. Ot[í]{} Floranes et al.[@oti2005] computed kernels for these quantities and demonstrated that they are, indeed, largely insensitive to the near-surface layers and they have seen widespread use in asteroseismic modelling. From the same underlying principles, Roxburgh[@roxburgh2015; @roxburgh2016] described methods to fit out a more general $\ell$-independent component of the frequency differences. These are too new to have been used widely.
These various parametrized methods have not yet been systematically compared with observations, though the community’s collective experience suggests that none generally leads to absurd results. Schmitt & Basu[@schmitt2015] conducted the most thorough study yet by inserting structural perturbations into stellar models across the HR diagram and then trying to fit the frequency differences using the solar-calibrated power law, the cubic and combined terms of Ball & Gizon[@ball2014] or the observed solar surface effect, rescaled by the large separation. The combined term by Ball & Gizon[@ball2014] appeared to fare best, although the scaled solar term also performed reasonably on the main sequence.
The parametrizations do not solve the problem of the surface effects but they at least allow us to exploit the reams of data already available while we work towards properly modelling the surface effects. The results should always be interpreted with the knowledge that the Sun remains the only star for which we can truly calibrate the frequency differences. Everything else depends on the confidence we place in how well our best-fitting models represent the stars under study.
![Frequency differences for the Sun between different combinations of models and data as a function of model frequencies. The lines show the frequency differences between the solar model before and after the near-surface layers are replaced by horizontally-averaged 3D RHD simulation data. The different lines correspond to different choices of averaging co-ordinate: geometric depth $z$ (solid blue), pressure $P$ (dotted green) or optical depth $\tau$ (dashed red). The white points show the same differences between observed and modelled frequencies as in Fig. \[f:problem\]. The orange squares show the frequency differences after the near-surface layers of the solar model have been replaced by 3D RHD simulation data averaged at constant geometric depth $z$. The overall extent of the surface effect is reduced to a few $\mu\mathrm{Hz}$ but a clear systematic difference remains.[]{data-label="f:3drhd"}](modelS.pdf){width="\hsize"}
![Kiel diagram (surface gravity against effective temperature) showing the parameters of the 3D RHD simulations used by [@sonoi2015 blue squares] and [@ball2016 orange circles]. The grey lines show solar-metallicity evolutionary tracks from BaSTI [@basti2004] for masses from $0.6$ to $2.0{\hbox{$\rm\thinspace \text{M}_{\odot}$}}$ in steps of $0.2{\hbox{$\rm\thinspace \text{M}_{\odot}$}}$. Both sets of simulations include a solar model (models A and G2 in Sonoi et al.[@sonoi2015] and Ball et al.[@ball2016]) and the hottest models (models B and F3 in Sonoi et al.[@sonoi2015] and Ball et al.[@ball2016]) also have comparable parameters.[]{data-label="f:HR"}](sonoi_ball_HR.pdf){width="\hsize"}
------------------------------------------ -----------------------------------------------------
{width="0.5\hsize"} {width="0.5\hsize"}
------------------------------------------ -----------------------------------------------------
Three-dimensional radiation hydrodynamics {#s:3drhd}
=========================================
I mentioned in Sec. \[s:problem\] that the surface effect is chiefly caused by improper modelling of near-surface convection. So why not use better models of near-surface convection? This is the idea behind recent efforts to combine stellar models with 3D RHD simulations. Several groups have simulated near-surface convection from first principles in stars of various spectral types [e.g. @ludwig2009; @beeck2013a; @magic2013]. These simulations are sufficiently realistic to reproduce most of the observed characteristics of the Sun’s near-surface convection [again, see @nordlund2009 for a review] and it is assumed that they are similarly realistic for other stars.
The process of replacing a stellar model’s near-surface layers with averaged simulation data is becoming known as *patching*. The frequency differences are then computed between the *unpatched* model (the original stellar model) and the *patched* model (with the near-surface layers replaced).
The idea of patching is not new. Rosenthal et al.[@rosenthal1999] restricted their study of solar oscillations to modes with angular degree $\ell>60$. These modes are trapped within the solar convection zone, so they could compare their averaged simulation data with envelope models computed using MLT. Their early results showed that replacing the equilibrium structure of the stellar model with the simulation data, averaged at constant geometric depth, already introduced a surface effect of similar magnitude to the observed effect, although a significant systematic effect remained. More recently, Piau et al.[@piau2014] used 3D RHD simulation data to compute surface effects in a complete solar model (not just the convective envelope), finding that the structural component of the surface effect reduced the remain discrepancy to a few ${\hbox{\rm\thinspace $\mu$Hz}}$. Finally, on the subject of the solar surface correction, Magic & Weiss[@magic2016] computed surface effects using simulations with different input magnetic field strengths and found that they could reproduce reasonably well the frequency shifts induced by the changing level of magnetic activity in the Sun.
The 12 months preceding this meeting saw the first papers to combine stellar models and 3D RHD simulations for the surface effects in other types of star. First, Sonoi et al.[@sonoi2015] combined stellar models from CESTAM [@morel1997; @morel2008] with simulations from the CIFIST atmosphere grid [@ludwig2009]. Second, Ball et al.[@ball2016] combined stellar models from MESA [@paxton2011; @paxton2013; @paxton2015] with simulations from the MURaM code [@beeck2013a]. The two groups independently performed nearly the same calculations using somewhat complementary sets of stellar models. The ten simulations used by Sonoi et al.[@sonoi2015] cover one red giant (around the red clump) and dwarfs and subgiants hotter than the Sun. The four simulations used by Ball et al.[@ball2016] span the main-sequence from spectral type F3 to K5. Fig. \[f:HR\] shows the atmospheric parameters for the two groups’ simulations.
These studies are not definitive. For a start, they only deal with the part of the surface effect caused by improving the structure of the equilibrium stellar model. The averaged simulation profiles include the turbulent pressure but it remains unclear what is the appropriate form of the perturbation to the turbulent pressure. Both Sonoi et al.[@sonoi2015] and Ball et al.[@ball2016] assume that the turbulent pressure varies with the total pressure: Rosenthal et al.[@rosenthal1999] dubbed this the *gas gamma one* approximation. Changing this assumption potentially affects the results by a factor of about two [@rosenthal1999]. Moreover, it is unclear exactly what is the appropriate horizontal average to take from the simulation data. Sonoi et al.[@sonoi2015] and Ball et al.[@ball2016] both used averages over constant geometric depth but different averages give surface effects that differ by a few ${\hbox{\rm\thinspace $\mu$Hz}}$ for the Sun (see Fig. \[f:3drhd\]).
Fig. \[f:3drhd\] shows how the mode frequencies of a standard solar model [Model S again, @modelS] are changed by the modification of the near-surface equilibrium structure. The white points are the same differences shown in Figs \[f:problem\] and \[f:param\]. The solid blue, dotted green and dashed red curves are the differences in the model frequencies before and after patching with the G2-type MURaM simulation averaged over constant geometric depth, pressure or optical depth. The spread in the curves shows that uncertainty above the appropriate average introduces an uncertainty in the frequency shifts of about $0.5$–$1.0{\hbox{\rm\thinspace $\mu$Hz}}$. Finally, the orange squares show the remaining difference between the patched model (with the simulation averaged at constant geometric depth) and the BiSON observations. The overall surface effect is reduced substantially, though clearly a large effect remains and the remaining difference is still a surface effect. It is not yet clear if the remaining trend is because the averaged near-surface structure is still not quite right, because non-adiabatic effects have been neglected, or (most likely) both.
With these uncertainties in mind, both teams found that the surface effect is larger in stars that are hotter. Based on their cooler dwarfs, Ball et al.[@ball2016] also noted that the overall shape of the frequency differences as a function of frequency is similar in the G2-, K0- and K5-type models, but some qualitative change sets in between the F3- and G2-type models. Sonoi et al.[@sonoi2015], with their greater coverage of surface gravity, also found that the surface effect increases with increasing surface gravity. Within their limitations, the two groups’ results are mutually consistent. They have two simulations with similar parameters and the results agree well. Fig. \[f:spectype\] shows the frequency shifts for all the simulations by Ball et al.[@ball2016 left] and the simulations A and B of Sonoi et al.[@sonoi2015 right], which have similar parameters to models G2 and F3 of Ball et al.[@ball2016].
Both teams also compared the parametric fits described in Sec. \[s:param\], though Sonoi et al.[@sonoi2015] only compared their modified Lorentzian with a power law. Fig. \[f:spectype\] shows that a simple power law does not describe the differences between the patched and unpatched models very well. Ball et al.[@ball2016] found that the combined term was consistently superior, notably in their F3 model, but a scaled solar frequency correction fits reasonably well in their three cooler models. Sonoi et al.[@sonoi2015] provided simple fits to the best-fit parameters as a function of surface properties, though they did not consider a scaled solar correction or either of the corrections by Ball & Gizon [@ball2014].
Further exploitation of the 3D RHD simulations is underway, notably on non-adiabatic effects, but these early results already give some indication of how much of a surface effect is introduced by improving the background stellar model. It remains to be seen if the conclusions hold up as further surface effects are considered.
The future {#s:future}
==========
To close, I briefly opine on how we might progress further on the problem of surface effects. The main theoretical path at this point is to further exploit the 3D RHD simulations. There is far more information available than simply the horizontally- and temporally-averaged profiles and this information can be used to investigate other components of the surface effect. But it should be remembered that even indirect conclusions drawn from the simulations can be useful. For example, the parametrizations of the surface effects tend to correlate with the mixing-length parameter in stellar models. There is good physical reason for this: both the surface effect and the mixing-length parameter are sensitive to the superadiabatic layer near the stellar surface. If the mixing-length parameter is constrained separately by the simulations [e.g. @ludwig1999; @trampedach2014b] then the surface effect is also better constrained.
Progress is more difficult from the observational side. The best solar-like oscillators from the nominal *Kepler* mission show modes oscillating at frequencies nearly low enough that they are unaffected by the surface effect. If just a few more radial orders could be detected, these low frequencies could potentially be used to fit models without a surface term, though at the cost of discarding the many higher-frequency modes that are available. Alas, no imminent mission will provide such high-quality data for single targets, so we may have to wait until PLATO [@plato] for higher-quality data on single targets.
From the ground, however, there is tremendous potential from the Stellar Oscillation Network Group [see e.g. @jcd_kasc9 these proceedings]. Because it observes in radial velocity, the background signal of granulation is weaker, which allows lower-frequency modes to be detected more easily. This could allow us to calibrate models directly to the unaffected frequencies and inspect the remaining frequencies to determine the surface effect after fitting the stellar model. One node of the network is fully operational and another partially so. The first results from the first node were reported at this meeting [@jcd_kasc9]. Adding nodes to the network probably represents our best chance of bringing tight observational constraints to bear on the problem of surface effects.
Acknowledgement {#acknowledgement .unnumbered}
===============
The author would like to thank the organizers for partial financial support to attend the meeting. He also acknowledges research funding by Deutsche Forschungsgemeinschaft (DFG) under grant SFB 963/1 “Astrophysical flow instabilities and turbulence”, Projects A18.
[^1]:
[^2]: The scaling relations are in essence empirical, which suppresses this effect. But if the surface effect varies significantly between different stars, it could be important.
[^3]: The cubic term is also mentioned by Libbrecht & Woodard[@libbrecht1990] and Goldreich et al.[@goldreich1991].
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: |
We report spatially resolved measurements of the oblique merging of two supersonic laboratory plasma jets. The jets are formed and launched by pulsed-power-driven railguns using injected argon, and have electron density $\sim
10^{14}$ cm$^{-3}$, electron temperature $\approx 1.4$ eV, ionization fraction near unity, and velocity $\approx
40$ km/s just prior to merging. The jet merging produces a few-cm-thick stagnation layer, as observed in both fast-framing camera images and multi-chord interferometer data, consistent with collisional shock formation \[E. C. Merritt et al., Phys. Rev. Lett. [**111**]{}, 085003 (2013)\].
author:
- 'Elizabeth C. Merritt'
- 'Auna L. Moser'
- 'Scott C. Hsu'
- 'Colin S. Adams'
- 'John P. Dunn'
- 'A. Miguel Holgado'
- 'Mark A. Gilmore'
title: Experimental evidence for collisional shock formation via two obliquely merging supersonic plasma jets
---
Introduction
============
We have conducted experiments on the oblique merging of two supersonic plasma jets[@Merritt-prl13] on the Plasma Liner Experiment[@hsu-pop12] (PLX) at Los Alamos National Laboratory. These experiments were the second in a series of experiments intended to demonstrate the formation of imploding spherical plasma liners via an array of merging supersonic plasma jets.[@hsu-ieee12; @cassibry12; @cassibry13] The latter has been proposed[@thio99; @thio01; @hsu-ieee12] as a standoff compression driver for magneto-inertial fusion[@lindemuth83; @kirkpatrick95; @lindemuth09] (MIF) and, in the case of targetless implosions, for generating cm-, $\mu$s-, and Mbar-scale plasmas for high-energy-density physics[@drake] research. In our first set of experiments, the parameters and evolution of a single propagating plasma jet were characterized in detail.[@hsu-pop12] The next step beyond this work, a thirty-jet experiment to form and assess spherically imploding plasma liners, has been designed[@hsu-ieee12; @cassibry13; @awe11] but not yet fielded. A related jet-merging study[@case13; @messer13; @wu13] was also conducted recently by HyperV Technologies.
The supersonic jet-merging experiments reported here are also relevant to the basic study of plasma shocks[@jaffrin] in a semi- to fully collisional regime. Related studies include counter-streaming laser-produced plasmas supporting hohlraum design for indirect-drive inertial confinement fusion[@bosch; @rancu; @wan] and for studying astrophysically relevant shocks,[@woolsey; @romagnani; @kuramitsu; @kugland; @ross] colliding plasmas using wire-array Z pinches,[@swadling13a; @swadling13b] and applications such as pulsed laser deposition[@luna07] and laser-induced breakdown spectroscopy.[@sanchez-ake] Primary issues of interest in these studies include the identification of shock formation, the formation of a stagnation layer[@hough09; @hough10; @yeates] between colliding plasmas, and the possible role of two-fluid and kinetic effects on plasma interpenetration.[@berger; @pollaine; @rambo94; @rambo95]
In this paper we present detailed measurements of the stagnation layer that forms between two obliquely merging supersonic plasma jets in a semi- to fully collisional regime. First, we briefly describe the experimental setup (Sec. \[sec:setup\]). Then we discuss observations of the stagnation layer emission morphology (Sec. \[emission\]) and density enhancements (Sec. \[density85\]). We also examine the observed stagnation layer thickness in the context of various estimated collision length scales and two-fluid plasma shock theory (Sec. \[width\]). Collectively, our observations are shown to be consistent with collisional shocks. We close with a discussion of the implications of our results on proposed imploding plasma liner formation experiments (Sec. \[discussion\]) and a summary (Sec. \[sec:summary\]).
Experimental setup {#sec:setup}
==================
Two plasma railguns are mounted on adjacent ports of a 2.7-m-diameter spherical vacuum chamber \[Fig. \[setup\](a)\], with a half-angle $\approx 12^\circ$ between the jet axes of propagation and a distance $\approx46$ cm between the gun nozzles. At the nozzle exit, individual jets have initial parameters of peak electron density $n_e\approx 2\times
10^{16}$ cm$^{-3}$, peak electron temperature $T_e\approx 1.4$ eV, diameter $=5$ cm, and axial length $\approx
20$ cm.[@hsu-pop12] In this series of experiments, the initial jet velocity $V_{\rm jet}\approx
40$ km/s and Mach number $M\equiv V_{\rm jet}/C_{\rm s,jet} > 10$, where $C_{\rm s,jet}$ is the sound speed in the jet. More details on the railguns and the characterization of single-jet propagation are reported elsewhere.[@hsu-pop12] The jets are individually very highly collisional (thermal mean free paths $\lambda_{i} \sim \lambda_{e}
\sim 100$ $\mu$m in a $\sim 20$-cm-scale plasma at initial jet merging), but the characteristic collision length ($\sim 1$ cm, see Sec. \[width\]) between counter-propagating jet ions is on the order of the thickness of the observed stagnation layer that forms between the obliquely merging jets.
![(a) Schematic showing the spherical vacuum chamber, two merging plasma jets, ($R$,$Z$) coordinates used in the paper, approximate interferometer (representing all 8 chords) and spectrometer (view ‘1’) lines-of-sight, and CCD camera field-of-view. (b) Location of interferometer chords (dots, $Z\approx 85$ cm, inter-chord spacing $= 1.5$ cm), and spectroscopy views (dashed circle, diameter $\approx 7$ cm) overlaid on a cropped CCD image of jet merging. Spectroscopy views ‘1’ and ‘2’ are located at $(R,Z) \approx (3.75~\mbox{cm}, 85~\mbox{cm})$ and $(12~\mbox{cm},
55~\mbox{cm})$, respectively.[]{data-label="setup"}](fig1.eps "fig:"){width="2.truein"}\
The key diagnostics for our merging experiments are a visible-to-near-infrared survey spectrometer (0.275 m focal length with 600 lines/mm grating and 0.45 $\mu$s gating on the 1024-pixel microchannel plate array detector), an eight-chord 561 nm laser interferometer,[@merritt-rsi12; @merritt-htpd12] and an intensified charged-coupled-device (CCD) visible-imaging camera (DiCam Pro, $1280\times1024$ pixels, 12-bit dynamic range). The CCD camera field-of-view extends from $Z \approx 0$–156 cm. The interferometer chords and spectrometer view ‘1’ intersect the stagnation layer at $Z \approx 85$ cm \[Fig. \[setup\](b)\], with an angle of $
\approx 36^\circ$ with respect to the jet-merging plane (into the page). The line formed by the interferometer chords is roughly transverse to the stagnation layer ($\approx
30^\circ$ with respect to the $R$ direction), with inter-chord spacing of $\approx 1.5$ cm, spanning $R \approx
0.75$–$11.25$ cm. The $\approx 30^\circ$ angle with respect to $R$ introduces slight temporal offsets ($\approx 0.2$ $\mu$s between adjacent interferometer chords) for interferometer data plots versus $R$. The $\approx 36^\circ$ angle between the chords and the merge plane may lead to underestimates of plasma density enhancements and overestimates of local density minima due to the chords intersecting both shocked and unshocked plasma regions. Spectrometer view ‘1’ is centered on the interferometer chord at $(R,Z) = (3.75~\mbox{cm},
85~\mbox{cm})$. Spectrometer view ‘2’ is located at $(R,Z) \approx (12~\mbox{cm}, 55~\mbox{cm})$ and is oriented $\approx 31^\circ$ relative to the merge plane. The collimated spectrometer field-of-view has a divergence of $2.4^\circ$ and a diameter $\approx7 \pm0.5 $ cm at the measurement position. Plasma jet velocity is determined via an array of intensified photodiode detectors.[@hsu-pop12] Figure \[pics\] shows a sequence of twelve CCD camera images (a different shot for each time; images are very reproducible) of the time evolution of jet merging and the formation of a stagnation layer along the jet-merging plane (midplane, horizontal in the images), with a double-peaked emission profile transverse ($R$ direction, vertical in the images) to the layer. Experiments were conducted with top jet only, bottom jet only, and both jets firing to enable the most direct comparison between single- and merged-jet measurements.
{width="6.truein"}\
We have measured the jet magnetic field strength (transverse to the jet propagation direction) using magnetic probes mounted at two locations along the exterior of the cylindrical railgun nozzle. The probe coils have nominal turns $\times$ area of 10 cm$^2$ (at the relevant frequency of 50 kHz, corresponding to the frequency of the gun current that produces the jet magnetic field), and the signals are passively integrated with a time constant of 0.322 ms. The jet is maintained at a constant diameter of 5 cm inside the nozzle. The field strength decreases from $\approx 0.14$ T at $Z=-16$ cm to $\approx 0.075$ T near the nozzle exit ($Z=0$ cm), with a decay time of 5.6 $\mu$s (see Fig. \[bfield\]). Extrapolating the decay to $t=24$ $\mu$s (i.e., $\approx 12$ $\mu$s after the jet exits the nozzle), the field would be approximately 0.01 T. Based on the parameters at initial jet merging ($n_e = 2 \times 10^{14}$ cm$^{-3}$, $T_e = 1.4$ eV,[@hsu-pop12] $B = 0.01$ T and $v = 40$ km/s), then the ratio of the jet kinetic energy density ($\rho v^2/2$) to the magnetic energy density ($B^2/2\mu_0$) is 270. The corresponding magnetic Reynold’s number $R_m\approx 1.4$ (using a jet radial length scale of 5 cm for diffusion and a propagation distance 40 cm for advection), consistent with strong resistive field decay. If instead of being spatially uniform, the axial current producing the measured transverse magnetic field is peaked and mostly contained within a radius $r_0<r_{\rm nozzle}=2.5$ cm, then the peak field inside the jet would be larger than the measured value by a factor $B_0 r_{\rm nozzle}/r_0$, where $B_0=0.0035$ T. If $r_0 = 1$ cm, then the peak $B=0.35$ T, which, extrapolating to $t=24$ $\mu$s, would give a kinetic-to-magnetic energy density ratio of $\approx 47$, still much larger than unity. We also point out that the inferred decay time of 5.6 $\mu$s ignores jet expansion and cooling, meaning that 5.6 $\mu$s is an upper bound. Thus, we ignore magnetic field effects in this paper. These magnetic field measurements were taken during hydrogen experiments (the rest of the paper reports argon results), but $T_e\approx 1.4$ eV, and thus the magnetic diffusivity, were similar in both cases.
![Peak magnetic field (transverse to rails and jet propagation direction) vs. time (shots 2444, 2445, 2446, 2448, 2449, 2450). The data (squares) are from magnetic probes mounted at two positions along the exterior of the cylindrical railgun nozzle.[]{data-label="bfield"}](fig3.eps "fig:"){width="2.1truein"}\
The argon plasma jets in these experiments likely had high levels of impurities. The post-shot chamber pressure rise for gas-injection-only was about 30% of that of a full railgun discharge, implying possible plasma impurity levels of up to 70%. Identification of bright aluminum and oxygen spectral lines in our data[@Merritt-prl13] suggests that impurities are from the zirconium-toughened-alumina (0.15 ZrO$_2$ and 0.85 Al$_2$O$_3$) railgun insulators. Because the exact impurity fractions as a function of space and time in our jets are unknown, we bound our analysis by considering the two extreme cases of (i) 100% argon and (ii) 30% argon with 70% impurities. For case (ii), we approximate the jet composition as 43% oxygen and 24% aluminum (based on their ratio in zirconium-toughened-alumina) for spectroscopy analysis.
Consistency of stagnation layer morphology with hydrodynamic shocks {#emission}
===================================================================
Oblique shock morphology
------------------------
Because inherently two-dimensional (2D) effects, such as non-uniform jet profiles, and time-dependent effects do not permit a tractable analytic treatment of our problem and require full 2D simulations, we use analytic 1D hydrodynamic theory to gain qualitative insight into the shock boundary morphology. The assumption of parallel, uniform flow within each jet \[see Fig. \[boundary\](a)\] reduces this to a 1D problem analogous to supersonic flow past a wedge or compression corner.[@landau; @nunn] Comparing with the 1D theory, we show that the observed emission layers (Fig. \[pics\]) are consistent with post-shocked plasma,[@Merritt-prl13] with their edges (at larger $|R|$) corresponding to the shock boundaries.
![(a) Simple schematic of the interaction of two obliquely interacting supersonic flows with initial Mach numbers $M_1$. Flows are incident on the midplane with angle $\delta$. A shock boundary forms at angle $\beta$ with respect to the original flow direction. Post-shock flows have Mach number $M_2$ and flow direction parallel to the midplane. (b) CCD image with postulated shock boundaries (solid white lines) and initial jet interaction distances $Z_i$ for shot 1128 at $t = 30$ $\mu$s ($Z_i\approx 30$ cm) and (c) shot 1120 at $t = 38$ $\mu$s ($Z_i \approx 21$ cm). The field-of-view is the same for both CCD images.[]{data-label="boundary"}](fig4.eps "fig:"){width="2truein"}\
Figure \[boundary\](a) shows a simple schematic of the jet interaction, where $\delta$ is the angle between the jet flow direction and the midplane, $M_1$ is the initial (pre-interaction) Mach number, and $\beta \equiv \beta(\delta, M_1)$ is the angle between the jet-flow direction and the position of an oblique shock boundary.[@nunn; @drake] Figure \[boundary\](b) shows a similar structure in a merged-jet CCD image. In this system, the turning angle $\delta \equiv \delta(Z_i)$ is given by $\tan \delta = (23 \mbox{ cm})/Z_i$, where $Z_i$ is the point at which the jets first interact, as determined by the appearance of emission \[as indicated in Figs. \[boundary\](b) and \[boundary\](c)\]. The shock boundary angle $\beta$ is given by[@landau] $$\begin{aligned}
\frac{23~\mbox{cm}}{Z_i} = 2 \cot \beta \left[ \frac{M_1^2 \sin^2 \beta - 1}{M_1^2( \gamma + \cos 2\beta) + 2} \right],
\label{betaeqn}\end{aligned}$$ and the opening angle of the shock relative to the midplane is $\beta - \delta$. Determination of $Z_i$ from plasma emission may slightly overestimate $Z_i$, but the errors introduced are small compared to the actual difference between predicted and observed values of $\beta - \delta$ (presented below). Also, a slight overestimate of $Z_i$ does not affect the discussion in Sec. \[transition\] regarding a possible shock transition.
Assuming $T_e = 1.4$ eV, mean charge $\bar{Z} = 0.94$ (both inferred from spectroscopy at $Z = 41$ cm),[@hsu-pop12] and specific heat ratio $\gamma = 1.4$,[@awe11] then a 100% argon plasma jet with $V_{\rm jet} = 40$ km/s has $M = 19$. For the 30%/70% mixture composition, $T_e = 1.4$ eV and $\bar{Z} = 0.92$ (see Sec. \[density85\]), which are similar to the 100% argon case. To place a stringent lower bound on $M$ for the 30%/70% case, we use an ion-to-proton mass ratio $\mu = \mu_{O}=16$ because oxygen is the lightest element in the impurity mixture. Thus, we estimate that $12<M<19$. We find that predicted $\beta-\delta$ values are very similar for $M=12$ and $M=19$ for a range of $Z$ \[Fig. \[angles\](a)\].
We observe that $Z_i$ falls from $Z_i \approx 45$ cm at $t = 26$ $\mu$s to $Z_i \approx 18$ cm at $t = 36$ $\mu$s \[Fig. \[angles\](b)\], consistent with jet axial expansion[@hsu-pop12] that reduces the velocity and thus increases the jet expansion angle for the rear portion of the jet. A second dip in $Z_i$ beginning at $t \approx 47$ $\mu$s is due to the arrival of a trailing jet (created by ringing in the underdamped railgun current[@hsu-pop12]) at the merge region. For early times $t \approx 24$–$33$ $\mu$s ($Z_i \approx 45$–$25$ cm), the 1D theory predicts oblique shock formation consistent with the observed wedge-shaped emission boundary, as illustrated in Fig. \[boundary\](b). In this case, the measured $\beta - \delta \approx
5^\circ$. For $M = 12$–$19$, the theoretically predicted $\beta - \delta \approx 11^\circ$, which is within approximately a factor of two of the experimentally inferred value. This is reasonable agreement given that the 1D prediction does not include 2D/3D nor plasma equation-of-state[@messer13] (EOS) effects.
![(a) Plot of $\delta$ and $\beta - \delta$ vs. $Z$ for both $M = 12$ and $M=19$ (from 1D hydrodynamic theory). The predicted threshold turning angle (also from the 1D theory), $\delta = \delta_{max} = 45^\circ$, and corresponding $Z_i(\delta_{max}) \approx 25$ cm are marked with horizontal and vertical dotted lines, respectively. (b) Plot of $Z_i$ vs. time for shots 1119–1143 and shots 1160–1182. Error bars correspond to a $\pm 7.5$ pixel offset along the $Z$ axis during image processing. The 1D theoretical cutoff for oblique shock formation, $Z_i \approx 25$ cm, is indicated by a horizontal dashed line.[]{data-label="angles"}](fig5.eps){width="2.truein"}
Speculation on shock transition {#transition}
-------------------------------
There is a possible emission morphology transition between earlier and later times \[i.e., Fig. \[boundary\](b) versus \[boundary\](c)\]. For the theoretical 1D problem with uniform, parallel flow \[with an angle $\delta$ relative to the “interaction boundary" in Fig. \[boundary\](a)\], there is a threshold $\delta_{\rm max} \approx 45^\circ$ ($Z_i \approx 25$ cm) beyond which no oblique shock forms. In 2D theory, this corresponds to a detached shock, which we treat in the 1D analysis by considering the limiting case of a normal shock. Predicting the exact structure of a detached shock in our 2D geometry, including spatial non-uniformities and time-dependence, is not a tractable analytic problem and requires 2D simulations beyond the scope of this paper. However, we can still compare our postulated morphology transition with $\delta_{\rm max}$ from the 1D theory. As shown in Fig. \[angles\](b), our observed $Z_i$ falls below (and hence $\delta$ rises above) the transition threshold (predicted by the 1D theory) around 33 $\mu$s, consistent with the approximate time of the possible morphology transition between Figs. \[boundary\](b) and \[boundary\](c). While it is far from conclusive that our observations show a shock transition or a detached shock, they are suggestive and motivate more detailed future work.
Observation of merged-jet densities exceeding that of interpenetration {#density85}
======================================================================
If the merged-jet emission layers are post-shocked plasma, then we expect an increase in density across the shock boundary during jet merging. The density increase across a 1D shock boundary should satisfy the 1D Rankine-Hugoniot relation[@drake] $$\begin{aligned}
\frac{n_2}{n_1} = \frac{(M_1 \sin \beta)^2 (\gamma + 1)}{(M_1 \sin \beta)^2 (\gamma - 1) +2},\end{aligned}$$ where $n_1$ and $n_2$ are the pre- and post-shock densities, respectively. We bound the theoretically predicted density changes in the system by using $M_1 = 12$–$19$, as well as the $\beta$ range corresponding to the observed $Z_i(\delta$). We use the conservative value of $\gamma=1.4$, as suggested by recent work in a similar parameter regime,[@awe11] as a simple way to model ionization and EOS effects in the theoretical estimates of Mach number and density enhancement. Because the $Z_i$ range encompasses $\delta > \delta_{\rm max}$, we consider the limiting case of detached shock formation (corresponding to a normal shock, i.e., $\beta=90^\circ$, in 1D theory) in addition to oblique shocks. For an oblique shock with $M = 12$, Eq. (\[betaeqn\]) gives $\beta = 34^\circ$–$59^\circ$ for measured $Z_i = 45$–$25$ cm. Thus, the range of $n_2/n_1 = 5.4$–$5.7$. Similarly, for $M = 19$ we find $\beta = 34^\circ $–$58^\circ$ and $n_2/n_1 = 5.7$–$5.9$. Assuming normal shocks ($\beta = 90^\circ$) for $Z_i <25$ cm, we find $n_2/n_1 = 5.8$–$5.9$ for $M = 12$–$19$. Thus, the overall range across the shock boundary is $n_2/n_1 = 5.4$–5.9, according to hydrodynamic 1D theory.
{width="4.5truein"}\
Next we compare $n_2/n_1$ from the 1D theory with the measured density enhancement $n_{\rm merged}/n_{\rm single}$ of the merged- over single-jet cases. We calculate the ion-plus-neutral density $n_{\rm tot}$ using an interferometer phase shift analysis accounting for multiple ionization states and the presence of impurities (see Appendix \[app:int\]). According to Eq. (\[phase-err\]), to determine $n_{\rm tot}$ we need the interferometer phase shift $\Delta \phi$, mean charge $\bar{Z}$, the correction $Err$ \[Eq. (\[eq:err\])\] accounting for all non-free-electron contributions to $\Delta \phi$, and the interferometer chord path length (approximated by the plasma jet diameter). The maximum correction $Err_{\rm max}$ is the largest scaled sensitivity, $C_{0,k}/C_e$ \[Eq. (\[eq:err-max\])\]. These are for Ar <span style="font-variant:small-caps;">i</span>: 0.08 ($\delta N_n^{STP}= 2.8 \times 10^{-4}$ at $\lambda = 561$ nm, $\rho^{STP} = 1.6$ g/L),[@crc-opt; @crc] for O <span style="font-variant:small-caps;">i</span>: 0.03 ($K_{OI}m_{O} \approx 4.4 \times 10^{-24}$ cm$^3$ for $5000\mbox{ K} < T < 10000$ K),[@ivanova] and for Al <span style="font-variant:small-caps;">i</span>: 0.007 ($\delta N_n^{STP} = 6.2 \times 10^{-2}$ at $\lambda = 561$ nm, $\rho^{STP} = 2.7$ g/cm$^3$).[@racik; @crc] Thus, $Err_{\rm max}=0.08$ (for Ar <span style="font-variant:small-caps;">i</span>).
First, we determine $\Delta \phi_{\rm single}$ and $\Delta\phi_{\rm merged}$ for the single- and merged-jet cases, respectively. The single-jet peak $\Delta\phi_{\rm single}$, averaged across chords for a single shot, is $\Delta\phi_{\rm single} \approx 4.0^\circ \pm 0.6^\circ$, where $0.6^\circ$ is the standard deviation \[Fig. \[phase\](a)\]. The peak $\Delta\phi_{\rm single}$ averaged over multiple top-jet-only shots at the $R=2.25$ cm chord is $\Delta \phi_{\rm single} =4.3^\circ \pm 0.3^\circ$ \[Fig. \[compare\](b)\], and thus we assume $\Delta\phi_{\rm single}=4^\circ$ for evaluating $n_{\rm tot,single}\equiv n_{\rm single}$. Merged-jet $\Delta\phi_{\rm merged}$ traces for a single shot show \[Fig. \[phase\](b)\] a non-uniform spatial profile with a peak near the midplane and peak magnitude $\Delta\phi \approx 14^\circ$. At $R = 2.25$ cm, the peak $\Delta \phi = 14.3 \pm 2.4^\circ$ averaged over multiple shots \[Fig. \[compare\](b)\]. Thus, we assume $\Delta\phi_{\rm merged}=14^\circ$ for evaluating $n_{\rm tot,merged}\equiv n_{\rm merged}$.
![(a) The difference between merged-jet and the sum of single-jet phase shifts $\Delta\psi$ vs. time for data averaged over shots 1117–1196 (merged-jet), shots 1277–1278 (bottom-jet) and shots 1265–1267 (top-jet). (b) Multi-shot (same data sets) averaged interferometer phase shift vs. time at $R = 2.25$ cm, for top-, bottom-, and merged-jet cases. Error bars indicate the standard deviation of $\Delta\psi$ or $\Delta\phi$ over the stated data set.[]{data-label="compare"}](fig7.eps "fig:"){width="2.3truein"}\
Before evaluating $n_{\rm single}$ and $n_{\rm merged}$, we examine $\Delta\phi$ enhancements for merged- over single-jet experiments by considering the quantity $$\Delta\psi = \Delta\phi_{\rm merged} - (\Delta\phi_{\rm top} + \Delta\phi_{\rm bottom}),$$ where $\Delta\phi_{\rm top}$ and $\Delta\phi_{\rm bottom}$ are from top-jet-only and bottom-jet-only shots, respectively. We use $\Delta \phi$ values averaged over multiple shots (Fig. \[compare\]) to reduce potential errors introduced by shot-to-shot variations. A $\Delta \psi > 0$ implies a density of the merged-jet beyond that of the sum of single jets and/or an increase in $\bar{Z}$ over that of a single jet. Merged-jet measurements over the data set considered (merged-jet: shots 1117–1196; bottom-jet: shots 1277–1278; top-jet: shots 1265–1267) show that $\Delta \psi > 0$ for $R \le 5.25$ cm \[Fig. \[compare\](a)\], implying that simple jet interpenetration cannot account for the observed stagnation layer $\Delta\phi_{\rm merged}$. For $R \ge 6.75$ cm, $\Delta \psi$ is small because this region is outside the stagnation layer.
Now we evaluate $n_{\rm single}$ and $n_{\rm merged}$ in order to estimate the density enhancement $n_{\rm single}/n_{\rm merged}$ at $Z\approx 85$ cm and $R\le 5.25$ cm. Having determined $\Delta\phi$ and $Err_{\rm max}$, we need only $\bar{Z}$ to estimate $n_{\rm single}$ and $n_{\rm merged}$. The $T_e$ and $\bar{Z}$ are determined by comparing spectral data[@hsu-pop12; @Merritt-prl13] with non-local-thermodynamic-equilibrium (non-LTE) spectral calculations in the optically thin limit using PrismSPECT.[@macfarlane03] To mitigate the impact of line-of-sight effects on our spectral analysis, we used the appearance (e.g., Ar <span style="font-variant:small-caps;">ii</span>) and absence (e.g., impurity Al <span style="font-variant:small-caps;">iii</span>) of spectral lines in the data (typically varying only in intensity) in the time range of interest to determine bounds on peak $T_e$ and $\bar{Z}$. A single jet (assuming 100% Ar) has a jet diameter $\approx 22$ cm at $Z \approx 80$ cm and $\bar{Z} $ = 0.94 ($T_e=1.4$ eV) at $Z\approx 41$ cm (the emission is too low at $Z = 85$ cm to infer $\bar{Z}$ there).[@hsu-pop12] Using $\bar{Z}=0.94$ with $\Delta\phi_{\rm single}=4.0^\circ$, we obtain $n_{\rm single}=2.1$–$2.3 \times
10^{14}$ cm$^{-3}$ (bounds provided by $Err=0$ and $Err_{\rm max}=0.08$). For the 30%/70% mixture case (at the same $T_e=1.4$ eV), $\bar{Z}=0.92$,[@Merritt-prl13] and therefore the $n_{\rm single}$ estimate changes by only a few percent.
To infer $\bar{Z} $ and $T_e$ for the merged-jet case, and therefore $n_{\rm merged}$, at $Z\approx 85$ cm, we examine spectral data from spectrometer view ‘1.’ For 100% argon, we infer that peak $T_e \ge 1.4$ eV and $\bar{Z} \ge 0.94$.[@Merritt-prl13] For the 30%/70% mixture, we infer that $2.2$ eV$\le$ peak $T_e < 2.3$ eV and $1.3\le \bar{Z} < 1.4$, with the upper bounds determined by the absence of an Al <span style="font-variant:small-caps;">iii</span> line in the data.[@Merritt-prl13] Thus, for the 100% argon case, we see little-to-no change in $\bar{Z}$ compared to the single-jet measurements, but the 30%/70% mixture calculation predicts an increase in $\bar{Z}$ during jet merging, accounting for some of the observed $\Delta\phi$ enhancement. Using $\Delta \phi = 14^\circ$, chord path length of 22 cm, and $\bar{Z} $ = 0.94 (100% argon case), we obtain $n_{\rm merged}=
7.5$–$8.2 \times 10^{14}$ cm$^{-3}$ (bounds provided by $Err=0$ and $Err_{\rm max}=0.08$). In this case the density increase $n_{\rm merged}/n_{\rm single} = 3.2$–3.8. For the most conservative $\bar{Z} =
1.4$ of the 30%/70% mixture case, $n_{\rm merged}= 5.0$–$5.3 \times
10^{14}$ cm$^{-3}$, and $n_{\rm merged}/n_{\rm single} = 2.1$–2.4. These values are summarized in Table \[ntable\].
The observed range of $n_{\rm merged}/n_{\rm single} = 2.1$–3.8 exceeds the factor of two expected for jet interpenetration, although it is smaller than the $n_2/n_1=n_{\rm shock}/n_{\rm unshocked} = 5.4$–5.9 predicted by 1D theory. Note that plasma diameter enhancement (along the interferometer chord direction) in the merged- over the single-jet case, which we have not characterized, and overestimates of $\bar{Z}$ (given that we do not have a direct measurement at $Z\approx 85$ cm) would both lead to reductions in our estimate of $n_{\rm merged}/n_{\rm single}$. The difference between the measured and predicted density jumps could be due to 3D (e.g., pressure-relief in the out-of-page dimension) and/or plasma EOS effects not modeled by 1D hydrodynamic theory.
100% Ar 30%/70%
--------------------------------- ----------------------------------- -----------------------------------
$T_{\rm e,merged}$ $\ge 1.4$ eV 2.2 eV$\le T_e <$2.3 eV
$\bar{Z}_{\rm single}$ 0.94 0.92
$\bar{Z}_{\rm merged}$ 0.94 1.4
$n_{\rm single}$ 2.1–2.3$\times 10^{14}$ cm$^{-3}$ 2.2–2.4$\times 10^{14}$ cm$^{-3}$
$n_{\rm merged}$ 7.5–8.2$\times 10^{14}$ cm$^{-3}$ 5.0–5.3$\times 10^{14}$ cm$^{-3}$
$n_{\rm single}/n_{\rm merged}$ 3.2–3.8 2.1–2.4
: Summary of the experimentally inferred jet density enhancement at $Z \approx 85$ cm for the two mixture cases: 100% Ar and 30% Ar/70% impurities. Single-jet and merged-jet densities are calculated using $\Delta\phi = 4^\circ$ and $\Delta\phi = 14^\circ$, respectively, jet diameter of 22 cm, and $Err_{\rm max} = 0.08$. Note that values for $\bar{Z}_{\rm single}$ are from $Z\approx 41$ cm.[@hsu-pop12][]{data-label="ntable"}
We point out a few additional features from the interferometry. The spatial profile for the merged-jet $\Delta\phi$, as seen in Fig. \[phase\](b), is peaked a few centimeters away from the midplane ($R=0$) and correlates with the peaked emission profile in the $R$ direction, as seen in the CCD images (Fig. \[pics\]). Figure \[phase\](a) shows evidence of variations in $\Delta \phi_{\rm peak} \approx 2.5^\circ$ over $\Delta t \approx 2$ $\mu$s in the merged-jet measurements that are not present in single-jet experiments. Assuming $V_{\rm jet}=40$ km/s, the width of the indicated structure is $\approx 8$ cm. The appearance of this $\Delta
\phi$ structure alternates between adjacent chords for chords at $R = 0.75$–3.75 cm, i.e., the $\Delta\phi$ rise in one chord corresponds to a fall in another chord at $\approx 1.5$ $\mu$s intervals. Because the inter-chord distance is 1.5 cm, the structure has a transverse velocity $\approx 15$ km/s. The underlying cause of these structures has not yet been determined.
Electron density results (determined via Stark broadening of the H-$\beta$ line) at spectrometer view ‘2’ ($Z\approx 55$ cm) also show a density enhancement: from $n_e \le 8.6 \times 10^{13}$ cm$^{-3}$ (shot 1106) for a top-jet-only case to $n_e \approx 1.6 \times 10^{15}$ cm$^{-3}$ (shot 1101) during jet merging (Fig. \[stark\]). The electron density was determined via[@hsu-pop12] $$\begin{aligned}
n_e = 6.05 \times 10^{14} [\mbox{FWHM}(\mbox{pixels})]^{3/2} \mbox{ cm}^{-3},\end{aligned}$$ where FWHM is the full-width-half-maximum of the Stark-broadened H-$\beta$ line (more details given in the caption for Fig. \[stark\]). For the top-jet-only shot (1106), the FWHM of the Lorentzian (with instrumental broadening removed) is 0.27 pixels, which is significantly less than the 1 pixel spectrometer resolution. So, we consider $n_e = 8.6 \times 10^{13}$ cm$^{-3}$ an upper bound, i.e., the density could be less but is too small to be resolvable. Thus, $n_{\rm e,merged}/n_{\rm e,single} \gtrsim 10$ at $Z \approx 55$ cm, which is significantly larger than the $n_{\rm merged}/n_{\rm single}$ observed at $Z \approx 85$ cm. Some of the $n_e=\bar{Z}n_{\rm tot}$ increase is likely due to increased ionization during jet merging, but unfortunately there was not enough information in the measured spectrum at $Z\approx 55$ cm to infer $\bar{Z}$. The $Z\approx 55$ cm measurements were taken at a larger distance from the jet axes than the $Z\approx 85$ cm measurements, which, along with possibly a different $\bar{Z}$ at $Z\approx 55$ cm, could contribute to the difference in density enhancements observed at the two different locations. Nevertheless, the magnitude of the $n_e$ enhancement suggests the presence of post-shocked density also at $Z\approx 55$ cm.
\]. Shown are the experimental data (diamonds with error bars $\pm \sqrt{\rm counts}$), an overlay of the measured instrumental broadening profile (dotted line, labeled as ‘psf’ for point spread function), and a Lorentzian H-$\beta$ profile (dashed line) that gives the best fit (minimum $\chi^2$) of the convolution (solid line) of the psf and the Lorentzian to the data.[]{data-label="stark"}](fig8.eps "fig:"){width="3.3truein"}\
Collisionality estimates and comparison to two-fluid plasma shock theory {#width}
========================================================================
Both the experimentally measured emission[@Merritt-prl13] and interferometer $\Delta \phi$ \[Fig. \[phase\](b)\] have the same gradient length scale (few cm) in the $R$ direction, and the $\Delta \phi$ dip at $R = 0.75$ cm and peak at $R = 2.25$–$3.75$ cm \[Fig. \[phase\](b)\] are well-aligned with the emission dip and peak, respectively.[@Merritt-prl13] In this section, we compare these observations with the expected scale sizes of collisional plasma shock formation via colliding plasmas. For the latter, the stagnation layer thickness is expected[@rambo94] to be on the order of the ion penetration length into the opposing jet. We find that, in our parameter regime, the limiting physics for ion penetration is frictional drag exerted by the ions of one jet on the counter-streaming ions of the other jet. This is evaluated using the slowing-down rate in the fast approximation,[@nrl-formulary] $$\nu^s_{ii^\prime} = 9.0\times10^{-8} n_i^\prime Z^2 Z^{\prime 2} \ln \Lambda \left(\frac{1}{\mu} + \frac{1}{\mu^\prime}
\right)\frac{\mu^{1/2}}{\epsilon^{3/2}},$$ where (see Appendix \[app:logL\]) $$\ln \Lambda=43- \ln \left[ \frac{Z Z^\prime (\mu + \mu^\prime)}{\mu \mu^\prime (v_{\rm rel}/c)^2}
\left(\frac{n_e}{T_e}\right)^{1/2}\right]
\label{logL-text}$$ is the Coulomb logarithm for counter-streaming ions (with relative velocity $v_{\rm rel}$) in the presence of warm electrons,[@nrl-formulary] $n_i$ and $n_e$ \[cm$^{-3}$\] the ion and electron densities, respectively, $Z$ the mean charge state, $T_e$ \[eV\] the electron temperature, $\epsilon$ \[eV\] the relative kinetic energy of the test particle, $c$ the speed of light, and the unprimed and primed variables correspond to a test particle from one jet and the field particles of the other jet, respectively. The ion penetration length is $$\lambda^s_i \approx \frac{v_{\rm rel}}{4 \sum_{i^\prime} \nu_{i i^\prime}^s},
%\lambda^s_i \approx v_{\rm rel}/ 4 \sum_{i^\prime} \nu_{i i^\prime}^s,$$ where the factor of 4 results from the integral effect of $v_{\rm rel}$ slowing down to zero,[@messer13] and the summation is over all field-ion species for the mixed-species jet case. We estimate $\lambda^s_i$ by considering jets of 100% argon and the 30%/70% mixture (specifically, 30% Ar, 43% O, 24% Al), in all cases using $v_{\rm rel} = 20$ km/s (corresponding to $\delta = 30^\circ$ and $Z_i \approx 40$ cm) and the plasma parameters listed in Table II, which also contains a summary of the ion-electron slowing-down distances $\lambda_{ie}^s$ calculated using the slow approximation for $\nu_{ie}^s$ and the Coulomb logarithm for ion-electron collisions.[@nrl-formulary] For inter-species collisions between mixed-species jets (due to impurities), we use $n_i = (\mbox{\% ion species}) \times n_{\rm tot}$.
100% Ar 30%/70% mixture
--------------------------- ---- -------------------- -- --------------------
$n_{\rm tot}$ (cm$^{-3}$) $8 \times 10^{14}$ $5 \times 10^{14}$
$T_e$ (eV) 1.4 2.2
$\bar{Z} $ Ar 0.94 1.2
Al 2.0
O 1.0
$\lambda^s_i$ (cm) Ar 1.8 0.8
Al 0.2
O 0.3
$\lambda^s_{ie}$ (cm) Ar 17.3 25.1
Al 6.6
O 14.1
: Summary of stopping lengths for inter-jet particle interactions, for both the 100% Ar and 30%/70% mixture cases.
\[lambda-table\]
We also estimate the inter-jet mean free path (mfp) of Ar$^{1+}$-Ar charge and momentum transfer. The assumption of $v_{\rm rel} = 20$ km/s gives a kinetic energy of $\approx 80$ eV, corresponding to charge and momentum transfer cross-sections $\sigma_{CT} \approx 0.3 \times 10^{-18}$ m$^{2}$ and $\sigma_{m} \approx 0.7
\times 10^{-18}$ m$^{2}$, respectively.[@phelps90] The total mfp for Ar$^{1+}$-Ar interaction is $\lambda_{in} =
1/\sigma_{\rm tot} n_n = 1/[(\sigma_{CT} + \sigma_m) n_n]$, where $n_n = (1-\bar{Z} ) n_{\rm tot}$ (for $\bar{Z} <1$) is the neutral density. For the pure-argon merged-jet parameters (an upper bound on $n_n$ because $n_n/n_{\rm tot}
< 10^{-2}$ for the mixture case), $\lambda_{in} \approx 2$ cm $\gtrsim \lambda^s_i$. Comparing all these length scale estimates with the observed few-cm-thick stagnation layer implies that our inter-jet merging is in a semi- to fully collisional regime.
Previously, we showed that the transverse ($R$) dynamics of our oblique jet merging compared favorably with 1D collisional multi-fluid plasma simulations of our experiment.[@Merritt-prl13] Specifically, reflected shocks in the simulation (propagating in the $R$ direction) gave rise to a double-peaked density profile (at $\pm R$) consistent with our density and emission profile measurements. Here, we consider our experimental observations in the context of two-fluid plasma shock theory.[@jaffrin] In the case of a high-$M$, two-fluid shock, differing ion and electron transport results in shock structures on multiple spatial scales.[@jaffrin] The length scale of ion viscosity and thermal conduction effects is on the order of the collisional mfp of the shocked ions, $\lambda_i = v_{th,i}/\nu_i$, where $v_{th,i}$ and $\nu_i$ are the ion thermal velocity and thermal collision frequency, respectively, while the length scale of electron viscosity and thermal conduction effects is on the order of $\lambda_i \sqrt{m_i/m_e}$.[@jaffrin] The downstream mfp in our system is estimated to be on the order of $8\times 10^{-3}$ cm based on the merged-jet parameters given in Table \[ntable\]. In order to bound the range of electron shock scale lengths, we use the limiting cases of $\mu = \mu_O=16$ and $\mu =\mu_{Ar}=40$, and obtain $\lambda_i \sqrt{m_i/m_e}\approx 1.4$–2.2 cm, which is of the same order as the gradient scale lengths of the observed emission[@Merritt-prl13] and $\Delta\phi$ profiles \[Fig. \[phase\](b)\]. This suggests that our observations are also consistent with collisional two-fluid plasma shocks in that the observed scales could be large enough to contain an electron-scale pre-shock.
On the use of merging plasma jets for forming spherically imploding plasma liners {#discussion}
=================================================================================
A key motivation for this work was to study two obliquely merging supersonic plasma jets as the “unit physics" process underlying the use of an array of such jets to form spherically imploding plasma liners. The latter is envisioned as a standoff driver for MIF.[@thio99; @thio01; @cassibry09; @hsu-ieee12; @hsu-pop12; @santarius12] The dynamics arising in the jet merging, e.g., shock formation, sets the properties of the subsequent, merged plasma that ultimately determines the liner uniformity and peak ram pressure ($\rho v^2$). These physics issues have been considered recently via theory and numerical modeling.[@parks08; @cassibry12; @kim13] In spherical plasma liner formation via an array of plasma jets, the initial merging would be among more than two jets, and the detailed merging geometry would depend on the port geometry of the vacuum chamber. In the case of PLX, a quasi-spherical arrangement of 60 plasma guns would result in twelve groups of five jets, with each group arranged in a pentagonal pattern.
A key figure of merit for implosion performance is the jet/liner Mach number $M$, i.e., a lower $M$ results in faster plasma spreading, density reduction, and lower ram pressure.[@parks08; @awe11; @davis12; @cassibry13] A concern is that jet merging would lead to shock formation and heating that would significantly decrease $M$ (compared to its initial value) and, thus, implosion performance. The results reported here are encouraging in that the experimentally inferred increases in $T_e$ \[by up to a factor of $(2.3~{\rm eV})/(1.4~{\rm eV})=1.64$\] and $\bar{Z}$ (by up to a factor of $1.4/0.94=1.49$) lead to an increase in $C_s\sim (\bar{Z}T_e)^{1/2}$ of 56% (we caution that more data is needed to establish a more accurate upper bound on $T_e$ in the merged case). We estimate the speed of the leading edge of the merged plasma to be $\approx 45$ km/s (see Fig. \[fig:merged-speed\]), which is close to the initial jet speed of $\approx 41$ km/s. An unchanged velocity after jet merging would result in a modest 36% reduction in $M$.
![\[fig:merged-speed\] Leading edge position of the merged jet, as determined visually from CCD images, versus time (shots 1120–1172). Diamonds are data points, and the black line is a linear fit giving the velocity of the merged-jet leading edge.](fig9.eps){width="2truein"}
With regard to uniformity, the outstanding questions are how the observed structure in two-jet merging would affect the uniformity of the leading edge of an imploding spherical plasma liner formed by multiple merging jets, and how much non-uniformity would be tolerable for compression of a magnetized plasma target for application to MIF. This problem has been studied recently in two simulation studies,[@cassibry12; @kim13] which reached opposing conclusions using two different codes employing very different numerical models and techniques. One study concluded that a series of shocks occurring during plasma liner convergence would degrade the implosion performance,[@kim13] while the other showed that initial non-uniformities arising from jet merging were largely smeared out by the time of peak compression.[@cassibry12] More detailed studies are needed to resolve the discrepancy. We envision a five-jet experiment on PLX followed by a 30- or 60-jet experiment to study this and other issues.
Summary {#sec:summary}
=======
We have made spatially resolved measurements, in a semi- to fully collisional regime, of the stagnation layer that forms between two obliquely merging supersonic plasma jets. CCD images show a double-peaked emission profile transverse to the layer, with the central emission dip consistent with a density dip observed in the interferometer data. The stagnation layer thickness is a few cm, which is of the same order as the ion penetration length (in our case determined by frictional drag between counter-streaming ions). The observed stagnation layer emission morphology shortly after jet merging is consistent with hydrodynamic oblique shock theory. The density increase from that of an individual jet to the density of the post-merge stagnation layer is greater than that of interpenetration, even accounting for the higher ionization estimates found for the high-impurity versus pure-argon analysis limits. The measured density increase is low compared to 1D theoretical hydrodynamic predictions, but discrepancies are expected due to multi- dimensional and plasma EOS effects in the experiment. We did not observe a strong rise in $T_e$ or $\bar{Z}$, which, coupled with little observed change in the jet velocity after merging, is encouraging for proposed plasma liner formation experiments.
Significant portions of this work are from E. C. Merritt’s doctoral dissertation. We acknowledge HyperV Technologies Corp. for extensive advice on railgun operation, T. P. Intrator and G. A. Wurden for sharing laboratory and diagnostic hardware, and J. T. Cassibry, J. Loverich, and C. Thoma for useful discussions. This work was supported by the U.S. Dept. of Energy.
Interferometer phase shift analysis {#app:int}
===================================
Previous interferometer phase shift analysis [@merritt-htpd12] for this experiment assumed a singly ionized argon plasma, which was adequate for our single-jet experiments.[@hsu-pop12] In these two-jet merging experiments, the observation of higher ionization states and significant impurity percentages required generalization of the phase shift analysis.
For a plasma with multiple gas species and ionization states, we can write $\Delta\phi$ as a superposition of the contributions from the electrons and all possible ionization states for each gas species in the plasma: $$\begin{aligned}
\Delta\phi_{\rm tot} &=& \Delta\phi_e - \sum_{j,k} \Delta\phi_{j,k} \\
&=& \int C_e n_e dl - \int \sum_{j,k} C_{j,k} n_{j,k} dl,\end{aligned}$$ where $C_e$ is the interferometer sensitivity constant for electrons and $C_{j,k}$ is the sensitivity constant for the $j$th ionization state ($j=0$ denotes neutrals) of the $k$th gas species.
For a species with ionization state $j$, the electron density due to that species is $n_{e,j} = j n_j$. The total electron density is then $n_e = \sum_{j,k} n_{e,(j,k)} = \sum_{j,k} j n_{j,k}$. The average ionization state of the plasma is then $$\begin{aligned}
\bar{Z} = \frac{n_e}{n_{\rm tot}} = \frac{\sum_{j,k} n_{e,(j,k)}}{\sum_{j,k} n_{j,k}} = \frac{\sum_{j,k} j n_{j,k}}{\sum_{j,k}
n_{j,k}},\end{aligned}$$ where $n_{\rm tot} = \sum_{j,k} n_{j,k}$ is the total ion-plus-neutral density of the plasma. The phase shift equation becomes $$\begin{aligned}
\Delta\phi_{\rm tot} &=&\int \left[C_e \bar{Z} n_{\rm tot} - \sum_{j,k} C_{j,k} n_{j,k}\right] dl \nonumber \\
&=& \int C_e \left[\bar{Z} - \sum_{j,k} \frac{C_{j,k}}{C_e}\frac{n_{j,k}}{n_{\rm tot}}\right] n_{\rm tot} dl \nonumber \\
&\approx& C_e [\bar{Z} - Err] \int n_{\rm tot} dl,
\label{phase-err}\end{aligned}$$ assuming a uniform $\bar{Z}$ along the path length through the plasma, and where $$Err = \sum_{j,k} \frac{C_{j,k}}{C_e}\frac{n_{j,k}}{n_{\rm tot}}.
\label{eq:err}$$
If all the $C_{j,k}$ and $n_{j,k}$ in the plasma are known, then $Err$ can be calculated exactly. However, this is typically prohibitive due to a lack of complete information for both $C_{j,k}$ and $n_{j,k}$. When $Err$ cannot be calculated exactly, it is useful to determine bounds on $Err$ (and thus $n_{\rm tot}$). Using Eq. (\[phase-err\]) and $Err=0$ (i.e., only electrons present), then the lower bound for $n_{\rm tot}$ is given by $$\begin{aligned}
\left(\int n_{\rm tot} dl\right)_{\rm min} = \frac{ \Delta\phi_{\rm tot}}{C_e \bar{Z} }.\end{aligned}$$ Similarly, if we can determine the maximum $Err=Err_{\rm max}$, then an upper bound on $n_{\rm tot}$ is given by $$\begin{aligned}
\left(\int n_{\rm tot} dl\right)_{\rm max} = \frac{ \Delta\phi_{\rm tot}}{C_e [\bar{Z} - Err_{\rm max}]}.\end{aligned}$$ One method for determining $Err_{\rm max}$ is to determine $C_{\rm max}$ for all $j,k$ present in the plasma, and then define $Err_{\rm max} \equiv C_{\rm max}/C_e$. Because $C_{\rm max} \ge C_{j,k}$ for all $j,k$ (by definition), then $$\begin{aligned}
Err \le \frac{C_{\rm max}}{C_e} \sum_{j,k} \frac{n_{j,k}}{n_{\rm tot}} = \frac{C_{\rm max}}{C_e}=Err_{\rm max}\end{aligned}$$ is always satisfied. The problem then reduces to finding $C_{\rm max}$ for the given plasma. The $C_{j,k} = (2 \pi K_{j,k} m_k)/\lambda$, where $K_{j,k}$ is the Slater screening constant, $m_k$ is the mass, and $\lambda$ is the interferometer laser wavelength. Since $K$ is proportional to the sum of mean square electron orbits for all bound electrons,[@alpher1] then for a given gas species $k$ the largest $K_{j,k}$ occurs for the neutral atom, i.e., $K_{\rm max} = K_{0,k}$. Thus, $C_{\rm max} = C_{0,k}$ for whichever gas species $k$ in the plasma has the largest neutral sensitivity constant. The maximum correction factor can be written as $$\begin{aligned}
Err_{\rm max} &=& \frac{(C_{0,k})_{\rm max}}{C_e} = \frac{2 \pi }{C_e \lambda} (K_{0,k} m_k)_{\rm max} \label{eq:err-max}\end{aligned}$$ or, using $K_{j,k}m_k = (\delta N_n^{STP}/n_n^{STP})_k$,[@kumar; @merritt-htpd12] $$\begin{aligned}
Err_{\rm max} &=& \frac{2 \pi }{C_e \lambda} \left( \frac{N_n^{STP} }{n_n^{STP}}\right)_{\rm k,max},\end{aligned}$$ where $n_n^{STP}$ is the neutral density of the species at standard temperature and pressure, $N_n$ is the refractive index of the neutral species, $\delta N_n = N_n-1$, and $C_e = \lambda e^2/(4 \pi \epsilon_0 m_e c^2)$.
Re-derivation of the Coulomb logarithm for counter-streaming ions in the presence of warm electrons {#app:logL}
===================================================================================================
We point out an inconsistency in the Coulomb logarithm for counter-streaming ions with relative velocity $v_D=\beta_Dc$ in the presence of warm electrons ($kT_i/m_i,kT_{i^\prime}/m_{i^\prime} < v_D^2 < kT_e/m_e$), as given in the NRL Plasma Formulary (2013 edition),[@nrl-formulary] $$\lambda_{ii^\prime}=\lambda_{i^\prime i}= 35 - \ln\left[\frac{ZZ^\prime(\mu+\mu^\prime)}{\mu\mu^\prime
\beta_D^2}\left(\frac{n_e}{T_e}\right)^{1/2}\right],
\label{logL}$$ where $T_e$ is in eV and units are cgs unless otherwise noted. Unprimed and primed variables refer to test and field particles, respectively. Equation (\[logL\]) affects ion collisionality estimates for counter-streaming plasmas.[@drake12]
We re-derive the Coulomb logarithm using the definition employed in the NRL Plasma Formulary,[@nrl-formulary] $$\lambda=\ln\Lambda=\ln\left(\frac{r_{\rm max}}{r_{\rm min}}\right),
\label{logL-definition}$$ where in this case $$r_{\rm max}=\lambda_{De}=\left(\frac{kT_e}{4\pi n_e e^2}\right)^{1/2}=
7.43\times10^2\left(\frac{T_e}{n_e}\right)^{1/2}
\label{lD}$$ is the electron Debye length, and $r_{\rm min}=ZZ^\prime e^2/(m_{ii^\prime} v_D^2)$ is the distance of closest approach between two counter-streaming ions with reduced mass $m_{ii^\prime}=m_i m_{i^\prime}/(m_i+m_{i^\prime})$ and relative speed $v_D$. We assume that $r_{\rm min}$ is greater than the de Broglie wavelength $\hbar/(2m_{ii^\prime} v_D)$. We re-write $r_{\rm min}$ by pulling numerical constants to the front:
$$r_{\rm min}=\frac{e^2}{m_p c^2} \frac{ZZ^\prime(\mu+\mu^\prime)}{\mu \mu^\prime (v_D/c)^2}=
\frac{(4.8032\times10^{-10})^2}{(1.6726\times 10^{-24})(2.9979\times10^{10})^2}\frac{ZZ^\prime
(\mu+\mu^\prime)}{\mu\mu^\prime \beta_D^2}=\\
1.5347\times 10^{-16}\frac{ZZ^\prime(\mu+\mu^\prime)}{\mu\mu^\prime\beta_D^2}.
\label{rmin}$$
Substituting Eqs. (\[lD\]) and (\[rmin\]) into Eq. (\[logL-definition\]), we obtain
$$\lambda_{ii^\prime}=
\ln\Lambda=\ln\left\{\frac{743(T_e/n_e)^{1/2}}{1.5347\times10^{-16}[ZZ^\prime(\mu+\mu^\prime)]/(\mu\mu^\prime\beta_D^2)}\right\}
=43-\ln\left[\frac{ZZ^\prime (\mu+\mu^\prime)}{\mu\mu^\prime \beta_D^2}\left(\frac{n_e}{T_e}\right)^{1/2}\right]
\label{logL-correct}$$
\[same as Eq. (\[logL-text\])\], which should supersede Eq. (\[logL\]).
The discrepancy between Eqs. (\[logL-correct\]) and (\[logL\]) is exactly accounted for if the constants $k=1.6022\times 10^{-12}$ erg/eV and $e^2=(4.8032\times10^{-10})^2$ statcoulomb$^2$ in Eq. (\[lD\]) are dropped.[@swadling-pc] This seems like a plausible mistake to make in arriving at Eq. (\[logL\]).
For counter-streaming Al-Al collisions with $\mu=\mu^\prime=27$, $Z=Z^\prime=2.0$, $v_D=20$ km/s, $T_e=2.2$ eV, and $n_e=6.5\times
10^{14}$ cm$^{-3}$ (corresponding to values in Table II), we calculate $\lambda_{ii} = 0.325$ and $\lambda_{ii}=8.3$ using Eqs. (\[logL\]) and (\[logL-correct\]), respectively. The latter is a more reasonable result for this weakly coupled example.
[55]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\
12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} in @noop [**]{}, (, , ) p. @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [**]{} (, , ) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [**** ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [**** ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [**]{} (, ) pp. @noop [**]{} (, , ) pp. @noop [**]{}, ed. (, ) pp. @noop [**]{}, ed. (, ) pp. @noop [****, ()]{} @noop [****, ()]{} in @noop [**]{}, (, ) p. @noop @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} **, @noop [Ph.D. thesis]{}, () @noop [****, ()]{} @noop
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'Service-based systems are software systems composed of autonomous components or services provided by different vendors, deployed on remote machines and accessible through the web. One of the challenges of modern software engineering is to ensure that such a system behaves as intended by its designer. The Reo coordination language is an extensible notation for formal modeling and execution of service compositions. Services that have no prior knowledge about each other communicate through advanced channel connectors which guarantee that each participant, service or client, receives the right data at the right time. Each channel is a binary relation that imposes synchronization and data constraints on input and output messages. Furthermore, channels are composed together to realize arbitrarily complex behavioral protocols. During this process, a designer may introduce errors into the connector model or the code for their execution, and thus affect the behavior of a composed service. In this paper, we present an approach for model-based testing of coordination protocols designed in Reo. Our approach is based on the input-output conformance (ioco) testing theory and exploits the mapping of automata-based semantic models for Reo to equivalent process algebra specifications.'
author:
- Natallia Kokash Farhad Arbab Behnaz Changizi
- Leonid Makhnist
bibliography:
- 'main.bib'
title: 'Input-output Conformance Testing for Channel-based Service Connectors'
---
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'We explore the unique and fascinating structure of neutron stars. Although neutron stars are of interest in many areas of Physics, our aim is to provide an intellectual bridge between Nuclear Physics and Astrophysics. We argue against the naive perception of a neutron star as a uniform assembly of neutrons packed to enormous densities. Rather, by focusing on the many exotic phases that are speculated to exist in a neutron star, we show how the reality is different and far more interesting.'
author:
- 'J. Piekarewicz'
bibliography:
- '../../ReferencesJP.bib'
title: The Nuclear Physics of Neutron Stars
---
[address=[Department of Physics, Florida State University, Tallahassee, FL 32306-4350, USA]{}]{}
Historical Context
==================
Almost 30 years after the seminal work by Oppenheimer and Volkoff on the structure of neutron stars[@Opp39_PR55], a young graduate student by the name of Jocelyn Bell detected a “bit of scruff” in the data arriving into her radio telescope. The arriving signal was “pulsing” with such an enormous regularity (once every 1.337302088331 seconds) that she and her research advisor Anthony Hewish were convinced that they had detected a beacon from an extraterrestrial civilization. Initially dubbed as “Little Green Man 1”, the source now known as radio pulsar “PSR B1919+21” was shortly identified as a rapidly rotating neutron star[@Hewish:1968].
The term “neutron star” appeared in writing for the first time in the 1933 proceedings of the the American Physical Society by Baade and Zwicky who wrote the now famous phrase: *With all reserve we advance the view that supernovae represent the transitions from ordinary stars into “neutron stars”, which in their final stages consist of extremely closed packed neutrons*[@Baade:1934]. Remarkably, this announcement came barely two years after the discovery of the neutron by Chadwick[@Chadwick:1932]. It appears, however, that the possible existence of dense stars that resemble “one giant nucleus” was already contemplated by Landau—even before the discovery of the neutron. For a detailed account of Landau’s role in the history of neutron stars see Ref.[@Yakovlev:2012rd].
Another towering figure in the history of neutron stars was Subrahmanyan Chandrasekhar (“Chandra”). As luck would have it, Chandra’s fundamental discovery—the prediction of a maximum white-dwarf mass known as the “Chandrasekhar limit”[@Chandrasekhar:1931]—came a year before the discovery of the neutron. Yet, fully aware of this fundamental discovery, Chandra wrote: *For a star of small mass the white-dwarf stage is an initial step towards complete extinction. A star of large mass cannot pass into the white-dwarf stage and one is left speculating on other possibilities*. We now know that these “other possibilities” refer to either a neutron star or a black hole. For his contribution to the understanding of physical processes of importance to the structure and evolution of stars, Chandra was recognized with the 1983 Nobel Prize.
Unfortunately, Jocelyn Bell was never awarded the Nobel Prize, although her advisor Anthony Hewish was recognized with the award in 1974 for “his decisive role in the discovery of pulsars”. The exclusion of Jocelyn Bell as co-recipient of the Nobel Prize was both controversial and roundly condemned by the astrophysics community. Still, Bell has always displayed enormous grace and humility in the face of this controversy. Indeed, she has stated: *I believe it would demean Nobel Prizes if they were awarded to research students, except in very exceptional cases, and I do not believe this is one of them.* It appears that Dr. Iosif Shklovsky, as well as many others, did not share her views. Dr. Shklovsky—recipient of the 1972 Bruce Medal for outstanding lifetime contributions to astronomy—told Jocelyn Bell: *Miss Bell, you have made the greatest astronomical discovery of the twentieth century.*
Introduction {#Introduction}
============
A neutron star is a gold mine for the study of physical phenomena that cut across a variety of disciplines, such as particle physics, nuclear physics, astrophysics, and general relativity among others. Although the most common perception of a neutron star is that of a uniform assembly of neutrons packed to densities that may exceed that of normal nuclei by up to an order of magnitude, the reality is far different and significantly more interesting. On the one hand, neutron stars are not expected to be particularly massive. Recent observations of two accurately measured neutron stars suggest masses near 2$M_{\odot}$. These represent the most massive neutron stars observed to date[@Demorest:2010bx; @Antoniadis:2013pzd]. Although more massive stars may very well exist, it is widely believed that the maximum mass of a neutron star will not exceed 3 solar masses[@Lattimer:2006xb]. On the other hand, neutron stars are extremely compact. While enormous progress has been made in constraining stellar radii, a simple estimate derived from the observed spin period of rapidly rotating neutron stars is highly insightful. The first discovered millisecond pulsar PSR 1937+21 is a remarkable neutron star that rotates almost 640 times a second with a spin period of $P\!=\!1.557\,806\,448\,872$milliseconds. This rate is nearly thousand times faster than Bell’s PSR B1919+21 pulsar. Given such an enormous spin frequency, the pulsar must be compact enough to allow the gravitational attraction to balance the immense centripetal acceleration. This simple argument places an upper limit on the stellar radius of $$R \lesssim \sqrt[3]{r_{s}^{\odot} \left(\frac{cP}{2\pi}\right)^{2}
\left(\frac{M}{2M_{\odot}}\right)} \approx 15\,{\rm km}\;,$$ where $r_{s}^{\odot}\!\approx\!3$km is the Schwarzschild radius of the Sun, namely, the radius at which the Sun would become a black hole. Naturally, an object with such exotic properties displays unique characteristics. In Table\[Table1\] we list some of the characteristics of a “canonical” $1.4\,M_{\odot}$ neutron star, such as the Crab pulsar—the compact remnant of the 1054 supernovae explosion that was witnessed and recorded in multiple Chinese and Japanese documents.
-------------------------------- -----------------------------
Name: PSR B0531+21 Constellation: Taurus
Distance: 2.2 kpc Age: 960 years
Mass: $1.4\,M_{\odot}$ Radius: 10km
Density: $10^{15}$g/cm${}^{3}$ Pressure: $10^{29}$atm
Surface Temperature: $10^{6}$K Escape velocity: 0.6c
Period: 33ms Magnetic Field: $10^{12}$ G
-------------------------------- -----------------------------
: Approximate characteristics of a “canonical” neutron star, such as the 960 year old Crab pulsar.[]{data-label="Table1"}
How does one describe the structure of such an exotic object. Spherically symmetric neutron stars in hydrostatic equilibrium satisfy the Tolman-Oppenheimer-Volkoff (TOV) equations[@Opp39_PR55; @Tol39_PR55], which represent an extension of Newton’s laws to the domain of general relativity. The TOV equations form a coupled set of first-order differential equations of the following form:
$$\begin{aligned}
& \frac{dP}{dr}=-G\,\frac{{\cal E}(r)M(r)}{r^{2}}
\left(1+\frac{P(r)}{{\cal E}(r)}\right)
\left(1+\frac{4\pi r^{3}P(r)}{M(r)}\right)
\left(1-\frac{2GM(r)}{r}\right)^{-1} \;,
\label{TOVa}\\
& \frac{dM}{dr}=4\pi r^{2}{\cal E}(r)\;,
\label{TOVb}
\end{aligned}$$
\[TOV\]
where $P(r)$, ${\cal E}(r)$, and $M(r)$ represent the pressure, energy density, and (enclosed) mass profiles of the star, and $G$ is Newton’s gravitational constant. For a given fluid element in the star, hydrostatic equilibrium is attained by adjusting the pressure gradient to exactly balance the gravitational pull. Note that the last three terms in Eq.(\[TOVa\]) (enclosed in parentheses) are of general-relativistic origin. For a neutron star with an escape velocity of about half of the speed of light, the effects from general relativity (GR) are essential. For example, ignoring GR effects would lead to a “Chandrasekhar limit” for a neutron star supported exclusively by neutron degeneracy pressure of $M_{\rm ch}\!\approx\!4\!\times\!1.4\,M_{\odot}\!=\!5.6\,M_{\odot}$. By properly incorporating GR corrections—which effectively enhance the pull from gravity—Oppenheimer and Volkoff determined a maximum neutron-star mass of only $M_{\rm max}\!\approx\!0.7\,M_{\odot}$[@Opp39_PR55]. Given that most accurately measured neutron-star masses fall in the $1.3\!-\!\!1.6\,M_{\odot}$ range[@Lattimer:2012nd], it is interesting to quote from Oppenheimer and Volkoff: “It seems likely that our limit of $\sim\!0.7M_{\odot}$ is near the truth”[@Opp39_PR55]. Although there is nothing wrong with the calculation by Oppenheimer and Volkoff, it does suffer from a critical omission: the role of nuclear interactions in modifying the equation of state (EOS) of a free Fermi gas of neutrons. Indeed, at the enormous densities encountered in a neutron star, the strong short-repulsion of the nucleon-nucleon (NN) force can not be ignored. It is precisely the critical role of nuclear interactions that makes neutron stars such a fruitful ground for the study of nuclear physics at the extremes of densities and isospin asymmetry. Remarkably, the only input that neutron stars are sensitive to is the equation of state of neutron-rich matter, namely, a relation between the pressure and the energy density $P\!=\!P({\cal E})$. Conversely and equally remarkable, each EOS generates a *unique* mass-[vs]{}-radius relation[@Lindblom:1992].
In hydrostatic equilibrium the neutron star is perfectly balanced by the action of two enormous forces: gravity and pressure. As indicated in Eq.(\[TOVa\]), hydrostatic equilibrium demands that the pressure gradient $dP/dr$ be negative, so the pressure decreases monotonically with distance until it vanishes at the edge of the star. In particular, the highest pressure—and density—at the center of the star must be enormous in order to be able to support the full weight of the star. This requires densities that may significantly exceed the one at the center of the nucleus (i.e., $\rho_{{}_{\!0}}\!=\!2.48\times 10^{14}{\rm g/cm^{3}}$). This implies that models of the EOS will have to encompass density regions—both high and low—inaccessible to experiment. What novel phases emerge under such extreme conditions is both fascinating and unknown. In what follows we embark on a journey of a neutron star that highlights such fascinating phases and discusses the tools that are required to uncover their observable signatures.
A Journey through a Neutron Star {#Tour}
================================
![(Left panel) An accurate rendition of the structure and phases of a neutron star—courtesy of Dany Page. (Right panel) An accurate depiction of the composition of the crust of a neutron star—courtesy of Sanjay Reddy.[]{data-label="Fig1"}](Fig1a.pdf "fig:"){width="0.35\columnwidth" height="6cm"} ![(Left panel) An accurate rendition of the structure and phases of a neutron star—courtesy of Dany Page. (Right panel) An accurate depiction of the composition of the crust of a neutron star—courtesy of Sanjay Reddy.[]{data-label="Fig1"}](Fig1b.pdf "fig:"){width="0.40\columnwidth" height="6cm"}
The structure of neutron stars is both interesting and complex. Fig.\[Fig1\]a provides what is believed to be an accurate rendition of the structure of a neutron star. The outermost surface of the neutron star contains a very thin atmosphere of only a few centimeters thick that is composed of Hydrogen, but may also contain heavier elements such as Helium and Carbon. The electromagnetic radiation that we observe may be used to constrain critical parameters of the neutron star. For example, assuming blackbody emission from the stellar surface at a temperature $T$ provides a determination of the stellar radius from the Stefan-Boltzmann law: $L\!=\!4\pi\sigma R^{2}T^{4}$. Unfortunately, complications associated with distance measurements and distortions of the black-body spectrum make the accurate determination of stellar radii a challenging task. Just below the atmosphere lies the $\sim\!100$m thick envelope that acts as “blanket” between the hot interior (with $T\!\gtrsim\!10^{8}$K) and the “cold” surface (with $T\!\gtrsim\!10^{6}$K)[@Page:2004fy]. Further below lies the non-uniform crust, a 1km region that is believed to be involved in pulsar glitches—the sudden increase in the rotational frequency of the star. The non-uniform crust sits above a uniform liquid core that consists of neutrons, protons, electrons, and muons. The core accounts for practically all the mass and for about 90% of the size of a neutron star. There is also a possibility, marked with a question mark in Fig.\[Fig1\]a, of an inner core made of exotic particles, such as hyperons, meson condensates, and quark matter[@Page:2006ud]. Given that at this time there is no clear observational evidence of their existence, such possibility will no longer be addressed here. Instead, we now continue with a detailed discussion of the structure of the crust and the *non-exotic* core.
The Outer Crust {#OuterCrust}
---------------
The outer crust is traditionally associated with the region of the star spanning about seven orders of magnitude in density; from about $10^{4}{\rm g/cm^{3}}$ to $4\!\times\!10^{11}{\rm g/cm^{3}}$ [@Baym:1971pw]. At these densities, the electrons—which form an essential component of the star in order to maintain charge neutrality—have been pressure ionized and move freely throughout the crust as a free Fermi gas. The dynamics of the outer crust is relatively simple, as it depends almost exclusively on nuclear masses. At these low sub-saturation densities, it is energetically favorable for nucleons to cluster into ${}^{56}$Fe nuclei that arrange themselves in a crystalline lattice. However, as the density increases ${}^{56}$Fe ceases to be the preferred nucleus. This is because the electronic contribution to the total energy increases rapidly with density. Thus, it becomes energetically advantageous for the energetic electrons to capture on protons and for the excess energy to be carried away by neutrinos. Thus at densities of about $10^{6}{\rm g/cm^{3}}$, ${}^{62}$Ni becomes the most stable nucleus. As the density continues to increase, the nuclear system evolves into a Coulomb lattice of progressively more exotic, neutron-rich nuclei[@RocaMaza:2008ja]; see Fig.\[Fig1\]b for an accurate depiction of the composition of the crust. Finally, at a “critical” density of about $4\!\times\!10^{11}{\rm g/cm^{3}}$, nuclei are unable to hold any more neutrons and the neutron drip line is reached. Most mass models used in the literature predict that the sequence of progressively more exotic nuclei will terminate with ${}^{118}$Kr—a nucleus with 36 protons and 82 neutrons. Note that the last isotope with a well measured mass is ${}^{97}$Kr, which is still 21 neutrons away from ${}^{118}$Kr! Thus, one most rely on mass models that are often hindered by uncontrolled extrapolations. In this regard, mass measurement on exotic nuclei are critical. A recent landmark experiment at *ISOLTRAP* measured for the first time the mass of ${}^{82}$Zn[@Wolf:2013ge]. The addition of this one mass value alone resulted in an interesting modification in the composition profile of the outer crust; see also Ref.[@Pearson:2011zz].
The Inner Crust {#Inner Crust}
---------------
The inner stellar crust comprises the region from neutron-drip density up to the density at which uniformity in the system is restored; about a third of nuclear matter saturation density $\rho_{{}_{\!0}}$. On the top layers of the inner crust nucleons continue (as in the outer crust) to cluster into a Coulomb crystal of neutron-rich nuclei embedded in a uniform electron gas. Now, however, the crystal is in chemical equilibrium with a superfluid neutron vapor. Note that although the precise details of the pulsar glitch mechanism are unclear[@Andersson:2012iu], the common perception is that glitches develop as a result of the tension created by the differential rotation between the superfluid component and the normal component, which continuously spins down due to the emission of electromagnetic radiation. As the density increases, the spherical nuclei that form the crystal lattice start to deform in an effort to reduce the Coulomb repulsion. As a result, the system starts to exhibit rich and complex structures that emerge from a dynamical competition between the short-range nuclear attraction and the long-range Coulomb repulsion. At the long length scales characteristic of the outer crust the system organizes itself into a crystalline lattice of well-separated spherical nuclei. At the other extreme of densities, uniformity in the core gets restored and the system behaves as a uniform Fermi liquid of nucleons and leptons. However, the transition region from the highly ordered crystal to the uniform liquid core is complex and poorly understood. Length scales that were well separated in both the crystalline and uniform phases are now comparable, giving rise to a universal phenomenon known as *Coulomb frustration*. In the bottom layers of the crust Coulomb frustration is manifested by the emergence of complex structures of various topologies dubbed “nuclear pasta”[@Ravenhall:1983uh; @Hashimoto:1984]. For some recent reviews on the fascinating structure and dynamics of the neutron-star crust see Refs.[@Chamel:2008ca; @Bertulani:2012], and references contain therein. In particular, in Fig.\[Fig2\] we display a snapshot obtained from a numerical simulation of a system of $Z\!=\!800$protons and $N\!=\!3200$neutrons that illustrates how the system organizes itself into neutron-rich clusters of complex topologies that are immersed in a dilute neutron vapor[@Horowitz:2004yf; @Horowitz:2004pv]. We note that a great virtue of the numerical simulations carried out in Refs.[@Horowitz:2004yf; @Horowitz:2004pv] is that pasta formation is studied in an unbiased way without assuming any particular set of shapes. Rather, the system evolves dynamically into these complex shapes from an underlying two-body interaction consisting of a short-range nuclear attraction and a long-range Coulomb repulsion.
![A snapshot of a Monte-Carlo simulation for a system consisting of 4000 nucleons at a baryon density of $\rho\!=\!\rho_{{}_{\!0}}/6$, a proton fraction of $Z/A\!=\!0.2$, and an effective temperature of $T\!=\!1$MeV[@Horowitz:2004yf; @Horowitz:2004pv].[]{data-label="Fig2"}](Fig2.pdf){width="0.45\columnwidth"}
Although the dynamics of the pasta is indeed fascinating, its existence has never been verified either experimentally or observationally. In the laboratory, heavy-ion collisions come closest at reproducing the conditions required for pasta formation. For example, low-energy heavy-ion collisions produce dilute neutron-rich matter that may closely resemble the conditions found in the bottom layers of the inner crust. However, the matter produced in these collisions is “warm” and theoretical models may be required to extrapolate to the fully catalyzed regime found in neutron stars. However, an intriguing fact may provide the first observational manifestation of the pasta phase in neutron stars. It appears that the spin period in a certain population of neutron stars known as *isolated X-ray pulsars* is constrained to at most 12 seconds long. That is, although magnetic fields estimates suggest that X-ray pulsars could slow down to periods of about 30 seconds in a few thousand years, none is observed with periods longer than 12 seconds[@Pons:2013nea]. Pons, Viganò, and Rea have shown that the existence of a highly resistive layer in the inner crust—*likely the pasta phase*—decreases the electrical conductivity, thereby resulting in a quenching of the dipolar magnetic field that ultimately limits the spin period to a maximum of about 20 seconds[@Pons:2013nea].
The Stellar Core {#Core}
----------------
At densities of about $10^{14}{\rm g/cm^{3}}$ the pasta phase will “melt” and uniformity in the system will be restored. At these densities (of about a third to a half of $\rho_{{}_{\!0}}$) the original perception of Baade and Zwicky[@Baade:1934], namely, that of a neutron star as a uniform assembly of extremely closed packed neutrons, is finally realized in the stellar core. Note, however, that in order to maintain both chemical equilibrium and charge neutrality a small fraction of about 10% of protons and leptons is required. Whereas the crust displays fascinating and intriguing dynamics, its structural impact on the star is rather modest. Indeed, more than 90% of the size and most of the mass reside in the stellar core. However, the equation of state of neutron-rich matter at the high densities attained in the core is poorly constrained by laboratory observables. Perhaps the cleanest constraint on the EOS at high-density will emerge as we answer one of the fundamental questions in nuclear astrophysics: *what is the maximum mass of a neutron star?* Or equivalently, *what is the minimum mass of a black hole?* In this regard, enormous progress has been made with the recent observation of two massive neutron stars by Demorest et al.[@Demorest:2010bx] and Antoniadis et al.[@Antoniadis:2013pzd]. Figure\[Fig3\]a displays the major impact of the mass measurement of PSR J164-2230 $(1.97\!\pm\!0.04\,M_{\odot})$ as this measurement alone can rule out EOS that are too soft to support a $2\,M_{\odot}$ neutron star—such as those with exotic cores. Undoubtedly, the quest for even more massive neutron stars will continue with the commissioning of the *Large Observatory for X-ray Timing* (LOFT) that will provide valuable insights into the behavior of ultra-dense matter.
![(Left panel) Predictions for the mass-vs-radius relation for a variety of models of the EOS using both exotic and non-exotic cores[@Demorest:2010bx]. (Right panel) Constraints on both stellar masses and radii extracted from various analyses of X-ray bursts[@Ozel:2010fw; @Steiner:2010fz; @Suleimanov:2010th]. Also shown are constraints obtained from the measurement of two massive neutron stars by Demorest[@Demorest:2010bx] and Antoniadis[@Antoniadis:2013pzd].[]{data-label="Fig3"}](Fig3a.pdf "fig:"){width="0.45\columnwidth"} ![(Left panel) Predictions for the mass-vs-radius relation for a variety of models of the EOS using both exotic and non-exotic cores[@Demorest:2010bx]. (Right panel) Constraints on both stellar masses and radii extracted from various analyses of X-ray bursts[@Ozel:2010fw; @Steiner:2010fz; @Suleimanov:2010th]. Also shown are constraints obtained from the measurement of two massive neutron stars by Demorest[@Demorest:2010bx] and Antoniadis[@Antoniadis:2013pzd].[]{data-label="Fig3"}](Fig3b.pdf "fig:"){width="0.48\columnwidth"}
In contrast, the accurate determination of stellar radii has proven more challenging. However, observations of a variety of spectroscopic phenomena in X-ray bursters appear to provide a promising approach for resolving both the mass and radius of a neutron star. Results from the analysis of three X-ray bursters by Özel, Baym, and Güver[@Ozel:2010fw] (shown in Fig.\[Fig3\]b) suggest very small radii—of $8$ to $\!10$km—that are difficult to reconcile with the predictions from models lacking exotic cores[@Fattoyev:2010rx]. Indeed, *none of the models* displayed in Fig.\[Fig3\] can account for such a small stellar radius and maximum masses of at least $2\,M_{\odot}$. However, it has been recognized that systematic uncertainties in the analysis of X-ray bursters, first by Steiner, Lattimer, and Brown[@Steiner:2010fz] and shortly after by Suleimanov et al.[@Suleimanov:2010th], may invalidate the results of Ref.[@Ozel:2010fw]. The results by Steiner et al.[@Steiner:2010fz], depicted in Fig.\[Fig3\]b by the two shaded areas that indicate 1$\sigma$ and 2$\sigma$ contours, suggest stellar radii in the $10$-$\!13$ km range. However, to complicate things even further, Suleimanov et al. have proposed a *lower limit* on the stellar radius of 14km for neutron stars with masses below 2.3M$_{\odot}$, concluding that neutron-star matter is characterized by a stiff EOS[@Suleimanov:2010th]. One must then conclude that whereas thermal emissions during X-ray bursts may become a powerful tool in the determination of stellar radii, at present such analyses are plagued by numerous uncertainties.
However, the situation seems to have improved recently with the study of accreting neutron stars during quiescence, the so-called *quiescent low-mass X-ray binaries* (qLMXBs). By studying the thermal spectra of five qLMXBs inside globular clusters, Guillot and collaborators reported a *common radius* for all five sources of $R_{NS}\!=\!9.1^{+1.3}_{-1.5}$km at a 90% confidence level[@Guillot:2013wu]. Such small radius seriously challenges our current understanding of the equation of state of dense matter. Indeed, it appears that few realistic models (if at all!) can accommodate both a small stellar radius and a large limiting mass[@Wiringa:1988tp]. One should mention that whereas the approach developed in Ref.[@Guillot:2013wu] provides a careful accounting of all uncertainties, some of the assumptions—such as a single common radius—and some of the adopted uncertainties have been called into question[@Lattimer:2013hma]. However, we are confident that with the launching of space missions such as LOFT and GAIA—with the latter providing *unprecedented positional measurements for about one billion stars in our Galaxy*—many of the current problems will be mitigated.
Whereas laboratory experiments are of marginal utility in constraining the limiting mass of a neutron star, they play an essential role in constraining stellar radii. This is because the radius of a neutron star is sensitive to the density dependence of the *symmetry energy* in the immediate vicinity of nuclear-matter saturation density[@Lattimer:2006xb]. Note that the symmetry energy represents the energy cost required to convert symmetric nuclear matter into pure nuclear matter. A critical property of the EOS that has received considerable attention over the last decade is the slope of the symmetry energy at saturation density[@Piekarewicz:2008nh]. The slope of the symmetry energy $L$ is directly related to the pressure of pure neutron matter at $\rho_{{}_{\!0}}$ and, as such, is strongly correlated to a myriad of neutron-star observables, such as its composition and cooling dynamics[@Fattoyev:2012rm]. Moreover, $L$ is also strongly correlated to the thickness of the neutron skin of heavy nuclei[@Brown:2000; @Furnstahl:2001un]—which is defined as the difference between the neutron and proton root-mean-square radii. The physical reason behind this correlation is particularly insightful. Heavy nuclei contain an excess of neutrons as a result of the repulsive Coulomb interaction among the protons. Energetically, it is advantageous—to both the surface tension and to the symmetry energy—to form an isospin symmetric ($N\!=\!Z$) core. So the basic question is *where do the extra neutrons go?* Placing them in the core reduces the surface tension but increases the symmetry energy. In contrast, moving them to the surface increases the surface tension but reduces the symmetry energy—which is lower in the dilute surface than in the dense core. So the thickness of the neutron skin emerges from a competition between the surface tension and the *difference* between the value of the symmetry energy at the surface relative to that at the center (i.e., $L$). If such a difference is large, then it is favorable to move many neutrons to the surface, thereby creating a thick neutron skin. This suggests a powerful correlation: *the larger the value of $L$ the thicker the neutron skin*[@Horowitz:2001ya]. Indeed, the strong correlation between $L$ and the neutron-skin thickness of ${}^{208}$Pb ($R_{\rm skin}^{208}$) is displayed in Fig.\[Fig4\]a where a large and representative set of density functionals were used to predict both[@RocaMaza:2011pm]. The strong correlation coefficient of $r\!=\!0.979$ suggests how a laboratory observable such as $R_{\rm skin}^{208}$ may serve to determine a fundamental property of the equation of state.
![(Left panel) Predictions from a large number of nuclear density functionals for the neutron-skin thickness of ${}^{208}$Pb and the slope of the symmetry energy $L$[@RocaMaza:2011pm]. Constraints from an updated PREX measurement (“PREX-II”) have been incorporated into the plot. (Right panel) Correlation coefficients between $R_{\rm skin}^{208}$ and a variety of neutron-star properties as predicted by the FSUGold density functional[@Todd-Rutel:2005fa].[]{data-label="Fig4"}](Fig4a.pdf "fig:"){width="0.5\columnwidth"} ![(Left panel) Predictions from a large number of nuclear density functionals for the neutron-skin thickness of ${}^{208}$Pb and the slope of the symmetry energy $L$[@RocaMaza:2011pm]. Constraints from an updated PREX measurement (“PREX-II”) have been incorporated into the plot. (Right panel) Correlation coefficients between $R_{\rm skin}^{208}$ and a variety of neutron-star properties as predicted by the FSUGold density functional[@Todd-Rutel:2005fa].[]{data-label="Fig4"}](Fig4b.pdf "fig:"){width="0.35\columnwidth"}
Recently, the Lead Radius Experiment (“PREX”) at the Jefferson Laboratory has provided the first model-independent evidence in favor of a neutron-rich skin in ${}^{208}$Pb[@Abrahamyan:2012gp; @Horowitz:2012tj]. Given that the neutral weak vector boson couples strongly to the neutron, parity violating electron scattering provides a clean probe of neutron densities that is free from strong-interaction uncertainties. As the proton radius of ${}^{208}$Pb is known extremely accurately from conventional (parity conserving) electron scattering, PREX effectively determined the neutron-skin thickness of ${}^{208}$Pb to be: $R_{\rm skin}^{208}\!=\!{0.33}^{+0.16}_{-0.18}\,{\rm fm}$[@Abrahamyan:2012gp]. In the future, an updated and approved PREX measurement—“PREX-II”—will improve the determination of $R_{\rm skin}^{208}$ by a factor of 3 (i.e., an error of about 0.06fm is anticipated). The impact of such a measurement in constraining the symmetry pressure $L$ is displayed with green error bars in Fig.\[Fig4\]a. Remarkably, the same pressure that is responsible for creating a neutron-rich skin supports a neutron star against gravitational collapse. Thus, models that predict thicker neutron skins often produce neutron stars with larger radii[@Horowitz:2000xj; @Horowitz:2001ya]. This makes possible to establish “data-to-data” relation between the neutron-rich skin of a heavy nucleus and the radius of a neutron star. This fact is nicely illustrated in Fig.\[Fig4\]b that shows the enormous reach of an accurate measurement of $R_{\rm skin}^{208}$. The correlations displayed in the figure were investigated through a covariance analysis[@Fattoyev:2012rm] using the accurately-calibrated FSUGold density functional[@Todd-Rutel:2005fa]. Besides the expected correlation between $R_{\rm skin}^{208}$ and the neutron-skin thickness of other neutron-rich nuclei, the figure displays a strong correlation with a variety of neutron-star observables—including neutron-star radii. For a detailed explanation of the physics behind this plot see Ref.[@Fattoyev:2012rm].
Conclusions {#Conclusions}
===========
Neutron stars provide a powerful intellectual bridge between Nuclear Physics and Astrophysics. In this manuscript we explored the fascinating structure of neutron stars and discussed how critical laboratory experiments and astronomical observations may constrained the EOS. In the particular case of the outer crust, we established the fundamental role that mass measurements of exotic nuclei at rare isotope facilities will play in elucidating its distinct composition. Next, we moved to the deeper inner crust with its complex pasta phase shapes that emerge from Coulomb frustration and which display unique dynamical features. Although finding clear signatures of its existence has proved elusive, we discussed how the lack of X-ray pulsars with long spin periods may be the first observable manifestation of the nuclear pasta phase. Finally, we moved into the deep stellar core and reported on a recent analysis of quiescent low-mass X-ray binaries that argues in favor of small neutron-star radii. This finding—suggesting a soft EOS—poses serious challenges to theoretical models that must at the same time account for the existence of massive neutron stars—which instead suggests a stiff EOS. Moreover, small stellar radii may also be at odds with the PREX finding of a rather large neutron-skin thickness in ${}^{208}$Pb (although the errors are large). If future laboratory experiments and astronomical observations confirm that both $R_{\rm skin}^{208}$ is thick and stellar radii are small, this would strongly suggest a softening of the EOS due to the onset of a phase transition. However, the EOS must eventually stiffen to account for the existence of massive neutron stars. Such extraordinary behavior will confirm the unique role of neutrons stars as gold mines for the study of the EOS of dense nucleonic matter.
The author thanks the organizers of the Seventh European Summer School on Experimental Nuclear Astrophysics for their kind hospitality. This work was supported in part by grant DE-FD05-92ER40750 from the Department of Energy.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'We analyze how thermal fluctuations near a finite temperature nematic phase transition affect the spectral function $A(\bk,\om)$ for single-electron excitations in a two-dimensional metal. Perturbation theory yields a splitting of the quasi-particle peak with a $d$-wave form factor, reminiscent of a pseudogap. We present a resummation of contributions to all orders in the Gaussian fluctuation regime. Instead of a splitting, the resulting spectral function exhibits a pronounced broadening of the quasi-particle peak, which varies strongly around the Fermi surface and vanishes upon approaching the Brillouin zone diagonal. The Fermi surface obtained from a Brillouin zone plot of $A(\bk,0)$ seems truncated to Fermi arcs.'
author:
- Hiroyuki Yamase
- Walter Metzner
title: Fermi surface truncation from thermal nematic fluctuations
---
The concept of nematic order in interacting electron liquids has attracted considerable interest over the last decade, mostly in the context of (quasi) two-dimensional systems [@fradkin10; @vojta09]. In a nematic state an orientational symmetry of the system is spontaneously broken, without breaking however the translation invariance. One route toward a nematic state is via partial melting of stripe order in a doped antiferromagnetic Mott insulator [@kivelson98]. Alternatively, a nematic state can be obtained from a Pomeranchuk [@pomeranchuk59] instability generated by forward scattering interactions in a normal metal [@yamase00; @halboth00]. On a square lattice, the most natural candidate for a Pomeranchuk instability has $d_{x^2-y^2}$ symmetry.
Signatures of nematic order with a $d_{x^2-y^2}$ symmetry have been observed in several strongly interacting electron materials. A nematic phase with a sharply defined phase boundary has been established for $\rm Sr_3 Ru_2 O_7$ in a strong magnetic field [@ruthenate]. Nematic order has also been observed in the high temperature superconductor $\rm Y Ba_2 Cu_3 O_y$ in transport experiments [@daou10] and neutron scattering [@hinkov]. Due to the slight orthorhombicity of the $\rm CuO_2$ planes one cannot expect a sharp nematic phase transition in $\rm Y Ba_2 Cu_3 O_y$. However, the strong temperature dependence of the observed in-plane anisotropy indicates that the system develops an intrinsic electronic nematicity, which drastically enhances the in-plane anisotropy imposed by the structure [@hackl09; @yamase06].
Nematic fluctuations close to a continuous nematic quantum phase transition naturally lead to non-Fermi liquid behavior [@oganesyan01; @metzner03; @garst10]. For a $d$-wave Pomeranchuk instability on a square lattice, the decay rate of electronic excitations is strongly momentum dependent along the Fermi surface. At the quantum critical point, the decay rate for single-particle excitations is proportional to $d_{\bk}^2 |\om|^{2/3}$, where $\om$ is the excitation energy and $d_{\bk}$ is a form factor with $d$-wave symmetry, such as $d_{\bk} = \cos k_x - \cos k_y$ [@metzner03; @metlitski10]. Landau quasi-particles are thus unstable everywhere on the Fermi surface except at the Brillouin zone diagonals, where $d_{\bk}$ vanishes. The temperature dependence of the decay rate in the quantum critical regime near the quantum critical point also differs strongly from Fermi liquid behavior [@dellanna06; @dellanna07].
Previous works on non-Fermi liquid behavior caused by nematic fluctuations focussed on the quantum critical point and the quantum critical regime at finite temperature. In this paper we analyze the spectral function for single-particle excitations in the [*thermal*]{} fluctuation regime near a nematic phase transition at finite temperatures. We show that a perturbative calculation of the self-energy in that regime leads to a splitting of the quasi-particle peak with a $d$-wave form factor in the single-particle excitation spectrum, reminiscent of a pseudogap. However, in a self-consistent calculation the split peak in the spectral function is replaced by a single broad peak. In the Gaussian fluctuation regime, a summation of vertex corrections to all orders is possible and confirms the self-consistent result. The Fermi surface obtained from the peak of the spectral function at zero excitation energy is thus smeared by a smooth broadening, which is most pronounced near the points $(\pi,0)$ and $(0,\pi)$ of the Brillouin zone, while it gradually decreases toward the Brillouin zone diagonal.
We consider a one-band system of electrons on a square lattice with a tight-binding dispersion $\eps_{\bk}$ and an effective interaction of the form [@metzner03] $$\label{H_I}
H_I =
\frac{1}{2L} \sum_{\bq} g(\bq) \, n_d(\bq) \, n_d(-\bq) \; ,$$ where $n_d(\bq) = \sum_{\bk,\sg} d_{\bk} \,
c_{\bk-\bq/2,\sg}^{\dag} c_{\bk+\bq/2,\sg}$ are $d$-wave density fluctuation operators, and $L$ is the number of sites. The function $g(\bq)$ is negative and peaked at $\bq = 0$, so that forward scattering dominates. An effective interaction of the form $H_I$ can be obtained from microscopic models such as the Hubbard or $t$-$J$ model [@yamase00; @halboth00].
For sufficiently negative values of $g = g(\b0)$ the interaction generates a $d$-wave Pomeranchuk instability leading to a nematic state with a spontaneously broken orientation symmetry [@metzner03; @kee03; @khavkine04; @yamase05]. A suitable order parameter characterizing the symmetry breaking is provided by the expectation value $\bra n_d(\b0) \ket$. Close to the transition (if continuous), strong $d$-wave density fluctuations with a long wavelength develop, which lead to a singular effective interaction. In the quantum critical regime the effective interaction is dynamical and of the form [@metzner03; @dellanna06] $$\label{D}
D_{\bk\bk'}(\bq,\nu_n) =
\frac{g \, d_{\bk} d_{\bk'}}
{(\xi_0/\xi)^2 + \xi_0^2 |\bq|^2 + |\nu_n|/(u|\bq|)} \; ,$$ where $\nu_n = 2\pi n T$ is a bosonic Matsubara frequency; $\xi$ is the nematic correlation length, while $\xi_0$ and $u$ are non-universal parameters determined by the momentum dependence of $g(\bq)$ and the band structure. In the thermal fluctuation regime near the finite temperature phase transition, quantum ($\nu_n \neq 0$) fluctuations are cut off by temperature, such that only the classical part of the effective interaction, $$\label{D_cl}
D_{\bk\bk'}(\bq) =
D_{\bk\bk'}(\bq,0) =
\frac{\tilde g \, d_{\bk} d_{\bk'}}{\xi^{-2} + |\bq|^2} \; ,$$ with $\tilde g = g/\xi_0^2$, is important.
The nematic transition on a square lattice belongs to the two-dimensional Ising universality class. A thermal phase transition at a critical temperature $T_c > 0$ is possible, since the dimensionality of the system is above the lower critical dimension (one). The correlation length $\xi$ diverges at $T_c$. Approaching the critical temperature, one first passes through a Gaussian fluctuation regime, where order parameter interactions are not important. Moving closer to $T_c$, one enters the Ginzburg region, where order parameter interactions become relevant, and the fluctuation propagator acquires an anomalous scaling dimension [@onsager44]. Close to the quantum critical point, the width of the Ginzburg region is of order $T_c/|\log T_c|$ [@bauer11].
The momentum resolved spectral function for single-particle excitations can be written as $$A(\bk,\om) = - \frac{1}{\pi} \Im G(\bk,\om) =
- \frac{1}{\pi}
\Im \frac{1}{\om - (\eps_{\bk} - \mu) - \Sg(\bk,\om)} \; ,
\label{A}$$ where $G(\bk,\om)$ and $\Sg(\bk,\om)$ are the retarded Green function and self-energy, respectively. We first compute the self-energy perturbatively to first order in the effective interaction. The contribution from classical fluctuations is given by [@dellanna06] $$\Sg(\bk,\om) = - T \int \frac{d^2q}{(2\pi)^2} \,
D_{\bk\bk}(\bq) \, G(\bk-\bq,\om) \; .
\label{Sigma}$$ In a non-selfconsistent evaluation of Eq. (\[Sigma\]) one approximates $G$ by the non-interacting Green function $G_0(\bk,\om) = [\om - (\eps_{\bk} - \mu) + i0^+]^{-1}$. The self-energy can then be computed analytically. The imaginary part has been obtained already previously [@dellanna06]. For momenta close to the Fermi surface and small frequencies one finds $$\Im\Sg(\bk,\om) = \frac{\tilde g d_{\bk}^2}{4 v_{\bk}} \,
T \xi \, l(\kappa) \; ,
\label{ImSigma}$$ where $v_{\bk} = |\nabla\eps_{\bk}|$ is the velocity of the electrons, $\kappa = [\om - (\eps_{\bk} - \mu)] \xi/v_{\bk}$, and $l(\kappa) = (1 + \kappa^2)^{-1/2}$. We assume that the Fermi surface does not cross van Hove points, such that $v_{\bk}$ is finite. The real part of the self energy is obtained either by a direct evaluation of Eq. (\[Sigma\]) or by a Kramers-Kronig transformation of the imaginary part as $$\Re\Sg(\bk,\om) = \frac{\tilde g d_{\bk}^2}{4\pi v_{\bk}}
\, \xi T \, l(\kappa) \, \ln \left|
\frac{1 - \kappa l(\kappa)}{1 + \kappa l(\kappa)}
\right| \; .
\label{ReSigma}$$ In Fig. 1 we show results for the spectral function as obtained from the non-selfconsistent first order calculation of the self-energy. Here and in all further numerical results we have chosen a dispersion $\eps_{\bk} = -2t (\cos k_x + \cos k_y) - 4t' \cos k_x \cos k_y
- 2t'' (\cos 2k_x + \cos 2k_y)$ with hopping amplitudes $t=1$, $t'=-0.3$, and $t''=0.15$. The lattice constant is one, and the chemical potential $\mu$ has been chosen such that the electron density is fixed at $n = 0.9$. The corresponding Fermi surface is closed around $(\pi,\pi)$. The coupling constant is $\tilde g = - 1.2$ and the temperature $T = 0.15$. We have not attempted to compute the correlation length, since it depends on model details such as the momentum dependence of $g(\bq)$. Instead we show results for various choices of $\xi$. One can see that a pronounced splitting of the quasi-particle peak develops for increasing $\xi$, which could be interpreted as a fluctuation precursor of the symmetry-broken state. However, we now show that the splitting disappears in a self-consistent calculation, and it is not restored by vertex corrections, at least in the Gaussian fluctuation regime.
{width="2.5in"}
In a self-consistent evaluation of Eq. (\[Sigma\]), with the interacting Green function on the right hand side, one has to solve an integral equation. The problem can be simplified by decomposing the momentum transfer $\bq$ in components $q_r$ and $q_t$ normal and tangential to the Fermi surface, respectively [@dellanna06]. The dependence on $q_t$ can be neglected in the momentum argument of $G$, such that the $q_t$-integral acts only on the fluctuation propagator $D_{\bk\bk}(\bq)$, yielding $$\bar D_{\bk}(q_r) =
\int \frac{dq_t}{2\pi} D_{\bk\bk}(\bq) =
\frac{\tilde g d_{\bk}^2}{2\sqrt{\xi^{-2} + q_r^2}} \; .
\label{Dbar}$$ The momentum dependence of the self-energy $\Sg(\bk,\om)$ can be parametrized by the Fermi momentum $\bk_F$ closest to $\bk$ and the oriented distance from the Fermi surface $k_r$, which carries the sign of $\xi_{\bk}$. One then obtains the one-dimensional integral equation $$\begin{aligned}
\Sg(k_r,\om) &=& \frac{\tilde g d_{\bk_F}^2 T}{4\pi}
\int \frac{dq_r}{\sqrt{\xi^{-2} + q_r^2}}
\nonumber \\
&\times&
\frac{1}{v_{\bk_F}(k_r - q_r) - \om + \Sg(k_r - q_r,\om)}
\; .
\label{selfcons}\end{aligned}$$ Note that $\Sg$ does not depend on $k_r$ and $\om$ independently, but only on the difference $\om - v_{\bk_F}k_r$. The dependence of $\Sg$ on $\bk_F$ enters only parametrically via $v_{\bk_F}$ and $d_{\bk_F}$ and has not been written explicitly. The integral equation (\[selfcons\]) can be solved numerically. The results for $A(\bk,\om)$ differ strongly from those suggested by non-selfconsistent perturbation theory. The quasi-particle splitting observed in the perturbative calculation (Fig. 1) is wiped out completely by the self-energy feedback into $G$. The spectral function exhibits only a single peak with a maximum at $\om = v_{\bk_F} k_r$, even for a very large correlation length $\xi$.
We now consider higher order contributions not contained in the self-consistent one-loop approximation (\[Sigma\]). The sum over all self-energy contributions generated by thermal fluctuations can be written as $$\begin{aligned}
\Sg(\bk,\om) &=& - T \int \frac{d^2q}{(2\pi)^2} \,
D_{\bk\bk}(\bq) \, G(\bk-\bq,\om)
\nonumber \\
&\times& \Lam(\bk-\bq/2,\om;\bq,0) \; ,
\label{Sg_exact}\end{aligned}$$ where $\Lam(\bk,\om;\bq,\nu)$ is the irreducible charge vertex including all vertex corrections. We exploit the fact that dominant contributions are due to small momentum transfers $\bq$ of order $\xi^{-1}$, to resum vertex corrections via an asymptotic Ward identity [@castellani94; @metzner98]. For small $\bq$, the charge vertex is related to the current vertex ${\bf\Lam}$ and the propagator via the Ward identity $\nu \Lam(\bk,\om;\bq,\nu) - \bq \cdot {\bf\Lam}(\bk,\om;\bq,\nu)
= G^{-1}(\bk+\bq/2,\om+\nu/2) - G^{-1}(\bk-\bq/2,\om-\nu/2)$. In the Gaussian fluctuation regime, interactions between order parameter fluctuations are not important. In a diagrammatic representation of perturbation theory, these interactions are generated by fermionic loops with more than two vertices. Neglecting Feynman diagrams with such loops leads to two simplifications. First, the effective interaction (\[D\_cl\]) remains unrenormalized. Second, diagrams contributing to the charge and current vertices involve only an open fermionic line. Since contributions with small $\bq$ dominate, the electron velocity $\bv_{\bk}$ entering the current operator is almost conserved such that the current vertex can be expressed by the charge vertex as ${\bf\Lam}(\bk,\om;\bq,\nu) = \bv_{\bk} \Lam(\bk,\om;\bq,\nu)$. The latter relation holds for each Feynman diagram without fermionic loops. Combining this with the Ward identity one obtains, in the static limit $\nu = 0$, $$\Lam(\bk,\om;\bq,0) =
\frac{ G^{-1}(\bk-\bq/2,\om) - G^{-1}(\bk+\bq/2,\om)}
{\bv_{\bk} \cdot \bq} \; .
\label{ward}$$ Inserting Eq. (\[ward\]) into Eq. (\[Sg\_exact\]), and using the Dyson equation $G^{-1} = G_0^{-1} - \Sg$, one obtains a closed system of equations for $\Sg$ and $G$. Decomposing the momenta $\bk$ and $\bq$ in radial and tangential components, and integrating $D_{\bk\bk}(\bq)$ over $q_t$ as before (self-consistent solution), one finds a one-dimensional linear integral equation for $G$, $$\begin{aligned}
&& (\om - v_{\bk_F} k_r + i0^+) \, G(k_r,\om) = \nonumber \\
&& \hskip 1cm 1 +
T \int \frac{dq_r}{2\pi} \frac{\bar D_{\bk_F}(q_r)}{v_{\bk_F} q_r}
\, G(k_r - q_r,\om) \; ,
\label{G_exact}\end{aligned}$$ with $\bar D_{\bk_F}(q_r)$ from Eq. (\[Dbar\]).
Note that vertex corrections cannot be summed by the above method at a nematic quantum critical point or for the related problem of non-relativistic fermions coupled to a $U(1)$ gauge field. This is because in these cases the dominant momentum transfers are almost tangential to the Fermi surface, so that the term $\bv_{\bk} \cdot \bq$ becomes subleading compared to contributions originating from fluctuations of the electron velocity [@metzner98].
The integral equation (\[G\_exact\]) can be converted to a linear differential equation by a Fourier transformation. The differential equation can be solved by standard methods. The result for the spectral function reads $$A(\bk,\om) = \int_{-\infty}^{\infty} dx \, \hat A(x) \,
e^{i(\om - v_{\bk_F} k_r)x/v_{\bk_F}} \; ,
\label{A_exact}$$ where $$\hat A(x) = \frac{1}{2\pi v_{\bk_F}} \,
\exp \left[ \int_0^x dx' \int_0^{x'} dx'' \,
T \frac{\tilde g d_{\bk_F}^2}{2\pi v_{\bk_F}^2} \, K_0(x''/\xi)
\right] \, .
\label{Ax}$$ $K_0$ is a modified Bessel function. Note that $A(\bk,\om)$ depends on $k_r$ and $\om$ only via the difference $\om - v_{\bk_F} k_r$.
In Fig. 2 we plot the spectral function $A(\bk,\om)$ for the same parameters as in Fig. 1. The function exhibits only a single peak with no trace of a splitting. The splitting present in Fig. 1 is therefore an artefact of the perturbation expansion, at least in the Gaussian regime. Vertex corrections do not change the self-consistent one-loop result qualitatively. Quantitatively they tend to sharpen the peak in $A(\bk,\om)$, but only moderately. For a large correlation length $\xi$ the width of the peak in $A(\bk,\om)$ is proportional to $\sqrt{\log\xi}$, corresponding to a peak in the imaginary part of the self-energy $\Im\Sg(\bk,\om) \propto \sqrt{\log\xi}$ at $\om = v_{\bk_F} k_r$. The width of the peak in $\Im\Sg$ is also proportional to $\sqrt{\log\xi}$ and therefore increases with $\xi$. This is very different from the perturbative result for $\Im\Sg$, Eq. (\[ImSigma\]), where the height of the peak increases rapidly with $\xi$, while its width shrinks.
{width="2.5in"}
In the quantum critical regime studied previously [@dellanna06] the spectral function also exhibits a single peak with a temperature dependent broadening. In that regime the width of the peak is proportional to $T\xi$, with a correlation length $\xi$ diverging as $(T |\log T|)^{-1/2}$ upon approaching the quantum critical point at $T = 0$.
It is striking that perturbation theory indicates a fluctuation precursor of the symmetry broken state at leading order, which is however not robust with respect to higher order contributions. One may compare with the case of a charge density wave with a finite wave vector, where symmetry breaking opens a gap. Perturbation theory indicates a pseudogap above the transition temperature in such systems, for example for the flux order studied in Ref. [@greco09]. It would be interesting to analyze the fate of the pseudogap in such systems in a calculation beyond perturbation theory.
Due to the $d$-wave form factor in the effective interaction, the broadening of the spectral function varies strongly in momentum space. To illustrate this, we plot an intensity map of $A(\bk,0)$ in the first quarter of the Brillouin zone in Fig. 3. $A(\bk,0)$ as obtained from Eq. (\[A\_exact\]) diverges on the Brillouin zone diagonal, since $d_{\bk}$ vanishes there. In a more complete model, where other interaction channels should be added, this divergence will be cut off at least by regular (Fermi liquid) contributions to $\Im\Sg$ of order $T^2$. We have therefore included such a regular contribution ($-T^2$). Due to the rapid increase of the broadening of the peak in $A(\bk,0)$ upon moving away from the Brillouin zone diagonal, and the corresponding decrease in the peak height, the Fermi surface seems truncated to arcs, albeit with fuzzy ends.
{width="1.8in"}
In summary, we have computed the spectral function $A(\bk,\om)$ for single electron excitations in the presence of critical fluctuations near a thermal nematic phase transition in a two-dimensional metal. Leading order perturbation theory indicates a splitting of the quasi-particle peak. However, a resummation of contributions to all orders reveals that the splitting is an artefact of perturbation theory, at least in the Gaussian fluctuation regime. The spectral function exhibits a pronounced broadening with a $d$-wave form factor, leading to features reminiscent of Fermi arcs in the Brillouin zone. The qualitative shape of the spectral function does not depend on the specific choice of parameters. Away from the Brillouin zone diagonal, the imaginary part of the self-energy has a peak at $\om = \eps_{\bk} - \mu$, in clear contrast to the conventional Fermi liquid form.
It is remarkable that the effect of Gaussian thermal fluctuations on electronic excitations could be treated non-perturbatively. The method used to sum contributions from thermal fluctuations to all orders is not restricted to nematic fluctuations, but could be applied equally well to other critical thermal fluctuations with a small wave vector, for example, close to a structural phase transition.
A continuous finite temperature phase transition is well established at the roof of the nematic dome found for $\rm Sr_3 Ru_2 O_7$ in a strong magnetic field [@ruthenate]. That system thus provides an opportunity to observe the fluctuation effects computed in this work, but also an experimental challenge, since the standard tool to measure the momentum resolved spectral function, that is, photoemission, is hampered by the magnetic field.
Fermi arcs have been observed in photoemission measurements of the spectral function $A(\bk,\om)$ in various high-$T_c$ cuprate compounds [@damascelli03]. However, the Fermi surface truncation in these materials is associated with a pseudogap formation, while we obtain only a strongly momentum dependent broadening of the spectral function. Although experiments indicate electronic nematicity at least in some cuprates [@fradkin10], our results thus show that another mechanism needs to be invoked to explain the photoemission data.
We are grateful to A. Chubukov, A. Greco, A. Katanin, B. Obert, and A. Rosch for valuable discussions, and to J. Bauer for a critical reading of the manuscript.
[99]{}
E. Fradkin et al., Annu. Rev. Condens. Matter Phys. [**1**]{}, 153 (2010).
M. Vojta, Adv. Phys. [**58**]{}, 699 (2009).
S. A. Kivelson, E. Fradkin, and V. J. Emery, Nature (London) [**393**]{}, 550 (1998).
I. J. Pomeranchuk, Sov. Phys. JETP [**8**]{}, 361 (1959).
H. Yamase and H. Kohno, J. Phys. Soc. Jpn. [**69**]{}, 332 (2000); ibid. [**69**]{}, 2151 (2000).
C. J. Halboth and W. Metzner, Phys. Rev. Lett. [**85**]{}, 5162 (2000).
S. A. Grigera et al., Science [**306**]{}, 1154 (2004); R. A. Borzi et al., Science [**315**]{}, 214 (2007); A. W. Rost et al., Science [**325**]{}, 1360 (2009).
R. Daou et al., Nature (London) [**463**]{}, 519 (2010).
V. Hinkov et al., Nature (London) [**430**]{}, 650 (2004); V. Hinkov et al., Nat. Phys. [**3**]{}, 780 (2007); V. Hinkov et al., Science [**319**]{}, 597 (2008).
A. Hackl and M. Vojta, Phys. Rev. B [**80**]{}, 220514(R) (2009).
H. Yamase and W. Metzner, Phys. Rev. B [**73**]{}, 214517 (2006); H. Yamase, Phys. Rev. B [**79**]{}, 052501 (2009).
V. Oganesyan, S. A. Kivelson, and E. Fradkin, Phys. Rev. B [**64**]{}, 195109 (2001).
W. Metzner, D. Rohe, and S. Andergassen, Phys. Rev. Lett. [**91**]{}, 066402 (2003).
M. Garst and A. V. Chubukov, Phys. Rev. B [**81**]{}, 235105 (2010).
Recent calculations indicate that the decay rate exponent $2/3$ is slightly changed by higher order corrections, see M. A. Metlitski and S. Sachdev, Phys. Rev. B [**82**]{}, 075127 (2010).
L. Dell’Anna and W. Metzner, Phys. Rev. B [**73**]{}, 045127 (2006).
L. Dell’Anna and W. Metzner, Phys. Rev. Lett. [**98**]{}, 136402 (2007); ibid. [**103**]{}, 159904 (E) (2009).
H.-Y. Kee, E.H. Kim, and C.-H. Chung, Phys. Rev. B [**68**]{}, 245109 (2003).
I. Khavkine et al., Phys. Rev. B [**70**]{}, 155110 (2004).
H. Yamase, V. Oganesyan, and W. Metzner, Phys. Rev. B [**72**]{}, 035114 (2005).
L. Onsager, Phys. Rev. [**65**]{}, 117 (1944).
J. Bauer, P. Jakubczyk, and W. Metzner, Phys. Rev. B [**84**]{}, 075122 (2011).
C. Castellani, C. Di Castro, and W. Metzner, Phys. Rev. Lett. [**72**]{}, 316 (1994).
W. Metzner, C. Castellani, and C. Di Castro, Adv. Phys. [**47**]{}, 317 (1998).
A. Greco, Phys. Rev. Lett. [**103**]{}, 217001 (2009); M. Bejas et al., Phys. Rev. B [**83**]{}, 014514 (2011).
A. Damascelli, Z. Hussain, and Z.-X. Shen, Rev. Mod. Phys. [**75**]{}, 473 (2003).
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
address: |
FB17, Mathematik-Informatik, Universitaet Paderborn,\
D-33098 Paderborn, Germany
author:
- 'W.X. MA [^1] and B. FUCHSSTEINER'
title: BINARY NONLINEARIZATION OF LAX PAIRS
---
\#1\#2\#3\#4[[\#1]{} [**\#2**]{}, \#3 (\#4)]{}
Introduction
============
Symmetry constraints become prominent because of the important role they play in the soliton theory [@AntonowiczWojciechowski] [@KonopelchenkoStrampp] [@OevelStrampp] [@RagniscoW]. A kind of very successful symmetry constraint method for soliton equations is proposed through the nonlinearization technique called mono-nonlinearization [@Cao] [@GengMa]. However, mono-nonlinearization involves only the Lax pairs of soliton equations. We would like to elucidate that the mono-nonlinearization technique can successfully be extended to the Lax pairs and the adjoint Lax pairs associated with soliton hierarchy. The corresponding symmetry constraint procedure is called a binary nonlinearization technique [@Ma] [@MaStrampp] [@Geng] because it involves the Lax pairs and the adjoint Lax pairs and puts the linear Lax pairs into the nonlinearized Lax systems. A kind of useful symmetries in our symmetry constraints is exactly the specific symmetries expressed through the variational derivatives of the potentials. The resulting theory provides a method of separation of variables for solving nonlinear soliton equations and exhibits integrability by quadratures for soliton equations. It also narrows the gap between finite dimensional integrable Hamiltonian systems and infinite dimensional integrable soliton equations. An illustrative example is carried out in the case of the three-by-three matrix spectral problem for AKNS soliton hierarchy.
Basic idea of binary nonlinearization
=====================================
This section reveals how to manipulate a binary nonlinearization procedure for a given soliton hierarchy along with a basic idea for the proof of the main result. Let $\cal B$ denote the differential algebra of differential vector functions $u=u(x,t)$, and write for $k\ge 0$ $$%{\cal V}^s_0 =\sum ^\infty _{k=0}
{\cal V}^s_{(k)}
=\{ (P^{ij}\partial ^k )_{s\times s}
\left.\right | P^{ij}\in
{\cal B} \},\
\widetilde {\cal V}^s_{(k)}=
{\cal V}^s_{(k
)}\otimes
C[\lambda,\lambda^{-1}],\ \partial =\frac d{dx} .$$ For $U=U(u,\lambda )\in
\widetilde{\cal V}_{(0)}^s$, we choose a solution to the adjoint representation equation $V_x=[U,V]$: $$V=V(u,\lambda )=\sum _{i\ge0}V_i\lambda ^{-i}, \ V_i\in
{\cal V}^s_{(0)}.$$ Suppose that the isospectral ($\lambda _{t_n}=0$) compatibility conditions $U_{t_n}-V^{(n)}_x+[U,V^{(n)}]=0,\ n\ge0,$ of the Lax pairs $$\left \{ \begin{array}{l}\phi _x=U\phi =U(u,\lambda )\phi ,\ U\in
\widetilde {\cal V}_{(0)}^s \vspace {2mm}
\\
\phi _{t_n}=
V^{(n)}
\phi =
V^{(n)}
(u,\lambda )\phi ,\
V^{(n)}=(\lambda ^{n}V)_++\Delta _n,\
\Delta _n\in \widetilde {\cal V}_{(0)}^s\end{array}
\right.$$ determine a soliton hierarchy $$u_{t_n}=K_n=JG_n=J\frac {\delta H_n}{\delta u},\ n\ge0.$$ If $\phi =(\phi _1,\phi_2,\cdots, \phi_s)^T$ and $\psi =(\psi _1,\psi_2,
\cdots, \psi_s)^T$ satisfy the spectral problem and the adjoint spectral problem $$\phi _x=U(u,\lambda )\phi,\ \psi _x=-U^T(u,\lambda )\psi,$$ and we set the matrix $\bar V =\phi \psi ^T= (\phi _k\psi _l)_{s\times s }$, then we have the following two basic results used in binary nonlinearization [@FokasAnderson] [@MaStrampp]:
\(i) the variational derivative of the spectral parameter $\lambda $ with respect to the potential $u$ may be expressed by = ,\[vdoflambda\]
\(ii) the matrix $\bar V $ is a solution to the adjoint representation equation $ V _x=[U, V]$, i.e. $\bar V _x=[U,\bar V]$.
Now introduce distinct eigenvalues $\lambda _1,\cdots, \lambda _N$ and let $$\phi^{(j)}=(\phi_{1j},\cdots,\phi_{sj})^T,
\ \psi^{(j)}=(\psi_{1j},\cdots,\psi_{sj})^T\ \, (1\le j\le N)$$ denote the eigenvectors and the adjoint eigenvectors corresponding to $\lambda _j\ (1\le j\le N)$, respectively. Make the Bargmann [**symmetry constraint**]{} K\_0=JG\_0=J\_[j=1]{}\^N E\_j G\_0=\_[j=1]{}\^N E\_j , where $E_j=
-\int _{-\infty}^\infty <\bar V(\lambda _j),\frac {\partial U}
{\partial \lambda _j
}>dx$, $\bar V(\lambda _j)=\phi^{(j)}\psi^{(j)T},$ $1\le j\le N$. The Bargmann constraint requires the covariant $G_0$ to be a potential function not including any potential differential and hence from the Bargmann symmetry constraint we may find an explicit nonlinear expression for the potential u=f (\^[(1)]{},\^[(2)]{},,\^[(N)]{};\^[(1)]{},\^[(2)]{}, ,\^[(N)]{}) .\[u\] Upon instituting (\[u\]) into the Lax pairs and the adjoint Lax pairs, we get two nonlinearized Lax systems, i.e. the nonlinearized spatial system {
[l]{} \_[jx]{}=U(f (\^[(1)]{},,\^[(N)]{};\^[(1)]{},,\^[(N)]{}) ,\_j) \_[j]{}, 1jN,\
\_[jx]{}=-U\^T(f (\^[(1)]{},,\^[(N)]{};\^[(1)]{},,\^[(N)]{}) ,\_j) \_[j]{}, 1jN;
. \[nxpart\] and the nonlinearized temporal systems for $n\ge 0$ {
[l]{} \_[jt\_n]{}=V\^[(n)]{}(f (\^[(1)]{},,\^[(N)]{};\^[(1)]{},,\^[(N)]{}) ,\_j) \_[j]{}, 1jN,\
\_[jt\_n]{}=-V\^[(n)T]{}(f (\^[(1)]{},,\^[(N)]{};\^[(1)]{},,\^[(N)]{}) ,\_j) \_[j]{}, 1jN.
. \[ntpartn\]
In order to discuss the integrability of (\[nxpart\]) and (\[ntpartn\]), we choose the symplectic structure $\omega ^2$ on ${\mbox{\rm I
\hspace{-0.9em} R}} ^{2sN}$ $$\omega ^2=\sum _{i=0}^s\sum _{j=0}^N d\phi_{ij}\wedge d\psi_{ij}=
\sum _{i=0}^sdP_i\wedge dQ_i,$$ where $P_i=(\phi_{i1},\cdots,\phi_{iN})^T,\
Q_i=(\psi_{i1},\cdots,\psi_{iN})^T$, $1\le i\le s$. We accept the following corresponding Poisson bracket for two functions $F,G$ defined over the phase space ${\mbox{\rm I
\hspace{-0.9em} R}} ^{2sN}$ $$\begin{aligned}
&&\{F,G\}=\omega ^2(IdG,IdF)=\omega ^2(X_G,X_F)\nonumber
\\&&=
\sum_{i=1}^s(<\frac {\partial F}{\partial Q_i},\frac {\partial G}
{\partial P_i}>-
<\frac {\partial F}{\partial P_i},\frac {\partial G}{\partial Q_i}>),
\label{poissonbracket}\end{aligned}$$ where $IdH=X_H$ represents the Hamiltonian vector field with energy $H$ defined by $i_{IdH}\omega ^2=i_{X_H}\omega ^2=dH$ and $<\cdot,\cdot>$ represents the standard inner product of ${\mbox{\rm I \hspace{-0.9em} R}} ^N$. Then we accept the following corresponding Hamiltonian system with the Hamiltonian function $H$ \_i={P\_i,H}=-, \_i={Q\_i,H}=, 1is.
[**Main Result:**]{} [*The nonlinearized spatial system (\[nxpart\]) is a finite dimensional integrable Hamiltonian system in the Liouville sense, and the nonlinearized temporal systems (\[ntpartn\]) for $n\ge0$ may be transformed into a hierarchy of finite dimensional integrable Hamiltonian systems in the Liouville sense, under the control of the nonlinearized spatial system (\[nxpart\]). Moreover the potential $u=f$ determined by the Bargmann symmetry constraint solves the $n$-th soliton equation $u_{t_n}=K_n$ in the hierarchy.*]{}
[**Idea of Proof:**]{} Note that we have $$(V(f,\lambda ))_{x}=[U(f,\lambda ), V(f,\lambda )],
\ (\bar V(\lambda _j))_{x}=[U(f,\lambda _j),\bar V(\lambda _j)]$$ and when $ u_{t_n}=K_n,$ we have $$( V(f,\lambda ))_{t_n}=[V^{(n)}(f,\lambda ), V(f,\lambda )],\
(\bar V(\lambda _j))_{t_n}=[V^{(n)}(f,\lambda _j),\bar V(\lambda _j)].$$ Therefore we may show that $F=\frac12 \textrm{tr}(V(f,\lambda ))^2$ is a common generating function for integrals of motion of (\[nxpart\]) and (\[ntpartn\]) since $F_x=\frac12 \textrm{tr}(V^2)_x
=\frac12\textrm{tr} [U,V^2]=0$ and $F_{t_n}=\frac12 \textrm{tr}(V^2)_{t_n}
=\frac12 \textrm{tr}[V^{(n)},V^2]=0$. A similar deduction may verify that $\bar F_j=\frac12 \textrm{tr}(\bar V(\lambda _j))^2$, $1\le j\le N$, are integrals of motion of (\[nxpart\]) and (\[ntpartn\]), too. Noticing F=\_[n0]{}F\_n\^[-n]{}, |F\_j=12 (\_[i=1]{}\^s \_[ij]{}\_[ij]{})\^2, 1jN, we get a series of explicit integrals of motion: $\bar F_j,\ 1\le j\le N,\ \{F_n\}_{n=0}^\infty,$ which may be proved to be involutive with respect to the Poisson bracket (\[poissonbracket\]). Further it is not difficult to show the Liouville integrability of (\[nxpart\]) and (\[ntpartn\]) when they can be rewritten as Hamiltonian systems with Hamiltonian functions being polynomials in $F_m,\ m\ge1$.
In addition, because the compatibility condition of (\[nxpart\]) and (\[ntpartn\]) is still the $n$-th soliton equation $u_{t_n}=K_n$, $u=f(\phi_j;\psi_j)$ gives an involutive solution to the $n$-th soliton equation $u_{t_n}=K_n$ once $\phi_j,\,\psi_j,\ 1\le j\le N,$ solve (\[nxpart\]) and (\[ntpartn\]), simultaneously. This sort of involutive solutions also exhibits a kind of separation of independent variables $x, t_n$ for soliton equations.
The case of AKNS Hierarchy
==========================
For AKNS hierarchy, we introduce a three-by-three matrix spectral problem $$\left(\begin{array}{c}
\phi _1 \\ \phi _2 \\ \phi _3\end{array}\right) _x=U\left(\begin{array}{c}
\phi _1 \\ \phi _2 \\ \phi _3\end{array}\right) =
\left(\begin{array}{ccc}
-2\lambda & \sqrt{2}q&0 \vspace{1mm}\\ \sqrt{2}r&0 &\sqrt{2}q\vspace{1mm}
\\ 0&\sqrt{2}r &2\lambda
\end{array}\right)\left(\begin{array}{c}
\phi _1 \\ \phi _2 \\ \phi _3\end{array}\right).$$ In this case, $ \phi =(
\phi _1 , \phi _2 , \phi _3)^T $ and $ u =
( q , r)^T$. A hierarchy of AKNS soliton equations [@MaStrampp] u\_[t\_n]{}= K\_n= (
[c]{} -2b\_[n+1]{}\
2c\_[n+1]{}
) = JL\^n (
[c]{} r\
q
)=J , n0 is the compatibility conditions of the Lax pairs \_x=U, \_t=V\^[(n)]{}, V\^[(n)]{}=(\^nV)\_+. Here the operator solution $V$ to $V_x=[U,V]$, the Hamiltonian operator $J$, the recursion operator $L$, and the Hamiltonian functions $H_n$ for $n\ge0$ read as $$V=\left(\begin{array}{ccc} 2a&\sqrt{2}b&0 \vspace{1mm}\\ \sqrt{2}c&0&\sqrt{2}b
\vspace{1mm}\\
0&\sqrt{2}c&-2a\end{array}\right)
%=\sum_{i=0}^\inftyV_i\lambda ^{-i}
=\sum _{i=0}^\infty
\left(\begin{array}{ccc} 2a_i&\sqrt{2}b_i&0\vspace{1mm}\\
\sqrt{2}c_i&0&\sqrt{2}b_i\vspace{1mm}\\
0&\sqrt{2}c_i&-2a_i\end{array}\right) \lambda ^{-i},$$ $$J=
\left(\begin{array}{cc} 0&-2\vspace{1mm}\\2&0\end{array}\right),\
L=
\left(\begin{array}{cc} \frac12 \partial -r\partial ^{-1}q&r\partial ^{-1}r
\vspace{1mm}\\
-q\partial ^{-1}q&-\frac12 \partial +q\partial ^{-1}r\end{array}\right) ,\
H_n=\frac {2a_{n+2}}{n+1}.$$ The operators $J$ and $JL$ constitute a Hamiltonian pair and $L^*$ is hereditary [@Fuchssteiner1].
In this AKNS case, the Bargmann symmetry constraint becomes K\_0=J =J \_[j=1]{}\^N (
[c]{} (\_[2j]{}\_[1j]{}+ \_[3j]{}\_[2j]{})\
(\_[1j]{}\_[2j]{}+ \_[2j]{}\_[3j]{})
), which engenders an explicit expression for the potential $u$ u =f(\_[ij]{};\_[ij]{})=(
[c]{} < P \_1,Q\_2>+< P \_2,Q\_3>\
< P \_2,Q \_1>+< P \_3,Q \_2>
).\[uakns\] Further besides $\bar F_j$, $1\le j\le N$, we can directly give the following explicit integrals of motion for the nonlinearized Lax systems $$\begin{array}{l}
F:=\frac12\textrm{tr}V^2=4(a^2+bc)=\sum_{m\ge0}F_m\lambda ^{-m},\vspace{2mm}\\
%F_m=4\sum _{i=0}^m(a_ia_{m-i}+b_ic_{m-i}),\vspace{2mm}\\
F_0=4,\ F_1=-8(<P_1 ,Q_1>-<P_3,Q_3>),\vspace{2mm}\\
%F_m=4\sum_{i=1}^{m-1}(a_ia_{m-i}+b_ic_{m-i})-8a_m\vspace{2mm}\\
F_m=4\sum _{i=1}^{m-1}\bigl[
(<A^{i-1}P _1,Q _1>-<A^{i-1}P_3,Q_3>)\times \vspace{2mm}\\
\qquad (<A^{m-i-1}P _1,Q _1>-<A^{m-i-1}P_3,Q_3>)\vspace{2mm}
\\ \qquad
+ 2(<A^{i-1}P_1,Q_2>+<A^{i-1}P_2,Q_3>)\times \vspace{2mm}\\
\qquad (<A^{m-i-1}P_2,Q_1>+<A^{m-i-1}
P_3,Q_2>)\bigr]\vspace{2mm}\\ \qquad
-8<A^{m-1}P _1,Q _1>-<A^{m-1}P_3,Q_3>,
\ m\ge 2,\end{array}$$ where $A=\textrm{diag}(\lambda _1,\lambda _2,\cdots,\lambda _N)$. The nonlinearized spatial system (\[nxpart\]) is rewritten as an integrable Hamiltonian system P \_[ix]{}={P\_i,H}=-, Q \_[ix]{}={Q\_i,H}=, i=1,2,3 with the Hamiltonian function $$\begin{array}{l}H=2(<A P _1,Q _1>-<AP _3,Q _3>)\vspace{2mm}\\
-2(<P _1,Q _2>+<P _2,Q _3>)(<P _2,Q _1>+<P _3,Q _2>)
,\end{array}$$ and under the control of the nonlinearized spatial system (\[nxpart\]), the nonlinearized temporal systems (\[ntpartn\]) for $n\ge 0$ can also be rewritten as the integrable Hamiltonian systems P \_[it\_n]{}={P\_i,H\_n}=-, Q \_[it\_n]{}={Q\_i,H\_n}=, i=1,2,3 with the Hamiltonian functions $$H_n= -\frac14
\sum _{m=0}^n \frac {d_m}{m+1}
\sum_{\begin{array}{c}{\scriptstyle{ i_1+\cdots+i_{m+1}=n+1}}
\vspace{-1mm}
\\ {\scriptstyle{ i_1,\cdots,i_{m+1}\ge1}}\end{array}}
F_{i_1}
\cdots F_{i_{m+1}}\, ,$$ where the constants $d_m$ are defined by $$\begin{array}{l}d_0=1,\ d_1=-\frac 18,\ d_2=\frac3{128},\vspace{2mm}\\
d_m=-\frac12
\sum_{i=1}^{m-1}d_id_{m-i}-\frac14d_{m-1}-\frac18
\sum_{i=1}^{m-2}d_id_{m-i-1},\
m\ge3.\end{array}$$ Moreover following the previous main result, the potential (\[uakns\]) with $$P_i(x,t_n)=g^x_{H}g^{t_n}_{H_n}P_i(0,0),\
Q_i(x,t_n)=g^x_{H}g^{t_n}_{H_n}Q_i(0,0),\ i=1,2,3,$$ gives rise to a sort of involutive solutions with separated variables $x, t_n$ to the $n$-th AKNS soliton equation $u_{t_n}=K_n$. Here $g^y_G$ denotes the Hamiltonian phase flow of $G$ with a parameter variable $y$ but $P_i(0,0),\,Q_i(0,0)$ may be arbitrary initial value vectors. A finite gap property for the resulting involutive solutions may also be shown.
Concluding remarks
==================
We remark that the finite dimensional Hamiltonian systems generated by nonlinearization technique depend on the starting Lax pairs. Thus the same equation may be connected with different finite dimensional Hamiltonian systems once it possesses different Lax pairs. AKNS soliton equations are exactly such examples [@MaStrampp].
We also point out that the Neumann symmetry constraint and the higher order symmetry constraints K\_[-1]{}=J\_[j=1]{}\^NE\_j, K\_m=JG\_m=J\_[j=1]{}\^NE\_j, (m1), may be considered. These two sorts of symmetry constraints are somewhat different from the Bargmann symmetry constraints because $K_{-1}$ is a constant vector and the conserved covariants $G_m,\, m\ge1, $ involve some differentials of the potential. This suggests that a few new tools are needed for discussing them [@Zeng]. Similarly, we can consider the corresponding $\tau$-symmetry (i.e. time first order dependent symmetry [@Chen]) constraints or more generally, time polynomial dependent symmetry [@Fuchssteiner2] constraints . Binary nonlinearization may also be well applied to discrete systems and non-Hamiltonian soliton equations such as the Toda lattice and the coupled Burgers equations [@Ma0]. Note that in the case of KP hierarchy, the similar Bargmann symmetry constraints have been carefully analyzed as well [@OevelStrampp], and the specific symmetries we use in constraints are sometimes called additional symmetries [@Dickey] and are often taken as source terms of soliton equations [@Mel'nikov]. It should also be noted that the nonliearized Lax systems are intimately related to stationary equations [@Tondo] and the more general nonliearized Lax systems can be generated from the linear combination of Bargmann symmetry constraints which will be shown in a late publication.
However, in the binary nonlinearization procedure, there exist two intriguing [**open problems**]{}. The first one is why the nonlinearized spatial system (\[nxpart\]) and the nonlinearized temporal systems (\[ntpartn\]) for $n\ge 0$ with the control of the nonlinearized spatial system (\[nxpart\]) always possess Hamiltonian structures? The second one is whether or not the nonlinearized temporal systems (\[ntpartn\]) for $n\ge 0$ are themselves integrable soliton equations without the control of the nonlinearized spatial system (\[nxpart\]). These two problems are important and interesting but need some further investigation.
Acknowledgments {#acknowledgments .unnumbered}
===============
One of the authors (W. X. Ma) would like to thank the Alexander von Humboldt Foundation for a research fellow award and the National Natural Science Foundation of China and the Shanghai Science and Technology Commission of China for their financial support. He is also grateful to Drs. W. Oevel, P. Zimmermann and G. Oevel for their helpful and stimulating discussions.
References {#references .unnumbered}
==========
[99]{} M. Antonowicz and S. Wojciechowski, [*J. Math. Phys.*]{} [**33**]{}, 2115 (1992). C.W. Cao, [*Sci. China*]{} A [**33**]{}, 528 (1990). H.H. Chen, Y.C. Lee and J.E. Lin, in [*Advances in Nonlinear Waves*]{}, Vol.2, ed. L. Debnath (Pitman, New York, 1985), p233. L.A. Dickey, [*Lett. Math. Phys.*]{} [**34**]{}, 379 (1995). A.S. Fokas and R.L. Anderson, [*J. Math. Phys.*]{} [**23**]{}, 1066 (1982). B. Fuchssteiner, . B. Fuchssteiner, . X.G. Geng, [*Phys. Lett.*]{} A [**194**]{}, 44 (1994). X.G. Geng and W. X. Ma, [*Nuovo Cimento*]{} A [**108**]{}, 477 (1995). B. Konopelchenko and W. Strampp, [*J. Math. Phys.*]{} [**33**]{}, 3676 (1992). W.X. Ma, [*J. Phys. A: Math. Gen.*]{} [**26**]{}, L1169 (1993). W.X. Ma, [*J. Phys. Soc. Jpn.*]{} [**64**]{}, 1085 (1995); Symmetry constraint of MKdV equations by binary nonlinearization, to appear in [*Physica*]{} A. W.X. Ma and W. Strampp, [*Phys. Lett.*]{} A [**185**]{}, 277 (1994). V.K. Mel’nikov, [*J. Math. Phys.*]{} [**31**]{}, 1106 (1990). W. Oevel and W. Strampp, [*Commun. Math. Phys.*]{} [**157**]{}, 51 (1993). O. Ragnisco and S. Wojciechowski, [*Inverse Problems*]{} [**8**]{}, 245 (1992). G. Tondo, [*Theor. Math. Phys.*]{} [**33**]{}, 796 (1994). Y.B. Zeng, [*Physica*]{} D [**73**]{}, 171 (1994).
[^1]: On leave of absence from Institute of Mathematics, Fudan University, Shanghai 200433, China
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
author:
- 'Kyoko Matsushita, Alexis Finoguenov and Hans Böhringer'
date: 'Received ; accepted '
subtitle: 'II. Abundance structure of the interstellar and intergalactic medium'
title: XMM observation of M 87
---
Introduction
============
The intracluster medium (ICM) contains a large amount of metals, which are mainly synthesized in early-type galaxies (e.g. Arnaud et al. 1992; Renzini et al. 1993). Thus, abundances of the metals are tracers of chemical evolution of galaxies and clusters of galaxies.
Based on the Si/Fe ratio observed with ASCA, a discussion on contributions from SN Ia and SN II to the metals has commenced. In a previous nucleosynthesis model of SN Ia, the Fe abundance is much larger than the Si abundance in the ejecta of SN Ia (Nomoto et al. 1984). Observations of metal poor Galactic stars indicate that average products of SN II have a factor of 2–3 larger abundance of $\alpha$-elements than Fe (e.g. Edvardsson et al. 1993; Nissen et al. 1994; Thielemann et al. 1996), although this ratio may depend on the initial mass function (IMF) of stars. From elemental abundance ratios in the ICM of four clusters of galaxies observed with ASCA, Mushotzky et al. (1996) suggested that the metals in the ICM were mainly produced by SN II. Since stars in early-type galaxies in galaxy clusters are very old (e.g. Stanford et al. 1995; Kodama et al. 1998), this means that most of the metals in the ICM are produced through star formations at high $z$. Fukazawa et al. (1998) systematically studied 40 nearby clusters and found that the Si/Fe ratio is lower among the low-temperature clusters, which indicates that SNe Ia products are also important among these clusters. From the observed radial dependence of the abundances, Finoguenov et al. (2000) found that the SNe II ejecta have been widely distributed in the ICM.
In the ICM, around a cD galaxy, the contribution of metals from the galaxy becomes important. Fukazawa et al. (2000) found a central increment of Fe and Si abundances around cD galaxies, which is due to SNe Ia from the cD galaxies. The supply of metals by stellar mass loss must be also considered (e.g. Matsushita et al. 2000)
In addition to the Si and Fe abundances, the XMM-Newton observatory enables us to obtain $\alpha$-element abundances such as for O and Mg, which are not synthesized by SN Ia. Böhringer et al. (2001) and Finoguenov et al. (2002) analyzed annular spectra of M 87 observed by XMM-Newton, and found a flatter abundance gradient of O compared to steep gradients of Si, S, Ar, Ca, and Fe. The stronger abundance increase of Fe compared to that of O indicates an enhanced SN Ia contribution in the central regions (in disagreement with our findings, Gastaldello and Molendi (2002) claim a similar gradient for O and Fe, which would not support this conclusion). The observed similar abundance gradients of Fe and Si and the large Si/O ratio at the center imply a significant contribution of Si by SN Ia that is a larger Si/Fe abundance ratio by SN Ia than obtained by the classical model of Nomoto et al. (1984). The larger Si/Fe ratio and the implication that the Si/Fe ratio supplied by SN Ia may change with radius in the M 87 halo may indicate a diversity of SN Ia explosions also reflected in a diversity of the light curves as discussed by Finoguenov et al. (2002). A similar abundance pattern of O, Si and Fe is observed around center of A 496 (Tamura et al. 2001).
A prerequisite of the abundance determination is a precise knowledge of the temperature structure of the ICM. Based on the XMM-Newton observation of M 87, Matsushita et al. (2002; hereafter Paper I) found that the intracluster medium has a single phase structure locally, except for the regions associated with radio jets and lobes (Böhringer et al. 1995; Belsole et al. 2001), where there is an additional $\sim$ 1 keV temperature component. The signature of gas cooling below 0.8 keV to zero temperature is not observed as expected for a cooling flow (e.g Fabian et al. 1984). The fact that the thermal structure of the intracluster medium is fairly simple and the plasma is almost locally isothermal facilitates the abundance determination enormously and helps to make the spectral modeling almost unique.
In this paper, based on the detailed study of the temperature structure, abundances of various elements are discussed. In section 2, we summarize the observation and data preperation. Sections 3 and 4 deal with the MEKAL (Mewe et al. 1995, 1996; Kaastra 1992; Liedahl et al. 1994) and the APEC model (Smith et al. 2001) fits to the deprojected spectra, and in section 5, abundance ratios are determined directly from the line ratios, considering their temperature dependences. In section 6, we evaluate an effect of resonant line scattering. Section 7 gives a discussion of the obtained results, and Section 8 summarizes the paper. We adopt for the solar abundances the values given by Feldman (1992), where the solar Ar and Fe abundances relative to H are 4.47$\times10^{-6}$ and 3.24$\times10^{-5}$ in number, respectively. These values are different from the “photospheric” values of Anders & Grevesse (1989), where the solar Ar and Fe abundances are 3.63$\times10^{-6}$ and 4.68$\times10^{-5}$, respectively. The solar abundances of the other elements are consistent with each other. Unless otherwise specified, we use 90% confidence error regions.
Observation
===========
M 87 was observed with XMM-Newton on June 19th, 2000. The effective exposures of the EPN and the EMOS are 30ks and 40ks, respectively. The details of the analysis of background subtraction, vignetting correction and deprojection technique are described in Paper I. When accumulating spectra, we used a spatial filter, excluding those regions where the brightness is larger by 15 % than the azimuthally averaged value in order to excise the soft emission around the radio structures (Figure 1 and 6 in Paper I), although it is not fully excluded within $2'$, due to its complicated structure and the limited spatial resolution of the XMM telescope.
For the EPN spectra, we employ the response matrix corresponding to the average distance from the readout for each accumulated region, epn\_fs20\_sY$i$\_thin.rmf from March 2001. Here, $i$ reflects the distance from the readout-node, which affects the energy resolution. For the EMOS data, we used m1\_thin1v9q19t5r5\_all\_15.rsp from June 2001. The spectral analysis uses the XSPEC\_v11.1 package. We fitted the EPN and EMOS data in the spectral range of 0.5 to 10 keV. For outside 8$'$, the energy range between 7.5 and 8.5 keV of the EPN spectra is ignored, because of strong emission lines induced by particle events (Freyberg et al. 2002).
Spectral fit with the MEKAL model
=================================
We have fitted the deprojected spectra with a single temperature MEKAL model with photoelectric absorption, in the same way as in Paper I. For the outermost region, we fitted the projected (annular) spectrum. For the spectra within 1$'$, a power-law component is added for the central active galactic nucleus. We fixed index of the power-law component to the best-fit value obtained from the spectrum within 0.12$'$ (Paper I) and normalized it using the point spread function of the XMM telescope. Abundances of C and N are fixed to be 1 solar and those of other elements are determined separately. Within 2$'$, we also fitted the spectra with a two temperature MEKAL model, where the abundances of each element of the two components are assumed to have the same values, since within this radius, there remains a small amount of the cooler component with a nearly constant temperature of 1 keV associated with the radio lobes (Belsole et al. 2001; Paper I). As discussed in Paper I, the two temperature model should reflect the actual temperature structure of the ICM, although within 0.5$'$, the XMM spatial resolution is not enough to resolve temperature components of the complicated structure of the ICM.
[rlllllllllr]{} $R$ & kT1 & kT2 &$N_{\rm{H}}$ & O & Si & S & Ar & Ca & Fe & $\chi^2/\mu^a$\
(arcmin) & (keV) & (keV) & & (solar) & (solar) & (solar) &(solar)& (solar)& (solar) &\
\
0.00- 0.35 &1.13& &4.1& 0.02& 0.37& 0.37& 1.01& 1.25& 0.20& 435/177\
0.35- 1.00 &1.47& &4.3& 0.00& 0.83& 0.83& 0.69& 1.84& 0.79& 563/177\
1.00- 2.00 &1.56$^{+0.03}_{-0.03}$& &3.2$^{+1.1}_{-1.1}$& 0.42$^{+0.14}_{-0.14}$& 1.29$^{+0.15}_{-0.14
}$& 1.09$^{+0.17}_{-0.15}$& 0.96$^{+0.37}_{-0.36}$& 2.66$^{+0.88}_{-0.85}$& 1.05$^{+0.08}_{-0.07}$& 177/175\
2.00- 4.00 &2.05$^{+0.04}_{-0.04}$& &1.3$^{+0.6}_{-0.6}$& 0.44$^{+0.09}_{-0.09}$& 1.24$^{+0.10}_{-0.10
}$& 1.12$^{+0.11}_{-0.10}$& 0.71$^{+0.20}_{-0.20}$& 1.33$^{+0.39}_{-0.39}$& 1.07$^{+0.06}_{-0.06}$& 172/175\
4.00- 8.00 &2.30$^{+0.04}_{-0.04}$& &1.4$^{+0.4}_{-0.4}$& 0.45$^{+0.07}_{-0.07}$& 0.92$^{+0.07}_{-0.07
}$& 0.71$^{+0.08}_{-0.07}$& 0.42$^{+0.14}_{-0.16}$& 1.01$^{+0.28}_{-0.27}$& 0.79$^{+0.04}_{-0.04}$& 199/175\
8.00-11.30 &2.51$^{+0.08}_{-0.08}$& &1.9$^{+0.7}_{-0.7}$& 0.37$^{+0.12}_{-0.11}$& 0.64$^{+0.11}_{-0.10
}$& 0.35$^{+0.12}_{-0.11}$& 0.37$^{+0.23}_{-0.24}$& 0.72$^{+0.42}_{-0.43}$& 0.62$^{+0.06}_{-0.06}$& 243/175\
11.30-13.50 &2.72$^{+0.06}_{-0.05}$& &0.6$^{+0.4}_{-0.4}$& 0.32$^{+0.07}_{-0.07}$& 0.61$^{+0.07}_{-0.06
}$& 0.20$^{+0.07}_{-0.07}$& 0.14$^{+0.15}_{-0.14}$& 0.49$^{+0.24}_{-0.27}$& 0.51$^{+0.03}_{-0.03}$& 275/175\
\
0.00- 0.35 &1.56$^{+0.19}_{-0.16}$& 0.75$^{+0.02}_{-0.02}$ &2.3$^{+1.5}_{-1.5}$& 0.24$^{+0.30}_{-0.23}$& 1.04$^{+0.62}_{-0.3
8}$& 1.07$^{+0.63}_{-0.43}$& 0.62$^{+1.07}_{-0.62}$& 2.03$^{+2.46}_{-2.03}$& 0.85$^{+0.39}_{-0.24}$& 138/176\
0.35- 1.00 &1.73$^{+0.04}_{-0.04}$& 0.91$^{+0.03}_{-0.03}$ &2.1$^{+0.8}_{-0.4}$& 0.65$^{+0.13}_{-0.12}$& 1.70$^{+0.17}_{-0.1
6}$& 1.53$^{+0.17}_{-0.16}$& 0.99$^{+0.28}_{-0.28}$& 1.38$^{+0.58}_{-0.57}$& 1.54$^{+0.12}_{-0.11}$& 227/176\
1.00- 2.00 &1.68$^{+0.06}_{-0.06}$& 0.97$^{+0.15}_{-0.15}$ &2.6$^{+1.1}_{-1.2}$& 0.55$^{+0.19}_{-0.16}$& 1.56$^{+0.26}_{-0.1
9}$& 1.25$^{+0.23}_{-0.18}$& 0.96$^{+0.41}_{-0.39}$& 2.38$^{+0.92}_{-0.87}$& 1.28$^{+0.09}_{-0.12}$& 162/176\
\
0.00- 0.35 &0.74& &9.0& 0.19& 0.32& 0.71& 3.15& 0.40& 0.08& 255/132\
0.35- 1.00 &1.44& &4.6& 0.28& 0.86& 0.86& 0.79& 2.06& 0.81& 296/132\
1.00- 2.00 &1.56$^{+0.05}_{-0.04}$& &3.5$^{+1.1}_{-0.5}$& 0.45$^{+0.15}_{-0.14}$& 1.09$^{+0.17}_{-0.16}$& 0.99$^{+0.22}_{-0.20}$& 1.30$^{+0.56}_{-0.54}$& 1.70$^{+1.28}_{-1.25}$& 0.98$^{+0.09}_{-0.08}$& 112/125\
2.00- 4.00 &1.98$^{+0.06}_{-0.06}$& &1.5$^{+0.7}_{-0.7}$& 0.40$^{+0.11}_{-0.10}$& 1.05$^{+0.14}_{-0.13}$& 0.95$^{+0.16}_{-0.15}$& 0.81$^{+0.35}_{-0.35}$& 1.63$^{+0.69}_{-0.68}$& 0.96$^{+0.08}_{-0.07}$& 125/125\
4.00- 8.00 &2.28$^{+0.06}_{-0.06}$& &1.4$^{+0.5}_{-0.6}$& 0.28$^{+0.09}_{-0.08}$& 0.71$^{+0.10}_{-0.10}$& 0.53$^{+0.13}_{-0.12}$& 0.21$^{+0.27}_{-0.21}$& 1.01$^{+0.50}_{-0.50}$& 0.69$^{+0.05}_{-0.05}$& 93/125\
8.00-11.30 &2.48$^{+0.10}_{-0.11}$& &1.7$^{+0.8}_{-0.9}$& 0.29$^{+0.16}_{-0.14}$& 0.58$^{+0.19}_{-0.17}$& 0.28$^{+0.21}_{-0.23}$& 0.41$^{+0.46}_{-0.41}$& 0.95$^{+0.81}_{-0.84}$& 0.53$^{+0.09}_{-0.07}$& 85/125\
11.30-16.0 &2.54$^{+0.07}_{-0.07}$& &0.5$^{+0.5}_{-0.4}$& 0.26$^{+0.07}_{-0.07}$& 0.37$^{+0.09}_{-0.08
}$& 0.33$^{+0.12}_{-0.11}$& 0.20$^{+0.24}_{-0.20}$& 0.38$^{+0.39}_{-0.38}$& 0.41$^{+0.04}_{-0.04}$& 190/131\
\
0.00- 0.35 &1.73$^{+0.16}_{-0.18}$& 0.73$^{+0.02}_{-0.02}$ &0.0$^{+1.4}_{0.0}$& 0.48$^{+0.39}_{-0.27}$& 1.23$^{+0.57}_{-0.49}$& 1.02$^{+0.74}_{-0.61}$& 1.27$^{+1.77}_{-1.27}$& 1.26$^{+3.31}_{-1.26}$& 1.08$^{+0.32}_{-0.32}$& 133/123\
0.35- 1.00 &1.73$^{+0.05}_{-0.05}$& 0.87$^{+0.03}_{-0.03}$ &0.9$^{+1.0}_{-0.9}$& 0.58$^{+0.14}_{-0.13}$& 1.47$^{+0.26}_{-0.23}$& 1.56$^{+0.29}_{-0.25}$& 0.96$^{+0.48}_{-0.47}$& 1.65$^{+0.96}_{-0.93}$& 1.55$^{+0.18}_{-0.17}$& 152/123\
1.00- 2.00 &1.77$^{+0.09}_{-0.08}$& 1.00 (fix) &0.8$^{+1.4}_{-0.6}$& 0.69$^{+0.23}_{-0.20}$& 1.73$^{+0.41}_{-0.35}$& 1.47$^{+0.42}_{-0.36}$& 1.60$^{+0.78}_{-0.71}$& 1.38$^{+1.52}_{-1.38}$& 1.49$^{+0.26}_{-0.23}$& 90/123\
\
$^a$ Degrees of freedom
The results are summarized in Table 1 and Figures 1–4. Representative spectra fitted with the MEKAL model are shown in Figure 2 and 7 of Paper I and Figure \[mos\_apec\]. There remain small discrepancies between the data and the best fit model. A residual structure exists at 0.8–1 keV of the deprojected spectra of the EPN of $R>$2$'$, when fitted with the single temperature MEKAL model, while the EMOS data have no such residual (Figure 2 of Paper I). Here, $R$ is the 3-dimensional radius. This problem will be discussed in the next paragraph. The Fe-L/Mg-K structure around 1.3 keV is not fitted well (Figure \[mos\_apec\]). Therefore, we do not show Mg abundances here and will discuss this in more detail below (Sec. 4 and 5.5-5.6). We also do not show Ne and Ni abundances, because their line strengths are much smaller than those of the Fe-L features at the same wavelength and a small uncertainty in the Fe-L complex gives large uncertainties in abundances of these elements (Masai 1987; Matsushita et al. 1997; 2000). There are also small discrepancies in the continuum level around K$\alpha$ lines of Si, S and Ar, which will be discussed in Sec. 5.
The derived abundances are mostly consistent between the EPN and the EMOS, although the EPN abundances of O, Si, S and Fe are systematically smaller than the EMOS results by 10–30%, when fitted with a single temperature MEKAL model. As will be shown in Sec. 5, the strengths of K$\alpha$ lines of H-like O, Si, S and He-like (+ Li like) Fe observed by the EPN and EMOS agree within several percent when considering the difference in the normalization between the two detectors. Therefore, these abundance discrepancies should be caused by the discrepancy at 0.8–1 keV, which may reflect uncertainties in an instrumental low energy tail of the strong Fe-L lines observed in EPN spectra, as discussed in Paper I. In addition, the EPN is characterized by a dependence of the energy resolution on the position on the detector. As a result, the observed Fe abundance of the EPN may be artificially decreased by these problems and the change of the continuum level also affects the O abundance. Those of Si and S also couple with abundances of other elements such as O and Fe due to the effect of free-bound emission and the increment of the strength of the Fe-L emission around the K$\alpha$ line of He-like Si. When we fix the O, Ne, Mg, Fe and Ni abundances to the best fit values derived from the EMOS spectra of the whole energy band and fitted the EPN spectra above 1.7 keV, we recover consistent Si and S abundances between the two detectors (Figure \[depro\_abund\]).
The derived abundances of Si, S, Ar, Ca, and Fe from deprojected spectra using the single temperature MEKAL model fit show strong negative gradients at $R>1'$, and drop sharply within this radius (Table 1, Figure \[depro\_abund\]) as already seen in the projected spectra (Böhringer et al. 2001; Finoguenov et al. 2001; Gastaldello & Molendi 2002). But now the absolute values are slightly larger than those from projected spectra. In contrast, the O abundance is nearly constant at $R>1'$.
The two temperature MEKAL fit on the deprojected spectra gives significantly larger abundances than the single temperature MEKAL fit (Figure \[depro\_abund\]-\[abund\_cont\_o\], Table 1) and the central abundance drops seen in the single MEKAL fit become very week in the results of the two temperature MEKAL fit, although contributions of the cooler component are small as shown in Paper I. For example, at $R$=0.5-1’, 10 percent of the cooler component in units of emission measure changes the Fe abundances by a factor of 2. The reason for this is the change of the temperature of the hot component, when the cold component is added. For a fixed Fe line feature this leads to a higher abundance required in the fit. In addition, abundances of O, Si and S also increase by a factor of 2. The increment of the O abundance is due to the change of the temperature structure. The Si and S abundances increase due to an increment of the free-bound emission and the strength of Fe-L lines around the He-like Si line. Therefore, considering the remaining 1 keV temperature component associated with the radio structures within 2$'$, abundances become nearly constant within 2$'$.
We have also tried a three temperature MEKAL model for spectra within 2$'$. Although the derived $\chi^2$ values have slightly improved, the derived abundances and their errors have nearly the same values as those from the two temperature MEKAL model fit.
Figure \[abund\_cont\_fesi\] shows the confidence contours for the Si, S and Fe abundances, derived from the deprojected spectra. For the outermost region, the results from the projected spectra are plotted. The elliptical shape of the confidence contours indicates that abundance ratios are better determined than the abundances themselves. The differences in the Si/Fe and S/Fe ratios between the EMOS and EPN are considerably smaller than the abundance differences. Furthermore, the two temperature MEKAL model fit gives almost the same values of the Si/Fe and S/Fe ratio with the single temperature MEKAL model fit, although absolute values of Si, S, and Fe are largely changed. This means that the abundance ratios do not strongly depend on the temperature structure. As a result, the Si/Fe ratio is determined to be nearly constant at 1.1$\pm$ 0.1 solar, with no radial gradient. In contrast, we find that the S abundance has a steeper gradient than Fe abundance.
The abundances of Si and Fe are nearly independent to the O abundance (Figure \[abund\_cont\_o\]). However, at the center, the two temperature MEKAL model gives the similar values of the O/Si and O/Fe ratio as the single temperature MEKAL model. The O/Si and O/Fe ratios are consistent between the two detectors, although the abundances of O, Si, and Fe between the EMOS and EPN differ by 10–30%. Both the Si/O and Fe/O ratios increase by a factor of 1.5 with radius. This result is in disagreement with the claim by Gastaldello & Molendi (2002), where they derived the constant Fe/O ratio. The background subtraction should not be a problem when determining O, Fe, and Si abundances, since even at the radius of 10$'$, the background is only a few percent between 0.5 to 2 keV. In addition, the background was subtracted according to the detailed study by Katayama et al. (2002). A more severe problem should be the uncertainties in the response matrix, since the O abundance was inconsistent between the EPN and the EMOS by a factor of 2 using an old response matrix (Böhringer et al. 2001), although the discrepancy becomes smaller now as shown in this section. Using an old version of the response matrix and fitting the projected spectra in the same way in Gastaldello & Molendi (2002), we achieve the same results.
In summary, from the deprojected spectra using the MEKAL model, the abundances of Si, S, Ar, Ca, and Fe have steep gradients at $R>2'$ and become nearly constant within the radius. The Si/Fe ratio is almost constant within the field of view, while the S/Fe ratio has a negative, and the O/Fe ratio has a positive gradient.
Spectral fits with the APEC model
=================================
A new plasma code, APEC (Smith et al. 2001), is now available in XSPEC\_v11. Within the energy resolution of the CCD detectors, the dominant change from version 1.0 to the APEC code in the XSPEC version 11.0.1 is not in the Fe-L lines, but in the K$\alpha$ lines. Using version 1.0 of the APEC code, strengths of the K$\alpha$ lines of H-like ions, which are the most fundamental lines, decrease by $\sim$ 50 % at 2 keV, and their temperature dependences also change. The strength of the 6.7 keV Fe-K line decreases, by a factor of 2 at 1 keV and 1.3 at 4 keV. The changes of the K$\alpha$ lines of He-like ions of Si, S, Ar and Ca are smaller. As a result, any single temperature APEC model cannot fit the deprojected spectra of M 87, especially for the ratio of the Fe-L and the Fe-K as shown in Paper I, although a single temperature MEKAL model can fit each deprojected spectrum at $R>2'$.
However, the strengths of K$\alpha$ lines of version 1.1 of the APEC code in version 11.1 of the XSPEC became more consistent with the MEKAL ones. Within the temperature range of the ICM of M 87, the largest differences except for the Fe-L structure are 20% changes in the strengths of the K$\alpha$ line of H-like O and the Fe line at 6.7 keV.
Using version 1.1 of the APEC code, we fitted the deprojected spectra of the EMOS. The results are summarized in Table 2 and Figure \[ratio\_apec\_mekal\]. In contrast to the worse $\chi^2$ values obtained from version 1.0 of the APEC code, version 1.1 gives similar $\chi^2$ values compared to those obtained by the MEKAL model. For the Fe-L/Mg-K structure around $\sim 1.4$ keV, the APEC (v1.1) model gives a better fit than the MEKAL model (Figure \[mos\_apec\]). The derived temperatures and hydrogen column densities are consistent with those from the MEKAL model. The contribution of the lower temperature component at $R$=1–2$'$ has slightly decreased from 4% by the MEKAL code to 2% by the APEC code, although the contribution within 1$'$ is consistent with each other. The few percent of the soft component using the MEKAL model at $R$=1–2$'$ may be an artifact due to the uncertainties in the Fe-L complex.
[rcccccccccc]{} $R$ & kT1 & kT2 & $N_{\rm{H}}$ & O & Mg &Si & S & Fe & $\chi^2/\mu^a$\
(arcmin) & (keV) & (keV)& &(solar) &(solar)& (solar)& (solar)& (solar) &\
\
0.00- 0.35 & 1.15 & & 2.5 & 0.06 & 0.0 & 0.29 & 0.41 &0.15 & 253/175\
0.35- 1.00 & 1.48 & & 2.4 & 0.97 & 1.17 & 1.80 & 1.79 &1.37 & 583/175\
1.00- 2.00 &1.59$^{+0.01}_{-0.03}$& &2.8$^{+1.2}_{-1.0}$& 0.56$^{+0.17}_{-0.08}$& 0.87$^{+0.17}_{-0.17}$& 1.47$^{+0.10}_{-0.08}$& 1.28$^{+0.15}_{-0.15}$& 1.15$^{+0.05}_{-0.05}$& 126/175\
2.00- 4.00 &2.01$^{+0.02}_{-0.01}$&&2.1$^{+0.3}_{-0.3}$& 0.54$^{+0.06}_{-0.09}$& 0.67$^{+0.12}_{-0.11}$& 1.19$^{+0.07}_{-0.06}$& 1.09$^{+0.08}_{-0.08}$& 1.03$^{+0.03}_{-0.02}$& 181/175\
4.00- 8.00 &2.28$^{+0.02}_{-0.03}$& &2.0$^{+0.2}_{-0.4}$& 0.52$^{+0.08}
_{-0.05}$& & 0.87$^{+0.06}_{-0.05}$& 0.68$^{+0.07}_{-0.06}$& 0.78$^{+0.02
}_{-0.02}$& 210/175\
8.00-11.30 &2.48$^{+0.03}_{-0.03}$&&2.2$^{+0.7}_{-0.7}$& 0.46$^{+0.1
5}_{-0.11}$& & 0.65$^{+0.08}_{-0.12}$& 0.36$^{+0.10}_{-0.14}$& 0.64$^{+0.
06}_{-0.05}$& 248/175\
11.30-13.50 &2.74$^{+0.05}_{-0.07}$& &0.8$^{+0.5}_{-0.5}$& 0.38$^{+0.13}
_{-0.08}$& & 0.62$^{+0.07}_{-0.06}$& 0.19$^{+0.09}_{-0.08}$& 0.53$^{+0.03
}_{-0.01}$& 305/175\
\
0.00- 0.35 &1.41$^{+0.27}_{-0.43}$& 0.78$^{+0.03}_{-0.02}$ &2.4$^{+1.1}_{-1.5}$& 0.26$^{+0.29}
_{-0.26}$& 0.75$^{+0.87}_{-0.31}$& 1.14$^{+0.96}_{-0.24}$& 1.18$^{+0.63}_{-0.42}$& 0.96$^{+0.10
}_{-0.10}$& 138/174\
0.35- 1.00 &1.78$^{+0.06}_{-0.05}$& 1.05$^{+0.01}_{-0.02}$ &2.0$^{+0.8}_{-0.6}$& 0.80$^{+0.13}
_{-0.14}$& 1.10$^{+0.14}_{-0.18}$& 1.78$^{+0.09}_{-0.16}$& 1.61$^{+0.11}_{-0.16}$& 1.58$^{+0.06
}_{-0.14}$& 190/174\
1.00- 2.00 &1.63$^{+0.01}_{-0.01}$& 1.00 (fix) &2.5$^{+0.6}_{-1.2}$& 0.64$^{+0.2
3}_{-0.20}$& 1.00$^{+0.32}_{-0.27}$& 1.61$^{+0.27}_{-0.23}$& 1.38$^{+0.26}_{-0.22}$& 1.26$^{+0.
17}_{-0.15}$& 125/174\
\
$^a$ Degrees of freedom
Figure \[ratio\_apec\_mekal\] compares the abundances derived from the APEC and the MEKAL model. The large abundance changes seen in APEC version 1.0 (Paper I) are not derived using the new APEC code, version 1.1. For Si, S, and Fe, we find consistent results within several percent. The O abundances from the APEC code are systematically larger than those from the MEKAL code by 20%, which is close to the difference of the line strength of the K$\alpha$ line of H-like O between the MEKAL and APEC model.
In contrast, the derived Mg abundances differ by a factor of 2 between the two codes. Since the APEC model can fit the Mg-K/Fe-L structure at $\sim$1.4 keV quite well, the APEC code may be better to derive a Mg abundance. There is another problem due to the strong instrumental line of the EMOS, which is located at a similar energy to the Mg line. At $r$=8–13.5$'$, the strength of the instrumental line is a factor of 10 larger than the Mg line of the ICM, and the uncertainty affects the spectrum at $R$=4–8$'$ through the deprojection. Here, $r$ is the projected radius from the center. Therefore, the Mg abundances are shown only within 4$'$. The derived Mg abundance is $\sim 1$ solar at $R<2'$ and shows a radial gradient outside this radius (Table 2).
The observed values of the Ne abundance derived from the APEC model are also a factor of 1.3 larger than those derived from the MEKAL model. For the Ni abundance, the APEC code gives 60% of the values from the MEKAL code at $R>2'$, but within $R<2'$, both codes give similar values. In contrast to the Mg lines, the Ni-L lines and the Ne-K lines cannot be distinguished in the spectra of M 87. Therefore, we do not present Ni and Ne abundances in this paper.
In summary, the new APEC code provides a better fit to the Fe-L structure, and the derived abundances of O, Si, and Fe are mostly consistent with those obtained from the MEKAL code. The Mg abundances are derived to be $\sim$ 1 solar at $R<2'$, since the Fe-L/Mg-K structure is well fitted with the APEC code.
Abundance determination using line ratios
=========================================
In this section, we try to obtain abundance ratios of elements directly from ratios of emission lines. When calculating the deprojected spectra, we make an assumption that spectra beyond the field of view are the same as the outermost spectrum, which is not valid since the temperature starts to decrease beyond the field of view of the detector (Shibata et al. 2001). Therefore, especially in the outer regions, there may be a bias due to this assumption, since emission lines have a stronger dependence on temperature than the continuum. In addition, abundances from deprojected spectra have a larger uncertainty due to poorer statistics than those derived from projected data. In order to derive abundances more accurately, we need projected spectra.
However, the projected spectra consist of multi-temperature components. As a result, some line ratios, such as the ratio of K$\alpha$ lines of He- and H-like S, of the projected spectra cannot be fitted with a single temperature MEKAL model (Figure \[mekalspec\]). We must be careful in using a multi-temperature model, e.g. a two temperature model, since emission lines and a continuum spectrum are different functions of temperature. In addition, the continuum spectra between S, Ar and Ca lines show small discrepancies between the data and the single temperature model. These small discrepancies sometimes give larger uncertainties in the strengths of faint lines such as K$\alpha$ lines of He-like Ar and Ca. For example, when fitting with a single temperature MEKAL model with the projected spectrum at 2–4$'$, temperature and Ar abundance are determined as 2.2 keV and 0.7 solar, respectively. However, when we restrict the energy range within 2.8–5.0 keV and fix the abundances of O, Ne, Mg, Si, S and Fe to the best fit values obtained from the whole energy band fit, the temperature becomes 2.4 keV and Ar abundance is determined as 1.1 solar. Here, the Ar abundance is changed by 60%, although the best fit temperatures differ by only 10%. This small discrepancy between the S and Ar lines is also seen in the deprojected spectra, although their errors are larger.
[ccccccccccccc]{} r & H-O/H-Si & O/Si (solar) & (H-S+He-S)/H-Si & S/Si (solar) &He-Ar/H-Si & Ar/Si (solar) & He-Ca/H-S & Ca/S (solar)\
\
0.0- 0.5 & $2.26\pm0.29$ & $0.43\pm0.06$ & $1.08\pm0.06$ & $1.14\pm0.06$ & $0.16\pm0.03$ & — & $0.20\pm0.04$ & —\
0.5- 1.0 & $2.16\pm0.15$ & $0.48\pm0.03$ & $1.02\pm0.04$ & $1.05\pm0.05$ & $0.14\pm0.02$ & $0.79\pm0.11$ & $0.23\pm0.03$ & $1.29\pm0.16$\
1.0- 2.0 & $1.90\pm0.11$ & $0.45\pm0.03$ & $0.94\pm0.03$ & $0.96\pm0.03$ & $0.18\pm0.03$ & $0.87\pm0.12$ & $0.28\pm0.03$ & $1.53\pm0.15$\
2.0- 4.0 & $1.84\pm0.13$ & $0.46\pm0.03$ & $0.93\pm0.03$ & $0.97\pm0.03$ & $0.17\pm0.01$ & $0.78\pm0.06$ & $0.23\pm0.02$ & $1.16\pm0.10$\
4.0- 8.0 & $2.38\pm0.10$ & $0.59\pm0.02$ & $0.82\pm0.03$ & $0.87\pm0.03$ & $0.14\pm0.02$ & $0.67\pm0.09$ & $0.24\pm0.02$ & $1.18\pm0.11$\
8.0-13.5 & $2.45\pm0.20$ & $0.61\pm0.05$ & $0.62\pm0.05$ & $0.67\pm0.05$ & $0.14\pm0.03$ & $0.71\pm0.16$ & $0.32\pm0.06$ & $1.56\pm0.29$\
\
0.0- 0.5 & $2.08\pm0.42$ & $0.35\pm0.07$ & $1.09\pm0.17$ & $1.16\pm0.18$ & $0.18\pm0.06$ & — & $0.18\pm0.08$ & —\
0.5- 2.0 & $2.06\pm0.23$ & $0.45\pm0.05$ & $0.96\pm0.04$ & $1.00\pm0.04$ & $0.16\pm0.02$ & $0.93\pm0.11$ & $0.29\pm0.04$ & $1.60\pm0.21$\
2.0- 5.6 & $1.77\pm0.12$ & $0.44\pm0.03$ & $0.90\pm0.04$ & $0.93\pm0.04$ & $0.18\pm0.02$ & $0.85\pm0.10$ & $0.24\pm0.03$ & $1.24\pm0.16$\
5.6-11.3 & $2.74\pm0.31$ & $0.68\pm0.08$ & $0.75\pm0.07$ & $0.80\pm0.08$ & $0.14\pm0.03$ & $0.67\pm0.15$ & $0.25\pm0.05$ & $1.26\pm0.26$\
\
0.0- 0.5 & $2.00\pm0.33$ & $0.38\pm0.06$ & $1.11\pm0.11$ & $1.17\pm0.11 $& $0.30\pm0.05$ & — & $0.24\pm0.09$ & —\
0.5- 1.0 & $1.87\pm0.20$ & $0.41\pm0.04$ & $1.04\pm0.06$ & $1.07\pm0.06 $& $0.15\pm0.03$ & $0.81\pm0.13$ & $0.24\pm0.06$ & $1.36\pm0.31$\
1.0- 2.0 & $1.90\pm0.18$ & $0.44\pm0.07$ & $0.92\pm0.05$ & $0.94\pm0.05 $& $0.18\pm0.02$ & $0.88\pm0.10$ & $0.28\pm0.04$ & $1.50\pm0.22$\
2.0- 4.0 & $1.87\pm0.23$ & $0.47\pm0.06$ & $0.93\pm0.05$ & $0.97\pm0.05 $& $0.17\pm0.02$ & $0.81\pm0.11$ & $0.35\pm0.05$ & $1.78\pm0.25$\
4.0- 8.0 & $2.09\pm0.26$ & $0.52\pm0.06$ & $0.76\pm0.05$ & $0.82\pm0.05 $& $0.08\pm0.02$ & $0.41\pm0.09$ & $0.34\pm0.09$ & $1.69\pm0.44$\
8.0-13.5 & $2.70\pm0.43$ & $0.67\pm0.11$ & $0.67\pm0.08$ & $0.73\pm0.09 $& $0.20\pm0.05$ & $1.02\pm0.23$ & $0.44\pm0.08$ & $2.14\pm0.39$\
Therefore, in this section, we determine strengths of emission lines, and then determine abundance ratios, considering the temperature dependence of line ratios. We have fitted the projected and deprojected spectra within an energy band around a given line with thermal bremsstrahlung and gaussians. Obtaining strengths of K$\alpha$ lines of He-like S and Ar, contributions of K$\beta$ lines of H-like Si and S were subtracted using the strengths of K$\alpha$ lines. Here, we used the K$\beta$ to K$\alpha$ ratio of the MEKAL code. Strengths of H-like K$\alpha$ lines of O were obtained from spectra within the energy range of 0.55–1 keV fitted with a gaussian and two APEC models with zero O abundance. Those of Si were also obtained from fitting with gaussians and the two temperature APEC model with zero Si abundance, using the 1.1–2.2 keV energy band. We selected line ratios whose temperature dependence is small, for accurate measurements of abundance ratios. The results are summarized in Table 3.
O vs. Si
--------
The abundance ratio of O/Si is better obtained from the ratio of the line strengths of K$\alpha$ lines of H-like ions. Although the O and Si abundances obtained from the EMOS and the EPN using the same single temperature model have small discrepancies of 20$\sim$ 30%, the line ratios of the two detectors agree well with each other (Figure \[osi\]).
The observed radial profile of the line ratio has a minimum at $\sim$2$'$ (Figure \[osi\]). Figure \[osi\] also shows the line ratios plotted against the best fit MEKAL temperatures from the whole energy band. When the abundance ratio of O/Si is constant, the line ratio is almost constant above 1.7 keV and it starts to increase sharply below the lowest temperature. Comparing the observed minimum value of $\sim$ 2 with the minimum value of $\sim4$ corresponding to the solar abundance ratio, no temperature distribution can reproduce an O/Si ratio larger than 0.5 solar at the radius of $\sim$2$'$.
Considering the temperature structure derived in Paper I, we can better constrain the O/Si ratio. The independence of the line ratio from the temperature above 1.7 keV means that the abundance ratio can be directly obtained dividing the line ratio by that of the solar abundance ratio when there is no temperature component below 1.7 keV. Thus, beyond $2'$, the gradient of the line ratio reflects a change of the O/Si ratio, since outside 2$'$, there is no temperature component below 1.7 keV (Paper I). From the line ratio at $r$= 2–4$'$, the O/Si is derived to be 0.46 solar, and beyond $r>8'$, it increases to 0.6–0.7 solar (Figure \[osi\]).
Within 2$'$, the presence of the additional $\sim$ 1 keV temperature component must be considered. The derived O/Si ratio assuming that the temperature is larger than 1.7 keV should be the maximum value of the O/Si ratio which is $\sim$ 0.5 solar within 2$'$ (triangles in the bottom panel of Figure \[osi\]). We also derive the O/Si ratio from the line ratio considering the fraction of the low temperature component, which is derived from the spectral fit using the best fit two temperature APEC model. Then, the O/Si ratio is derived to be 0.45 solar at 1–2$'$ and 0.4 solar within 0.5$'$. Since the line ratio is a steep function when the temperature is around 1 keV, we have also checked a case when the lower temperatures is 0.6 keV, which is 0.2 keV smaller than the central temperature. For the innermost region, the derived O/Si ratio with the temperature of 0.6 keV is consistent within 10 percent, since the temperature of the softer component becomes lower, the fraction of the component also becomes lower which is derived from the spectral fitting from Fe-L.
For comparison, we also plotted the O/Si ratio derived from the deprojected spectra through the spectral fits and the line ratio. Although the O/Si ratios derived from the line ratios are systematically larger by $\sim$ 20% than those derived from spectral fit with the MEKAL model, the gradient of the O/Si ratio is consistent between the two, and the O/Si ratio changes by a factor of 1.4 from the center to 10$'$.
S vs. Si
--------
Although the ratios of K$\alpha$ lines of He-like or H-like S to that of H-like Si are both steep functions of temperature, the ratio of the sum of the two K$\alpha$ lines of S to the H-like Si line is nearly constant within 10% between 0.9 to 5 keV when the S/Si ratio is constant (Figure \[ssi\]). The important point is that we can derive the abundance ratio almost independently from the temperature structure of the ICM. Considering that there is no temperature component below 1 keV, as in the case of the O/Si ratio, even for the projected data, the S/Si ratio can be directly calculated from the line ratio. As in the previous subsection, the effect of the lower temperature component is estimated to be less than a few percent.
The observed profile of the line ratio shows a negative gradient and at $r$=10$'$, it is about half of the central value. This gradient reflects the change of the S/Si ratio, because of the small dependence on temperature. Converted to the abundance ratio, the S/Si ratio is 1.0 solar within 2$'$, and drops to 0.67 solar at $\sim$ 10$'$ (Figure \[ssi\]). These results are systematically larger than those derived from the spectral fits. Especially, at $r>8'$, the spectral fits on the projected spectra give the S/Si ratio to be 0.3–0.5 solar (Table 1, Figure 2), while the line ratio gives the value of 0.6–0.7 solar. For the projected data, the S abundance from the spectral fit should not be correct since the single temperature model cannot fit the two K$\alpha$ lines simultaneously as shown in Figure 6.
Ar vs. Si
---------
Although the line ratio of the K$\alpha$ lines of He-like Ar to H-like Si is a steeper function of temperature than S/Si or O/Si ratio, within 1.7 to 2.8 keV, at a given ratio of Ar/Si, the change of the line ratio is less than 10 % (Figure \[arsi\]). In Paper I, we found that at $R>0.5'$, the ICM temperature at a given radius is dominated by a single temperature component whose temperature is larger than 1.7 keV. From Figure 7 in Paper I, the contributions from the 1 keV component to Ar and H-like Si lines are negligible. Therefore, as in case of the S to Si abundance ratio, we can obtain the abundance ratio of Ar/Si from the line ratio. The results indicate that the Ar/Si is 0.7–0.8 solar.
The derived Ar/Si ratios are systematically larger than the results of the spectral fit on the projected data (Finoguenov et al. 2002) by 30% within 3$'$ and by a factor of 2 at $r>8'$. This large discrepancy is caused by the failure to fit the continuum between the S and Ar lines illustrated in Figure 6.
Ca vs. S
---------
The ratio of the K$\alpha$ line of the He-like Ca to that of the H-like S also show a small temperature dependency (Figure \[cas\]). Because these lines are dominated by a hotter temperature component even around the center, we can calculate the Ca/S ratio in the same way. The Ca/S ratio is $\sim$ 1.5 solar within the whole region.
Mg vs. O
--------
The energy of the K$\alpha$ line of H-like Mg is slightly shifted from the peak of the Fe-L emission at $\sim$ 1.49 keV (Figure \[mgkspec\]). Thus, the K$\alpha$ line can be distinguished within the energy resolution of the CCDs. The problem with the spectral fitting is the modeling of the Fe-L structure around the Mg line. Due to the problem of the strong instrumental line of the EMOS, we do not present the results on the Mg-K line for $r>8'$.
The strengths of H-like and He-like K$\alpha$ lines of Mg were obtained from spectra within 1.1–1.6 keV, fitted with gaussians and a two temperature MEKAL or APEC model with zero Mg abundance. The results are summarized in Table 4.
Figure \[mgo\] shows the ratios of the K$\alpha$ line of the H-like Mg to that of H-like O. The strengths of the Mg line obtained using the APEC Fe-L model are $\sim$ 30% larger than those derived using the MEKAL model. Since the APEC fit to the Fe-L/Mg-K feature of 1.4 keV is better than the MEKAL model fit, the APEC code may be better to describe the Fe-L feature, although the discrepancy between the data and model of the MEKAL plus gaussians is only several percent.
[cccccccc]{} r & H-Mg/H-O & Mg/O [(solar)]{} & H-Mg/H-Si & Mg/Si [(solar)]{}\
\
0.0- 0.5 & $0.23\pm0.04$ & $1.46\pm0.23$ & $0.53\pm0.05$ & $0.63\pm0.07$\
0.5- 1.0 & $0.22\pm0.02$ & $1.33\pm0.11$ & $0.48\pm0.03$ & $0.64\pm0.04$\
1.0- 2.0 & $0.21\pm0.02$ & $1.26\pm0.14$ & $0.40\pm0.04$ & $0.57\pm0.06$\
2.0- 4.0 & $0.22\pm0.02$ & $1.30\pm0.14$ & $0.40\pm0.03$ & $0.59\pm0.05$\
4.0- 8.0 & $0.19\pm0.02$ & $1.16\pm0.10$ & $0.46\pm0.04$ & $0.69\pm0.06$\
\
0.0- 0.5 & $0.27\pm0.07$ & $1.73\pm0.48$ & $0.56\pm0.13$ & $0.61\pm0.14$\
0.5- 2.0 & $0.20\pm0.03$ & $1.19\pm0.17$ & $0.41\pm0.04$ & $0.54\pm0.05$\
2.0- 4.0 & $0.20\pm0.05$ & $1.21\pm0.28$ & $0.42\pm0.08$ & $0.66\pm0.10$\
\
0.0- 0.5 & $0.31\pm0.06$ & $1.90\pm0.36$ & $0.61\pm0.08$ & $0.72\pm0.10$\
0.5- 1.0 & $0.22\pm0.04$ & $1.31\pm0.22$ & $0.41\pm0.06$ & $0.54\pm0.07$\
1.0- 2.0 & $0.19\pm0.03$ & $1.14\pm0.20$ & $0.36\pm0.06$ & $0.52\pm0.08$\
2.0- 4.0 & $0.23\pm0.05$ & $1.41\pm0.31$ & $0.44\pm0.08$ & $0.65\pm0.13$\
4.0- 8.0 & $0.24\pm0.06$ & $1.45\pm0.38$ & $0.50\pm0.12$ & $0.75\pm0.18$\
8.0-13.5 & $0.15\pm0.04$ & $0.94\pm0.26$ & $0.41\pm0.10$ & $0.63\pm0.15$\
\
The projected EMOS spectra in the energy range of 1.35 to 1.6 keV are also fitted with a bremsstrahlung model plus two gaussians at 1.47 keV and 1.49 keV, since the K$\alpha$ line of H-like Mg at 1.47 keV and a Fe-L peak at 1.49 can be distinguished within the EMOS energy resolution (Figure \[mgkspec\]). Freeing the strength of 1.49 keV peak of the Fe-L enable us to constrain the Mg line strength. The derived strengths of the Mg line are consistent with those derived using the APEC code for the modeling of the Fe-L. This result supports the Mg line strengths from the APEC code than those from the MEKAL code.
As in Sec. 5.1-5.4, we have plotted the line ratio from the APEC code against the best fit MEKAL temperature (Figure \[mgo\]). The constant Mg/O ratio gives a constant line ratio within 10% above 1 keV. As in Sec. 5.1, we have estimated the effect of the temperature component below 1 keV, which is less than 10% for the innermost region. The derived Mg/O ratios are about 1.2–1.3 solar with no radial gradient at $r>0.5'$, when we adopt the Fe-L modeling for the APEC (Figure 13). Within 0.5$'$, a higher Mg/O ratio is allowed, since adding a temperature component blow 1 keV, the Mg/O ratio increases.
The O and Mg abundances obtained from the RGS spectrum through a spectral fit are 0.49$\pm0.04$ solar and 0.9$\pm0.2$ solar, respectively (Sakelliou et al 2002). This Mg/O ratio derived from the RGS is a factor of 1.5 larger than that at $r>0.5'$ derived from the EMOS spectra using the APEC code for modeling the Fe-L. Since most of the photons detected by the RGS come from the very center, where a larger Mg/O ratio is also allowed by the EMOS data, the Mg/O ratio may increase at the center.
Mg vs. Si
---------
The derived ratios of the K$\alpha$ lines of H-like Mg and Si are summarized in Table 4 and Figure \[mgsi\]. Here, only the result using the APEC code for the Fe-L modeling is shown. The temperature dependence of the line ratio is quite similar to that of the K$\alpha$ lines of O and Si, and the Mg/Si ratio was derived in the same way as in Sec. 5.1. Within 2$'$, the contribution of the 1 keV component is taken into account. The derived Mg/Si ratio is $\sim 0.6$ solar, which agrees well with the ratio derived from the spectral fit using the APEC model in Sec 4.
Fe vs. Si
---------
In contrast to the temperature dependences of the line ratios in the previous subsections, the ratio of the Fe-K line at 6.7 keV (which includes He and Li like ions) to the K$\alpha$ line of the H-like Si line strongly depends on temperature. For example, at 2 keV, a 10 % increment of temperature increases the Fe/Si ratio by 36 %. As a result, small uncertainties in the temperature structure give large uncertainties in the abundance ratios. Therefore, the Si/Fe ratio cannot obtained from the line ratio as in previous subsections. Since this ratio is quite important, we have calculated the Si/Fe line ratio of the projected data using the temperature structure obtained in Paper I and compared it with the observed profile (Figure \[fesiratio\]). We used the relation of $kT=1.68 (1+(R/1.6')^2)^{0.115}$ keV, since the hotter component dominates these lines even within 2$'$. Assuming the Fe/Si ratio is approximated by $a+bR$, where a and b are free parameters, we fitted the profile of the line ratio. The radial profile of the line ratio is well fitted with the model ($\chi^2$=7.67 for 9 degrees of freedom) and the Fe/Si ratio is determined to be $\sim$ 0.9 solar within the whole field of view (Figure \[fesiratio\]), although the abundance ratio changes by 20% when the whole temperature is shifted by 5%. Considering that the value of 0.9 solar is quite close to the ratio derived from the spectral fitting of the deprojected data, where the Fe abundance is determined by the Fe-L emission, the Fe/Si ratio should be $\sim$ 0.9 solar and its radial gradient is less than 10–20%.
Effect of resonance line scattering
===================================
Some resonant lines may become optically thick in the dense cores of clusters. Shigeyama (1998) calculated the effect on M 87 and several X-ray luminous elliptical galaxies, and found that it should be important at the center. Böhringer et al. (2001) suggested that the observed central abundance drop may be due to the effect. In the case of an X-ray luminous elliptical, NGC 4636, Xu et al. (2002) discovered direct evidence of resonant scattering using a line ratio between optically thick and thin lines.
The profile of optical depth for resonant lines of some prominent emission lines are calculated in Böhringer et al. (2001). Using the observed abundance profiles (Table 1), the temperature profile used in Sec. 5.7 and the density profile derived in Paper I, optical depths from the center for K$\alpha$ lines of H-like Si and O are calculated to be 1 and 0.4, respectively. Therefore, we must evaluate the effect of the scattering, although the central abundance drop is very small when fitted with the two temperature MEKAL model.
Ignoring turbulent motion, we have calculated the effect of resonant line scattering on the K$\alpha$ lines of H-like ions, using Monte-Carlo simulations.
The results are summarized in Figure \[res\]. The scattered profiles of lines are consistent with those ignoring scattering beyond 1$'$. Within this radius, the brightness of the lines decrease. The reduction rates at 0.5$'$ are 10, 30, 15% for O, Si, and S, respectively. The effect is important only within 0.5$'$ on the Si line, where the uncertainty in temperature structure also gives an uncertainty in abundance determination.
Discussion
==========
Summary of abundances
---------------------
The abundance profiles of O, Si, S, Ar, Ca and Fe are obtained from the deprojected spectra, based on the temperature study in Paper I. Although version 1.0 of the APEC code gives significantly different results (Paper I), version 1.1 of the APEC code gives consistent results with those of the MEKAL code for these elements. Within 2$'$, considering the two temperature nature of the ICM, the abundance profiles from deprojected spectra become almost constant. Within this radius, the Si, S, Ar, Ca and Fe abundances are $\sim$ 1.5 solar, while the O abundance is a factor of 2 smaller. Using the APEC code, the Mg abundance is derived to be $\sim$ 1 solar at the center, since the APEC code gives a better fit on the Fe-L/Mg-K structure at 1.4 keV. Beyond 2$'$, Si, S, Ar, Ca and Fe show strong negative gradients, while the abundance gradient of O is smaller. The Si/Fe ratio is 1.1 solar, with no gradient in the field of view.
The projected and deprojected profiles of abundance ratios are also obtained through line ratios. Since some line ratios among $\alpha$-elements are nearly constant within the temperature range of the ICM around M 87, the abundance ratios are obtained with better accuracy. From the temperature dependence of the line ratio, we can conclude that the low O/Si ratio cannot be explained by any temperature model, and the ratios of O/Si are 0.4–0.5 solar at the center and 0.6–0.7 solar at the outer region. The Mg/O ratio is determined to be 1.2–1.3 solar. The Ar and S abundances thus obtained are systematically larger than the spectral fit results due to small disagreements in the continuum fits around these lines. The S/Si is 1.1 solar at the center and 0.7 solar at outer regions. The Si/Fe ratio derived from the projected profile of the K$\alpha$ line ratio is consistent with that derived from the deprojected spectra through the spectral fit, where the Fe abundance is determined by the Fe-L structure.
The observed abundance pattern is similar to those obtained for A 496 (Tamura et al. 2001), and NGC 4636 (Xu et al. 2002). Therefore, this abundance pattern may be uniform around the central galaxies of groups or clusters.
Comparison of the abundance pattern with Galactic stars
-------------------------------------------------------
Figures \[stars\] and \[stars2\] show the observed abundance ratios of O, Mg, Si, Ca and Fe compared to those of the Galactic disk stars (Edvardsson et al. 1993). The average abundance ratios of low metallicity (i.e. \[Fe/H\]$<$-0.8) stars of the Galaxy by Clementini et al. (1999), which should reflect the average abundance pattern of SN II products of our Galaxy are also plotted. Other papers on Galactic low metallicity stars give similar abundance ratios. \[O/Fe\] ratios for stars with \[Fe/H\]$<-1$ derived by Nissen et al. (1994) and Peimbert (1992) are 0.48$\pm0.16$, and 0.5, respectively. Gratton & Sneden (1991) gives \[Si/Fe\] of 0.3 for these stars.
The observed abundance pattern of M 87 is located at a simple extension of that of Galactic stars, although the observed Fe/O range of M 87 is systematically larger.
The Mg/O ratio of the ICM is plotted assuming the ratio is constant within the observed region. We adopted the ratio derived from the line ratio, using the Fe-L modeling with the APEC code, since APEC gives better fits around the Mg-K/Fe-L region at 1.3–1.5 keV. \[Mg/O\] of the Galactic stars is almost constant at the same value as the ICM around M 87. The Galactic \[Ca/Si\] tends to a slightly smaller value than the that of M 87, but the difference is only 0.1 dex.
We note that Xu et al. (2002) also discovered similar values of the Fe/O and Mg/O ratio for the ISM in an elliptical galaxy, NGC 4636. The Si/Fe ratio of this galaxy observed by ASCA is also determined to be $\sim$ 1 solar (Matsushita et al. 1997; 2000). Thus, the O/Mg/Si/Fe pattern of NGC 4636 is similar to that of M 87. Therefore, the abundance pattern of the ICM of M 87 is not peculiar, but consistent, not only with an elliptical galaxy NGC 4636, but also with that of our Galaxy.
Nucleosynthesis from SN Ia
--------------------------
The observed radial change of the O/Fe ratio indicates that the contribution from SN Ia increases toward center.
Although the abundance pattern of ejecta of SN II may differ between early-type and late-type galaxies, and that of SN Ia also may not be a constant (Umeda et al. 1999; Finoguenov et al. 2002), for a first attempt we have assumed (Fe/Si)$_{\rm SN Ia}$, (Si/O)$_{\rm SN II}$, and (Fe/O)$_{\rm SN II}$ to be constants. Here, (Fe/Si)$_{\rm SN Ia}$ is the Fe/Si ratio of ejecta of SN Ia, and (Si/O)$_{\rm SN II}$ and (Fe/O)$_{\rm SN II}$ are the Si/O ratio and the Fe/O ratio of the ejecta of SN II, respectively. Since O is not synthesized by SN Ia, the Fe/O ratio is then expressed by, $$\frac{\rm{Fe}}{\rm{O}} = \left(\rm{\frac{Fe}{Si}}\right)_{\rm{SN Ia}}\left( \left(\frac{\rm{Si}}{\rm{O}}\right)-\left(\rm{\frac{Si}{O}}\right)_{\rm{SN II}}\right)+\left(\rm{\frac{Fe}{O}}\right)_{\rm{SN II}}$$ Thus, (Fe/Si)$_{\rm SN Ia}$ is the inclination in the Si/O vs. Fe/O plot (Figure \[stars\]). Even within the M 87 data, the gradient indicates that (Fe/Si)$_{\rm SN Ia}$ at the center is smaller than $\sim$ 1.3 solar. When we adopt the abundance pattern of the Galactic metal poor stars by Clementini et al. (1999) as that of SN II, (Fe/Si)$_{\rm SN Ia}$ around M 87 is determined to be $\sim 1.1 $ solar (Table \[sn1\] ). In addition, the M 87 data are located at the extension of those of Galactic stars, and the whole trend is consistent with (Fe/Si)$_{\rm SN
Ia}$ to be $\sim$ 1.1 solar. This means that also for metal rich stars in the Galaxy, the (Fe/Si)$_{\rm SNIa}$ is determined to be $\sim$ 1.3 solar. These values are a factor of 2 smaller than the standard W7 model (Nomoto et al. 1984; Table \[sn1\]), and closer to the WDD1 ratio (Iwamoto et al. 1999), which considers slow deflagration.
In the same way, we have fitted the S and Ar to Si ratio (Figure \[sosio\]) and derived the S/Si and Ar/Si ratio of SN Ia ((S/Si)$_{\rm SN Ia}$ and (Ar/Si)$_{\rm SN Ia}$) to be $\sim $ 1.5 solar and 1.3 solar, respectively (Table \[sn1\]). In contrast to the Fe/Si ratio, W7 and WDDs give nearly the same values for the ratio between the intermediate elements, S/Si and Ar/Si, and the observed ratios are slightly larger than the theoretical values.
------- --------------- ------ ------ ------ ------
M 87 SNI W7 WDD1 WDD2 WDD3
S/Si 1.57$\pm0.32$ 1.05 1.14 1.14 1.14
Ar/Si 1.25$\pm0.45$ 0.76 0.89 0.91 0.93
Fe/Si 1.10$\pm0.25$ 2.61 1.34 2.07 2.99
------- --------------- ------ ------ ------ ------
: The abundance pattern of SN Ia obtained from the ICM of M 87 and those from the nucleo synthesis model of W7 (Nomoto et al. 1984) and WDDs (Iwamoto et al. 1999). []{data-label="sn1"}
Diversity of abundance pattern in SN Ia?
----------------------------------------
The light curves of observed SN Ia are not identical but display a considerable variation (e.g. Hamuy et al. 1996). In SN Ia, the mass of synthesized Ni$^{56}$ determines the luminosity of each SN. Since the mass of the progenitor should be constant at 1.4 $M_{\odot}$, the ratio of mass of intermediate group elements from Si to Ca, to the mass of Fe and Ni, should depend on the luminosity of SN Ia. The observed luminosity of SN Ia correlates with the type of the host galaxy, and is suggested to be related to the age of the system; SNe Ia in old stellar system may have smaller luminosities (Iwanov et al. 2000), and hence are suggested to yield a smaller Fe/Si ratio (e.g. Umeda et al. 1999).
As discussed in Finoguenov et al. (2002), the smaller Fe/Si ratio observed for the ICM around M 87 may reflect the fact that M 87 is an old stellar system. When we compare the SN Ia abundance pattern for the central and outer regions of M 87, the outer zone SN Ia abundance pattern depends more heavily on the adopted abundance pattern of SN II. As in Finoguenov et al. (2002), we use the SN II pattern derived in Iwamoto et al. (1999) for which a Salpeter IMF was adopted. As indicated in Figure 17, the (Fe/Si)$_{\rm SN Ia}$ of the outer region is determined to be $\sim$ 1.7 solar, while the central region data yield a value of $\sim$ 1.2 solar (Figure \[stars\]). In contrast, when we adopt the abundance pattern of metal poor Galactic stars as a SN II pattern, the difference of (Fe/Si)$_{\rm SN Ia}$ of the central and outer regions becomes smaller. The average products of SN II may also differ between the ICM and late-type galaxies, since the IMF of stars may differ between early-type and late-type galaxies. Therefore, (Fe/Si)$_{\rm SNIa}$ of the outer region still has some uncertainty, while the central value is determined to be $\sim1.2$ solar, which is not affected by the adopted SN II pattern very much.
However, the metal poor stars, i.e. those with lower Fe/O, tend to be located around larger (Fe/Si)$_{\rm SNIa}$ values (Figure \[stars\]). This result suggests that the SN Ia were dominated by those with larger (Fe/Si)$_{\rm SNIa}$ when the Galaxy was a young stellar system. Furthermore, based on the ASCA survey of the Virgo cluster, Shibata et al. (2002) discovered that the Fe/Si ratio becomes larger at $R>30'$, where the SN Ia products are accumulated from high $z$. Although the O abundance cannot be measured, this result may reflect the fact that the Fe/Si ratio in SN Ia depends on the age of the system.
The ICM in the central and outer regions should have a different origin of metals. As discussed in Paper I, considering the gas mass and stellar mass loss rate, most of the Si and Fe at the center come from present SN Ia in M 87 in the last few Gyr. In contrast, in the outer regions, metals from SN Ia in M 87 are accumulated over a longer time scale, and in addition, other galaxies should contribute to the metals. Since the SN Ia rate is expected to be much larger in the past (Renzini et al. 1993), the SN Ia ejecta which occurred in much younger systems should be important in the outer region.
The indication that the abundance pattern of SN Ia in the outer region of M 87 features a larger (Fe/Si)$_{\rm SNIa}$ ratio compared to the inner region as found above, can nicely interpreted in line with of these previous findings. The larger (Fe/Si)$_{\rm SNIa}$ ratio originates from a younger stellar population, so SN Ia products that have been produced earlier in the history of M 87 and the Virgo cluster are more widely distributed in the ICM.
Nucleo synthesis from SN II
---------------------------
We can also constrain the abundance pattern of SN II in the ICM.
Since Mg and O are not synthesized by SN Ia, the Mg/O ratio reflects those of the products of SN II. The observed Mg/O ratio of $\sim$1.25 solar at $r>0.5'$ derived from the line ratio agrees well with that of the Galactic stars, although a higher Mg/O ratio is also allowed within $r<0.5'$ due to the uncertainties in the temperature component below 1 keV. This value is also quite similar to the Mg/O ratio of 1.3$\pm0.2$ solar of the ISM of an X-ray luminous elliptical galaxy, derived from the RGS data (Xu et al. 2002). The central Mg/O reflects the stellar abundance ratio of M 87, while that of outer region is the ratio of of metals in the ICM which is an accumulation of old SN II products ejected from galaxies. Therefore, at least for the Mg/O ratio, there is no observable difference in nucleosynthesis products of SN II between the elliptical galaxies, M 87 and NGC 4636, the ICM and the Galaxy.
The observed S/Si ratio decreases when the O/Si ratio increases (Figure 8, \[sosio\]). Considering that the observed Fe/Si ratio is constant, this indicates that S/Si ratio is anti-correlated to the O/Fe ratio, that is the contribution from SN II/SN Ia. From equation (1), using S instead of Fe, the S/O ratio in ejecta of SN II ((S/O)$_{\rm SNII}$) is obtained from the observed relation in Figure \[sosio\]. The Si/O ratio of metal poor stars by Clementini et al. (1999) is $\sim 0.7$ solar, while the nucleo synthesis model by Iwamoto et al. (1999) gives the value of 0.9 solar. Therefore assuming that (Si/O)$_{\rm SNII}$ is less than 1 solar, the extension of the observed values in Figure 19 gives (S/O)$_{\rm SNII}$ less than 0.5 solar. In the same way, (Ar/O)$_{\rm SNII}$ is derived to be less than 1 solar.
The observed S abundances of Galactic stars are generally consistent with the Si abundances (e.g. Chen et al. 2002, Takeda-Hidai et al. 2002), although S lines are very week. In contrast, the observation of type II planetary nebulae yielded a S/O and Ar/O ratio to be 0.4$\pm 0.2$ solar and 1.1$\pm 0.5$ solar (Kwitter & Henry 2001). The nucleosynthesis model of SN II by Nomoto et al. (1997) assuming Salpeter’s IMF gives the S/O ratio of 0.6 solar, while another model by Woosley & Weaver (1995) gives a larger ratio of 1 solar. Thus, there may be some uncertainty in the S/O ratio synthesized by SN II even in the Galaxy, and it is also not clear that the Galaxy and the ICM have the same ratio. The fact that the S/Si ratio of the ICM observed with ASCA generally decreases toward outer radius (Finoguenov et al. 2000) supports the low S/Si ratio in SN II, although ASCA cannot constrain O abundance. Therefore, it is important to study outer regions of clusters with the XMM-Newton where SN II may become dominant (e.g. Fukazawa et al. 2000, Finoguenov et al. 2000).
In summary, the SN II abundance pattern observed from the ICM mostly agrees well with that of our Galaxy. Since the abundance pattern of SN II depends on the IMF of stars, this consistency is able to provide strong constraints on it.
Comparison with the stellar metallicity profile
-----------------------------------------------
Figure \[ovsmg\] shows the observed Mg profile, with the stellar metallicity profile from the Mg$_2$ index (Kobayashi & Arimoto 1999). In addition to the Mg abundance profile derived from the APEC model fit, that from the Mg/Si line ratio and deprojected Si abundance profile is also plotted. The Mg abundance profile of the ICM has a slightly smaller normalization by 20$\sim$30% than the stellar metallicity profile at the same radius.
Although the Mg abundance may have some uncertainties due to the Fe-L atomic data, the APEC code can well fit the Fe-L/Mg-K structure. The uncertainty of the Mg abundance due to the uncertainty of the temperature structure should be also small, at least at $R$=0.5–2$'$, since the three temperature model does not change the result, although the abundances from the two- and single-temperature results differ by a factor of 1.5–2. Considering these uncertainties and the difference in observational techniques, we can conclude that the Mg abundance of the ICM is consistent with the stellar metallicity profile at the same radius. Since we are comparing abundances in two distinct media, stars and ISM, which could have very different histories, the abundance results do not have to agree in general. But this agreement is consistent with the picture where the central gas in M 87 comes from stellar mass loss as discussed below.
The SN Ia rate of M 87
----------------------
The metals in the ICM are a sum of metals ejected from the galaxy and those contained previously within the ICM. The observed Fe abundance becomes nearly constant within 2$'$, which indicates that the gas is dominated by gas ejected from the galaxy. The observed Mg and Fe abundances within 0.35–2$'$ are $1.1\pm0.2$ solar and 1.54$\pm0.1$ solar, respectively. Both the nucleosynthesis model by Nomoto et al. (1997) and observation of Galactic stars (e.g. Clementini et al. 1999; Nissen et al. 1994) indicate that (Fe/Mg)$_{\rm SN II}$ is 0.3–0.5 solar. Therefore, within $R<2'$, the Fe abundance synthesized by SN II should be 0.3–0.6 solar. Subtracting it, the Fe abundance synthesized by SN Ia is derived to be $\sim$ 0.9–1.4 solar.
The metallicity of gas ejected from the galaxy are a sum of stellar metallicity and the contribution from SN Ia, and can be expressed as, $$\begin{aligned}
\nonumber z^i =\frac{\alpha_*y^i_{*}+\alpha_{\rm SN}y_{\rm
SN}^i}{\alpha_*+\alpha_{\rm SN}}
\sim y_{*}^i+\langle\frac{\alpha_{\rm SN}}{\alpha_{*}} \rangle y_{\rm SN}^i\\
= y_{*}^i+\langle\frac{\theta_{\rm SN} M_{\rm SN}^i}{\alpha_{*}} \rangle\end{aligned}$$ (Loewenstein and Mathews 1991; Ciotti et al. 1991), where $z^i$ is the mass fraction of the $i$th element, $\alpha_*$ and $\alpha_{\rm
SN}$ are the mass loss rates of stars and SNe respectively, and $y_{*}^i$ and $y_{\rm SN}^i$ are their yields. $\theta_{\rm SN}$ is the SN Ia rate and $M_{\rm SN}^i$ is the mass of the $i$th element synthesized by one SN Ia. Converting to abundance in unit of the solar ratio, Fe abundance synthesized by SN Ia of gas ejected from an elliptical galaxy becomes
$$\begin{aligned}
\nonumber {{{A_*^{\rm{Fe}}}_{\rm SN Ia}}}+\left(\frac{\theta_{\rm SN} \rm M_{\rm SN}^{\rm Fe}}{\alpha_*}\right)\left(z^{\rm Fe}_{\rm solar} \right)^{-1} \\
={{{A_*^{\rm{Fe}}}_{\rm SN Ia}}}+2.1 \frac{
{\left(\frac{\theta_{\rm SN}}{1.0\times10^{-13}/{\rm yr}/L_{\rm B}}\right)}
\left(\frac{M^{\rm Fe}_{\rm SN}}{0.4M_\odot}\right)}
{{\left(\frac{\alpha_*}{1.5\times10^{-11}M_\odot /{\rm yr}/L_{\rm B}}\right)}}\end{aligned}$$
Here, ${{{A_*^{\rm{Fe}}}_{\rm SN Ia}}}$ is stellar Fe abundance synthesized by SN Ia and $M^{\rm Fe}_{\rm SN}$ is the mass of Fe synthesized by one SN Ia. $z^{\rm Fe}_{\rm solar}$ is the Fe metallicity of gas with the solar abundance. Considering the fact that recent observations of M 87 and A 496 (Tamura et al. 2001) indicate a smaller Fe/Si ratio synthesized by SN Ia than usually assumed previously, $M_{\rm Fe}$ of one SN Ia should be $\sim 0.4M_\odot$. From a theoretical stellar evolutionally model for a stellar population, the stellar mass loss rate is approximated by $1.5\times10^{-11}L_B
t_{15}^{-1.3} M_\odot /yr$, where $t_{15}$ is the age in unit of 15 Gyr, and $L_B$ is the B-band luminosity (Ciotti et al. 1991). An infrared observation, which can directly trace mass loss rate from asymptotic giant branch stars, gives a consistent value within a factor of 2 with the theoretical value (Athey et al. 2002). The SN Ia rate of elliptical galaxies is estimated to be 0.13$\pm$0.05 ${h_{75}}^2$ SN Ia/100yr/10$^{10}L_\odot$ (Cappellaro et al. 1997). Here, $h_{75}$ is a Hubble constant in unit of 75 km s$^{-1}$ Mpc$^{-1}$. Using the Hubble constant of 72$\pm8$ km s$^{-1}$ Mpc$^{-1}$ (Freedman et al. 2001), the present contribution from SN Ia to the Fe abundance becomes 1.4–3.6 solar.
Considering the large error due to a statistical error and uncertainties in bias corrections of SN Ia from optical observation, this value is consistent with the Fe abundance synthesized by SN Ia within 2$'$, 0.9–1.4 solar, which is a sum of ejecta of present SN Ia and that was trapped in stars and comes from stellar mass loss. The contribution from stars to Fe synthesized by SN Ia may be small, since stars in giant elliptical galaxies are thought to have a SN II like abundance pattern (e.g. Worthey et al. 1993).
Implication of S abundance
--------------------------
In Sec.7.5 we indicated that the S/Si ratio in SN II is much lower than 1 solar. If this fact is valid universally in the ICM, then the S/Si ratio may be a better indicator for the contribution from the SN Ia/SN II than the Fe/Si ratio. Of course, the best indicators of SN II are O, Ne and Mg, which are not synthesized in SN Ia. However, emission lines of Ne and Mg are often hidden in the strong Fe-L lines, and the O abundance can be measured only in relatively cool clusters.
As discussed in Sec.7.4 the Fe/Si ratio of SN Ia may be expected to show a variation, since the luminosity of the SN Ia should reflect the amount of Ni$^{56}$ synthesized by SN Ia. The pure SN II ejecta may provide the smaller Fe/Si ratio. Then SN Ia in young systems occur, which may have a larger Fe/Si ratio and so the Fe/Si ratio increases. Finally, SN Ia in old systems decrease the Fe/Si ratio again. In contrast, at least in Iwamoto et al. (1999), the S/Si ratio in the various SN Ia is almost constant. In addition, as shown in Sec. 5.2, the S/Si line ratio does not depend on the plasma temperature, while that of the Fe/Si is a very strong function. This means that, the derived S/Si does not depend on the temperature structure of the ICM. This is a very important advantage for the cool cores of the cluster center, where many temperature components exists. The problem is that both Si and S are synthesized in the same nucleo process, i. e. in the same region in the SN. Therefore, it is surprising that the S/Si ratio shows a radial variation. Unfortunately, the S abundance measured of the Galactic stars has very large uncertainties, and the S/Si ratio of the SN II may also depend on the IMF of stars. Therefore, it is quite important to calibrate the S/Si vs. O/Fe pattern in luminous clusters.
Summary and Conclusion
======================
Based on the temperature structure studied in Paper I, abundance profiles of 7 elements of the ICM around M 87 are obtained using the deprojected spectra. We have discussed the use of one- and two-temperature models in the fit and the application of the MEKAL and APEC codes. Two temperature models are only important in the region, R$<2$ arcmin. The previous problems with an earlier version of the APEC code are now resolved in the present version and this code provides now better fitting solutions in general. In addition, using line ratio profiles of the projected and deprojected data, the abundance ratio profiles are obtained with smaller uncertainties, considering the temperature dependence of the line ratio, The main results obtained are as follows,
- Abundance profiles of 7 elements are obtained with high accuracy (e.g. $\delta$Fe$\sim\delta$Si$\sim$5%, $\delta$O$\sim$10%).
- The Si and Fe abundance profiles have steep gradients at $>2'$ with a Fe/Si ratio of $0.9\pm0.1$ solar, and become constant within this radius at 1.5–1.7 solar.
- The S/Si ratio is about 1 solar at $r<$ 4$'$ and decreases to $\sim$ 0.7 solar at $r>10'$.
- The Ar/Si and Ca/Si are about 0.8 solar and 1.5 solar, respectively, although errors are relatively large.
- The O and Mg abundances are smaller than the abundances of Fe and the intermediate elements. The O/Si ratio is less than half solar at the center and increases with radius. The Mg/O ratio is 1.25 solar and consistent with no radial gradients at least at $r>0.5'$.
- The observed abundance pattern among O, Mg, Si, Ca and Fe of the ICM is located at an extension of that of Galactic stars. Therefore, the abundance pattern of the ICM is not peculiar and it should strongly constrain the products of SN Ia and SN II of early-type and late-type galaxies.
- The Si/O/Fe diagnostics shows large Si contributions by SN Ia, which may reflect the observational finding that SN Ia in old stellar systems are fainter.
- The observed Mg abundance of the ICM is consistent with the stellar metallicity profile from the Mg$_2$ index at the same radius. This result is consistent with stellar mass loss as the source of the gas in the very central region of M 87.
- The central abundance can be explained with the observed SN Ia rate and SN Ia model yields.
- The different radial profiles between Si and S suggest that the S/Si ratio synthesized from SN II is much smaller than the solar ratio. Then, the S/Si ratio may be a better indicator of the relative contribution from SN Ia and SN II, when O and Mg abundances cannot be observed.
We are actually very fortunate to have M 87 and the Virgo cluster as the closest galaxy cluster in our neighborhood, since it offers an ICM in just this low temperature range where the spectra are richest in spectral lines for this type of abundance diagnostics. Therefore it would be worthwhile to use this unique case for even deeper observations and more detailed investigations by X-ray spectroscopy in the future.
We would like to thank Nobuo Arimoto and Kuniaki Masai for valuable suggestions on this work. This work was supported by the Japan Society for the Promotion of Science (JSPS) through its Postdoctoral Fellowship for Research Abroad and Research Fellowships for Young Scientists.
The paper is based on observations obtained with XMM-Newton, an ESA science mission with instruments and contributions direct by funded by ESA Member States and the USA (NASA). The XMM-Newton project is supported by the Bundesministerium für Bildung und Forschung, Deutsches Zentrum für Luft und R aumfahrt (BMBF/DLR), the Max-Planck Society and the Haidenhain-Stiftung.
[DUM]{} Anders E., & Grevesse N., 1989, Geochim. Cosmochim. Acta, 53,197 Arnaud M., Rothenflug R., Boulade O., et al. 1992, A&A, 254, 49 Athey A., Bregman,J., Bregman J., & Temi, P. 2002, ApJ, 571, 272 Belsole E., Sauvageot, J.L., Böhringer, H., et al. 2001, A&A, 365, L188 Böhringer H., Nulsen P.E.J., Braun R., & Fabian A.C., 1995, MNRAS, 274, 67 Böhringer, H., Belsole, E., Kennea, J., et al. 2001, A&A, 365, L181 Cappellaro E., Turatto M.; Tsvetkov, D. Yu.; Bartunov O. S.; Pollas C.; Evans R.; Hamuy M., 1997, A&A, 322, 431 Chen Y.Q., Nissen, P.E., Zhao, G.,& Asplund, M., 2002, A&A, 390, 225 Clementini, G., Gratton, R.G., Carretta, E., & Sneden, C, 1999, MNRAS, 302, 22 Ciotti L., Pellegrini S., Renzini A., & D’Ercole A. 1991, ApJ, 376, 380 Edvardsson, E., Andersen, J., Gustafsson B., et al. 1993, 275, 101 Fabian A.C., et al. 1984, Nature, 307, 343 Feldman, U., 1992, Physica Scripta 46, 202 Finoguenov, A., David, L.P., & Ponman, T.J, 2000, ApJ, 544, 188 Finoguenov A., Matsushita, K., Böhringer, H., et al. 2002, A&A 381, 21 Freyberg, M. J., Briel, U.G., Dennerl, K., Haberl, F., Hartner, G., Kendziorra, E., Kirsch M., 2002, in Symp. New Visions of the X-ray Universe in the XMM-Newton and Chandra Era (ESA SP-488; Noordwijk: ESA) Freedman, W. L., Madore, B.F., Gibson, B.K., et al. 2001, ApJ, 553, 47 Fukazawa, Y., Makishima, K., Tamura, T., et al. 1998, PASJ, 50, 187 Fukazawa, Y., Makishima, K., Tamura, T., et al. 2000, MNRAS, 313, 21 Gastaldello, F., & Molendi, S., 2002, ApJ, 572, 160 Gratton, R.G., & Sneden, C., 1991, A&A, 241, 501 Hamuy, M., Philips, M.M., Suntzeff, N.B., et al. 1996, AJ, 112, 2438 Iwamoto, K., Brachwitz, F., Nomoto, K., et al. 1999, ApJS, 125,439 Iwanov, V., Hamuy, M., & Pinto, P.A., 2000, ApJ. 542,588 Kaastra J.S. 1992, An X-Ray Spectral Code for Optically Thin Plasmas (Internal SRON-Leiden Report, updated version 2.0) Katayama, H., Takahashi, I., Ikebe, Y., Matsushita, K., & Freyberg, M., 2002., submitted to A&A (astroph/0210135) Kobayashi, C., & Arimoto, N., 1999, ApJ. 527, 573 Kodama, T., Arimoto, N., Barger, A., et al. 1998, A&A, 334, 99 Kwitter, K.B., & Henry, R.B.C, 2001, ApJ, 562, 804 Liedahl D. A., Osterheld A.L., & Goldstein W.H. 1995, ApJ, 438, L115 Loewenstein M., & Mathews W.G. 1991, ApJ, 373, 445 Masai K. 1997, A&A, 324, 410 Matsushita, K., Makishima, K., Rokudanda, E., Yamasaki, N. Y., & Ohashi, T., 1997, ApJ, 488, 125 Matsushita K., Ohashi T., & Makishima K., 2000, PASJ, 52, 685 Matsushita, K., Belsole, E., Finoguenov, A., & Böhringer, H., 2002, A&A, 386, 77 (Paper I) Mewe R., Gronenschild E.H.B.M., & van den Oord G.H.J. 1985, A&AS, 62,197 Mewe R., Lemen J.R., & van den Oord, G.H.J. 1986, A&AS, 65,511 Mushotzky, R., Loewenstein, M., Arnaue, K.A., et al. 1996, ApJ, 466 Nissen, P.E., Gustafsson, B., Edvardsson, B., et al. 1994, A&A, 285, 440 Nomoto, K., Thielemann, F-K., & Wheeler, J.C., 1984, ApJ, 279, 23 Nomoto, K., Hashimoto, M., Tsujimoto, T., et al. 1997, Nuclear Physics A., 616, 79 Peimbert, M., 1992, in Elements and the Cosmos, ed. M.G.Edmunds & R.J. Terlevich (Cambridge: Cambridge Univ. Press), 196 Renzini A., Ciotti, L., D’Ercole, A., & Pellegrini, S. 1993, ApJ, 419, 52 Sakelliou, I., Peterson, J.R., Tamura, T., et al. 2002, A&A, 391, 903 Smith, R.K., Brickhouse, N.S., Liedahl,D.A., & Raymond, J.S., 2001, ApJ, 556, 91 Shibata, R., Matsushita, K., Yamasaki, N.Y., et al. 2001, ApJ, 549, 228 Shibata, R., Matsushita, K.,. Yamasaki, N.Y., et al. 2002, in preperation Shigeyama T., 1998, ApJ, 497, 587 Stanford, S. A., Eisenhardt, P. R. M., & Dickinson, M., 1995, ApJ, 450, 512 Takeda-Hidai, M., Sato, S., Honda, S., et al. 2002, ApJ, 573, 614 Tamura, T., Bleeker, J.A.M., Kaastra, J.S, et al. 2001, A&A, 379, 107 Thielemann, F-K., Nomoto, K., & Hashimoto, M. 1996, ApJ, 460, 408 Umeda H., Nomoto K., & Kobayashi C. 1999, ApJ, 522, L43 Woosley, S.E., & Weaver, T.A., 1995, ApJS, 101, 181 Worthey, G., Faber, S.M., & González, J.J. 1992, ApJ, 398, 69 Xu, H., Kahn, S., Peterson, J.R., Beher E., Paerels, F.B.S., Mushotzky, R. F., Jernigan, J.G., & Makishima, K., 2002, ApJ, 579, 600
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'We study the relationship between the metallicity of gamma-ray burst (GRB) progenitors and the probability distribution function (PDF) of GRB host galaxies as a function of luminosity using cosmological hydrodynamic simulations of galaxy formation. We impose a maximum limit to the gas metallicity in which GRBs can occur, and examine how the predicted luminosity PDF of GRB host galaxies changes in the simulation. We perform the Kolmogorov-Smirnov test, and show that the result from our simulation agrees with the observed luminosity PDF of core-collapse supernovae (SNe) host galaxies when we assume that the core-collapse SNe trace star formation. When we assume that GRBs occur only in a low-metallicity environment with $Z\lesssim 0.1 {Z_\odot}$, GRBs occur in lower luminosity galaxies, and the simulated luminosity PDF becomes quantitatively consistent with the observed one. The observational bias against the host galaxies of optically dark GRBs owing to dust extinction may be another reason for the lower luminosities of GRB host galaxies, but the observed luminosity PDF of GRB host galaxies cannot be reproduced solely by the dust bias in our simulation.'
author:
- |
Yuu Niino , Jun-Hwan Choi , Masakazu A. R. Kobayashi ,\
Kentaro Nagamine , Tomonori Totani and Bing Zhang
bibliography:
- 'reference\_list.bib'
title: |
Luminosity Distribution of Gamma-Ray Burst Host Galaxies at redshift $\lowercase{z}=1$\
in Cosmological Smoothed Particle Hydrodynamic Simulations:\
Implications for the Metallicity Dependence of GRB
---
INTRODUCTION
============
Long gamma-ray bursts (GRBs) are the brightest astronomical transient events, and they are important laboratories of high energy astrophysics in extreme conditions, as well as the tools to probe the high-redshift universe. For example, people have used GRBs to estimate the cosmic star formation rate (SFR) density at very high redshifts [e.g., @Kistler:09a]. The observational association of some of the long GRBs with energetic Type Ic supernovae [SNe, e.g., @Hjorth:03a; @Stanek:03a] is considered to be the evidence that at least some of the GRBs originate from core-collapses of very massive stars. @Zhang:09a applied multiple criteria to identify GRBs whose progenitors are related to core collapses of massive stars (the so-called “Type II” GRBs), and found that they mostly correspond to the traditional ‘long’ GRB population. Hereafter ‘GRB’ means Type II GRBs, and ‘SN’ means core-collapse SN, unless otherwise stated.
Although GRBs may originate from the core-collapses of very massive stars, the occurrence rate of GRBs is much lower than that of normal SNe, and the conditions required for a GRB to occur from a SN still remain as one of the most outstanding questions in current astrophysics. A better understanding of required conditions for a GRB to occur would constrain the physical mechanisms of GRBs, and allow us to predict the GRB detection rate for the future observations more reliably.
Some theoretical studies on the origin of GRBs using stellar evolution models suggest that a low metallicity environment may be a necessary condition for a GRB to occur [e.g., @MacFadyen:99a; @Yoon:05a; @Woosley:06a]. It has also been suggested from the observations that the metallicity distribution of GRB host galaxies at redshift $z < 0.25$ is significantly biased towards low metallicities compared to the expectation when GRBs are unbiased tracers of star formation [@Stanek:06a; @Modjaz:08a]. However, reliable spectroscopic estimates of metallicities are available only for galaxies at low redshifts [$z \lesssim 0.5$; @Savaglio:09a] while the majority of GRBs occur at higher redshift.
Furthermore, some observations suggest that the GRB host galaxies are systematically fainter than those of the core-collapse SNe [@Le-Floch:03a; @Fruchter:06a hereafter F06], indicating that the GRBs may preferentially occur in low metallicity environment, because fainter and lower mass galaxies generally have lower metallicities. These interpretations have also been supported by other theoretical studies using the models of galaxy formation and evolution [@Nuza:07a; @Lapi:08a; @Campisi:09a].
It is also reported that the GRB host galaxies at $z > 2$ have larger Ly$\alpha$ equivalent widths compared to general star-forming galaxies at similar redshifts [@Jakobsson:05a; @Fynbo:02a; @Fynbo:03a], which is possibly a result of the stronger ionizing flux emitted from low-metallicity stellar population in GRB host galaxies [@Niino:09a]. Although the Ly$\alpha$ emission property of GRB host galaxies could be used as a metallicity indicator in the studies of GRB progenitors, our current understanding of Ly$\alpha$ transfer in the interstellar medium (ISM) is still inadequate to draw a robust conclusion [@Niino:09a], and the samples of GRB host galaxies with Ly$\alpha$ detections is currently very small.
The differences between the observed probability distribution functions (PDF) of GRB and SN host galaxies at $z \sim 1$ as a function of luminosity (hereafter ‘luminosity PDF’) or size is consistent with the hypothesis that GRBs occur preferentially in low-metallicity environment (F06). However, it is not straightforward to connect the difference in the luminosity and/or size of host galaxies to the metallicity difference of GRB and SN progenitors. In fact, some studies of GRB host properties claim that GRBs can be produced in higher metallicity environments than suggested by the stellar evolution models [@Wolf:07a; @Kocevski:09a]. However, their results suffer from some uncertainties (see §\[sec:previous\]), and more tests are required to establish the connection between the metallicity dependence of GRBs and the luminosity PDF of GRB host galaxies.
In this paper, we investigate the relationship between the low metallicity preference of GRBs and the luminosity of GRB host galaxies at $z \sim 1$, where relatively large sample of observed GRB/SN host galaxies is available, using cosmological hydrodynamic simulations. We compare the predictions of our simulations with the observations to test if the low metallicity preference of GRBs predicted by the stellar evolution models is quantitatively consistent with the observed luminosity PDF of GRB host galaxies. Several studies on this topic using galaxy formation models have already been carried out [@Nuza:07a; @Lapi:08a; @Campisi:09a]. However, quantitative comparisons between the predicted luminosity PDF of GRB host galaxies and the observed luminosity PDF have not been performed yet.
This paper is organized as follows. In §\[sec:simulation\], we briefly describe our simulation code and show some relevant properties of simulated galaxies, such as the luminosity function (LF), dust extinction, and luminosity–metallicity relationship. In §\[sec:result\], we describe the modeling of GRB/SN event rate in the simulated galaxies, and compare the resulting luminosity PDF of the GRB/SN host galaxies to the observations. In §\[sec:discussion\], we discuss the effect of possible observational bias by the optically dark GRBs on the luminosity PDF of GRB host galaxies. Then we compare our results with previous studies. Our conclusions are summarized in §\[sec:conclusion\].
NUMERICAL METHOD AND BASIC DATA {#sec:simulation}
===============================
Simulations and Galaxy Identification {#sec:code}
-------------------------------------
We use the modified version of the tree-particle-mesh smoothed particle hydrodynamics (SPH) code GADGET-3 [originally described in @Springel:05a]. In this code, the SPH calculation is performed based on the entropy conservative formulation [@Springel:02a]. Our conventional code includes radiative cooling by H, He, and metals [@Choi:09a], heating by a uniform UV background of a modified @Haardt:96a spectrum, star formation, supernova feedback, phenomenological model for galactic winds, and a sub-resolution model of multiphase ISM and star formation [@Springel:03a].
In this paper, we adopt the following fiducial cosmology which is consistent with the latest WMAP result: $\Omega_m = 0.26$, $\Omega_{\Lambda} = 0.74$, $\Omega_b = 0.044$, $h=0.72$, $n_{s}=0.96$, and $\sigma_{8}=0.80$ [@Komatsu:09a; @Komatsu:10a]. We use two runs with different box sizes and resolution: N400L34 and N400L100 run. See Table \[table:sim\] for the simulation parameters.
We identify galaxies in the simulation at $z=1.0$ using a simplified variant of the SUBFIND algorithm [@Springel:01a; @Choi:09a]. In more detail, the code first computes a smoothed baryonic density field to identify candidate galaxies with high density peaks. The full extent of these galaxies are found by adding gas and star particles to the galaxies in the order of declining density. If all $N_{min}$ nearest neighbor particles have lower densities, this group of particles is considered as a new galaxy. Here, $N_{min}$ is the minimum number of gas and star particles that constitute one isolated galaxies. In this paper we set $N_{min} = 32$. If there is a denser neighbor, the particle is attached to the galaxy to which its nearest denser neighbor already belongs to. If two nearest neighbors belong to different galaxies and one of them has less than $N_{min}$ particles, then the two galaxies are merged. If two nearest neighbors belong to different galaxies and both of them has more than $N_{min}$ particles, the particles are attached to the larger galaxy, leaving the other galaxy intact. In addition, the gas particles in galaxies should be denser than $0.01 \rho_{th}$, where $\rho_{th}$ is the star formation density threshold [@Springel:03a]. In the current simulations, we use the SF threshold density of $n_{th}=0.6$cm$^{-3}$ [@Choi:09c].
[cccccc]{} N400L34 & 33.75 & $2\times 400^3$ & $3.60 \times 10^7$ & $7.33 \times 10^6$ & 3.375 N400L100 & 100.0 & $2\times 400^3$ & $9.12 \times 10^8$ & $1.91 \times 10^8$ & 6.45 \[table:sim\]
Distribution of Gas and Stars {#sec:distribution}
-----------------------------
We first examine the distribution of gas and stars in some simulated galaxies. Figure \[fig:example34\] is an example of a typical faint GRB host galaxy with ${M_{\rm UV}}=-19.0$, while Figure \[fig:example100\] is a more luminous galaxy with ${M_{\rm UV}}=-20.3$. In the lower panels of Figures \[fig:example34\] and \[fig:example100\], the distribution of gas with high- and low- metallicity is shown separately. In both examples, the high-metallicity gas is concentrated at the center of its host galaxy, and the low-metallicity gas is distributed more broadly.
It is reported that the observed GRBs are primarily located in the brightest regions of galaxies compared to SNe (F06), and the concentration of high-metallicity gas at the center of galaxies in our simulation may seem to contradict with the low-metallicity preference of GRBs. However, if the GRBs occur in the young star clusters that are not heavily enriched yet [@Larsson:07a], then their occurrence in bright regions would be unrelated to metallicity. We also see from this figure that our simulations do not have adequate resolution to resolve the disk structure with spiral arms, therefore it is difficult to discuss the spatial distribution of GRB sites within a single galaxy using our current cosmological simulations.
![ Distribution of gas ([*top left*]{}) and stars ([*top right*]{}) in an example galaxy from the N400L34 run. This galaxy has ${M_{\rm UV}}=-19.0$, $M_B=-19.3$, ${M_\star}=3.5\times 10^{9} {M_\odot}$, and ${Z_{\rm galaxy}}= 0.33 {Z_\odot}$. Distributions of gas with $Z>2.0{Z_{\rm galaxy}}$ and $Z<0.2{Z_{\rm galaxy}}$ are shown in the [*bottom left*]{} panel and the [*bottom right*]{} panel, respectively. []{data-label="fig:example34"}](figure1.eps)
![ Same as Fig.\[fig:example34\], but for a more luminous galaxy. This galaxy is from the N400L100 simulation, and has the following properties: ${M_{\rm UV}}=-20.3$, $M_B=-21.8$, ${M_\star}=4.5\times 10^{11} {M_\odot}$, and ${Z_{\rm galaxy}}= 1.2{Z_\odot}$. []{data-label="fig:example100"}](figure2.eps)
Dust Extinction Model {#sec:modelEbv}
---------------------
We compute the spectra and luminosities of simulated galaxies by applying the GALAXEV population synthesis code [@Bruzual:03a] to each constituent star particles and employing a simple model of dust extinction described below. To estimate the degree of extinction in each simulated galaxy, we assume that $E_{B-V}$ is proportional to the metal mass column density of the ISM: $\sigma_{\rm ISM} \times Z_{\rm ISM}$, where $\sigma_{\rm ISM}$ and $Z_{\rm ISM}$ are the gas column density and metallicity of each galaxy, respectively. The proportionality constant is fixed so that it agrees with the empirical estimate of gas-to-dust ratio in the Milky Way [$N_{\rm H}/E_{B-V} = 5.8\times 10^{21}$cm$^{-2}$; @Bohlin:78a] when $Z_{\rm ISM} = {Z_\odot}$: $$E_{B-V} = \frac{f_{\rm H}}{5.8\times10^{21}}
\frac{\sigma_{\rm ISM}}{m_p} \frac{Z_{\rm ISM}}{{Z_\odot}},$$ where $f_{\rm H} = 0.75$ is the hydrogen mass fraction and $m_p$ is the proton mass. This formulation also agrees with the gas-to-dust ratio in the Small Magellanic Cloud (SMC) $N_{\rm H}/E_{B-V} \sim 10^{23}$ cm$^{-2}$ [e.g., @Bouchet:85a; @Fitzpatrick:85a; @Tumlinson:02a], given that the metallicity in SMC is $\sim 10^{-1} {Z_\odot}$. We use the @Calzetti:97a extinction curve to calculate the extinction at each wavelength.
To estimate $\sigma_{\rm ISM}$ for each simulated galaxy, we assume that the ISM in each simulated galaxy follows the @Hernquist:90a mass profile with a total gas mass $M_{\rm ISM}$ of the galaxy. Then the mean $\sigma_{\rm ISM}$ within a radius $r$ is $$\sigma_{\rm ISM} = \frac{M_{\rm Hernquist}(< r)}{\pi r^2}
= \frac{M_{\rm ISM}}{8\pi (r+r_c)^2},$$ where the characteristic radius of the Hernquist mass profile $r_c$ is fixed such that $M_{\rm Hernquist}(< l_{\rm smooth,0}) = m_{\rm part,0}$. Here, $l_{\rm smooth,0}$ and $m_{\rm part,0}$ are the SPH smoothing length and the mass of the central gas particle with the highest gas density in the galaxy. For simplicity, we further assume $\sigma_{\rm ISM} = M_{\rm ISM} / 8\pi r_c^2$, which gives the value in the central region of the galaxy ($r \ll r_c$), and set $Z_{\rm ISM}$ to the metallicity of the central gas particle, because the highest density region of the galaxy is likely to be the main contributor to both the UV luminosity and GRB/SN production. Note that a single gas particle in our simulation represents a gas mass of $\sim 10^7 - 10^8 {M_\odot}$, and $l_{\rm smooth,0}$ is typically on the order of kpc.
In Figure \[fig:Z-Ebv\], we show the distribution of simulated galaxies on the metallicity vs. extinction plane. Note that the metallicity used here is the mean gas metallicity of the whole galaxy (${Z_{\rm galaxy}}$). Only galaxies with ${M_\star}> 10^{9.5} {M_\odot}$ are shown in this figure, because the dust extinction in most of the lower mass galaxies are negligible ($E_{B-V} \lesssim 0.01$). The galaxies in the N400L34 run have lower $E_{B-V}$ than those in the N400L100 run for a given metallicity, indicating that $\sigma_{\rm ISM}$ is lower in the N400L34 run on average. This is consistent with the finding by @Choi:09a that the galaxies with a lower resolution run have higher gas mass fractions, because a higher resolution run can resolve higher density peaks and hence more gas is consumed for star formation.
![ Extinction vs. metallicity for the simulated galaxies. The solid (red) and dashed (blue) contours are for the N400L34 and N400L100 runs, respectively. The contours represent $d^3n_{\rm gal}/dVdZdE_{B-V}$ = 0.01, 0.1,10, 100, and 1000 \[$h^3$Mpc$^{-3}$mag$^{-1}$\].[]{data-label="fig:Z-Ebv"}](figure3.eps)
Luminosity Function of Simulated Galaxies {#sec:luminosity}
-----------------------------------------
Figure \[fig:LF\] shows the LFs of simulated galaxies in the N400L100 and N400L34 runs, computed using the GALAXEV population synthesis code [@Bruzual:03a] and the dust extinction model described in §\[sec:modelEbv\].
Each simulation can resolve dark matter halos and galaxies only in a limited mass range due to its limited box-size and resolution. In terms of the galaxy properties, the N400L100 run contains a larger number of massive, brighter galaxies than the N400L34 run due to its larger box size. The N400L34 run can resolve lower-mass, fainter galaxies better than the N400L100 run due to its higher resolution.
The rest-frame $B$-band LFs are shown in the [*top left*]{} panel (dust extinction effect included) of Figure \[fig:LF\] with Poisson error bars. It is clear that the N400L100 run misses the faint galaxies with $M_B \gtrsim -19$, while the volume density of bright galaxies with $M_B \lesssim -20$ in the N400L34 run is smaller than that in the N400L100 run.
Since the two runs probe galaxies with different masses and luminosities, we combine the two LFs by taking the larger of the two runs at each $M_B$, and compare the result with the observation by @Faber:07a in the [*top right*]{} panel of Figure \[fig:LF\]. The combined LF agrees quite well with the observation when the dust extinction effect is included.
We also compare the rest-frame UV LF with the observation [@Dahlen:07a] in the [*bottom*]{} panels of Figure \[fig:LF\]. Contrary to the case of rest-frame $B$-band LF, the discrepancy between the observation and simulation is significant at the bright-end of the rest-frame UV LF. We discuss the effects of this overprediction on our luminosity PDF of GRB host galaxies in §\[sec:bright\].
Besides the bright-end overprediction, according to the comparison in Figure \[fig:LF\], the simulation seems to overestimate the galaxy LF at any ${M_{\rm UV}}$ by a factor of 1.5 at $z=1.0$. However, this over-estimation is probably within the range of uncertainties in our modeling and the cosmic variance in observations. It should be noted that the vertical shift of LF (e.g., the uniform overprediction at any ${M_{\rm UV}}$) does not affect our luminosity PDF of GRB/SN host galaxies.
Metallicity of Simulated Galaxies {#sec:metal}
---------------------------------
The distribution of simulated galaxies on the stellar mass vs. metallicity (${M_\star}- Z$) plane is shown in the [*left panel*]{} of Figure \[fig:Zgalaxy\]. Here the metallicity means the mean gas metallicity of a galaxy. The N400L34 galaxies are distributed around the empirical formula at $z=1.0$ proposed by @Savaglio:05a, which agrees well with the observational data points by @Perez-Montero:09a. The N400L100 galaxies are distributed on the more massive and luminous side than the N400L34 galaxies as expected from the luminosity functions. In both runs, more massive galaxies have higher metallicities than lower mass galaxies.
We note that the N400L100 galaxies have somewhat lower metallicities on average than those in the N400L34 run for the same ${M_\star}$. This can be understood as follows. In our simulation, chemical enrichment is calculated as an instantaneous feedback from star formation, and the metallicity of a galaxy increases as the gas is turned into stars. Therefore the metallicity of a galaxy is anti-correlated with the gas mass fraction (${f_{\rm gas}}$) of the galaxy. As mentioned in §\[sec:modelEbv\], a higher resolution simulation can consume gas in star formation more efficiently than a lower resolution simulation. Furthermore, if we compare galaxies with similar ${M_\star}$, galaxies with larger ${f_{\rm gas}}$ have larger total (star and gas) mass, and thus reside in lager simulation box. The N400L100 run has a higher ${f_{\rm gas}}$ compared to the N400L34 run due to its larger box size and a lower resolution, leading to a lower metallicity of galaxies. As discussed in §\[sec:luminosity\], the two simulation boxes complement each other, and each of the simulation cannot reproduce the observations alone for a wide dynamic range. If we combine the two distributions of galaxies shown in the [*left panel*]{} of Figure \[fig:Zgalaxy\], the overall distribution of the simulated galaxies agrees with the observations.
It should be noted that the mean gas metallicity (${Z_{\rm galaxy}}$) of a simulated galaxy depends on the threshold density of gas particles to be included in galaxies, which is a parameter in the galaxy finding algorithm described in §\[sec:code\]. If we include lower density gas particles than the current density threshold, each simulated galaxy may have lower ${Z_{\rm galaxy}}$, because gas particles in the outskirts of a galaxy have lower metallicity than those near the center (see §\[sec:distribution\]). However, the dependence of ${Z_{\rm galaxy}}$ on the grouping density threshold does not affect our luminosity PDF of GRB/SN host galaxies. Low density gas particles are not forming stars, and hence they do not contribute to the GRB production and UV luminosity.
The distribution of the simulated galaxies on the $L_{\rm UV}-Z$ plane is also shown in the right panel of Figure \[fig:Zgalaxy\], which is very similar to the ${M_\star}-Z$ distribution.
RESULTS {#sec:result}
=======
Metallicity of GRB Sites {#sec:GRBmetal}
------------------------
In this section, we explore the metallicity of GRB sites. We first assume that the GRB occurrence rate in each simulated galaxy is proportional to the total SFR without a metallicity limit. Then we examine the models in which the GRB rate is proportional to the total SFR in gas particles with $Z<{Z_{\rm crit}}$. Our interest is in the luminosity PDF of GRB/SN host galaxies, therefore we do not need to consider the absolute value of the event rate density.
![ SFR-weighted metallicity distributions of gas particles in the simulated galaxies with log$_{10}({Z_{\rm galaxy}}/{Z_\odot}) = 0.0\pm0.15,\ -1.0\pm0.15$, and $-2.0\pm0.15$ are plotted with solid (red), dashed (green), and dotted (blue) lines, respectively. In our model, we assume that the GRB rate is proportional to the sum of SFR for the gas particles with $Z<{Z_{\rm crit}}$. The top and bottom panels are showing the distributions for the N400L34 and N400L100 simulations. The large fluctuation in the particle metallicity distribution in galaxies with ${Z_{\rm galaxy}}= 0.01 {Z_\odot}$ (blue dotted line) is due to the small number of gas particles included in such low-metallicity galaxies. Colors are only for the online version. []{data-label="fig:Zpart"}](figure6.eps)
In order to investigate the effect of metallicity limit, we first examine the SFR-weighted distribution of gas particles as a function of metallicity. In Figure \[fig:Zpart\], we divide the galaxy population into three different samples according to the mean gas metallicity of the host galaxy (${Z_{\rm galaxy}}$). We can see that the metallicity of star forming gas particles has a wider range than that of ${Z_{\rm galaxy}}$, showing the large dispersion of gas metallicity in each galaxy. We note that the high metallicity gas is concentrated in the high density regions of its host galaxy (see §\[sec:distribution\]) where SFR is high. Therefore the peak metallicity of SFR-weighted distribution is higher than the mean ${Z_{\rm galaxy}}$ of the sample. The distributions for both N400L34 and N400L100 runs are very similar to each other.
Assuming that the GRB rate is proportional to the sum of SFR in gas particles with $Z<{Z_{\rm crit}}$, we can see in Figure \[fig:Zpart\] that, if we set ${Z_{\rm crit}}= 0.1\,{Z_\odot}$, then the GRB event rate in galaxies with ${Z_{\rm galaxy}}\sim 1.0\,{Z_\odot}$ (0.1${Z_\odot}$) would be reduced to a few percent (40%) compared to the case of no ${Z_{\rm crit}}$. With ${Z_{\rm crit}}= 0.1\,{Z_\odot}$, the event rate in galaxies with ${Z_{\rm galaxy}}\sim 0.01\,{Z_\odot}$ would be almost unchanged from the case of no ${Z_{\rm crit}}$. The large fluctuation in the distribution for galaxies with ${Z_{\rm galaxy}}\sim 0.01{Z_\odot}$ is due to the small number of gas particles included in such low-metallicity galaxies.
Rest-frame UV Luminosity of the Host Galaxies {#sec:host}
---------------------------------------------
Next, in order to discuss the luminosity PDF of host galaxies, we consider the SFR-weighted fraction of galaxies as a function of rest-frame UV luminosity as shown in Figure \[fig:hostLF\]. The luminosity PDF in the case of no ${Z_{\rm crit}}$ is plotted in [*panel (a)*]{} for the N400L34 and N400L100 runs. As mentioned in §\[sec:luminosity\], the two simulations have different resolving power of galaxies with different mass scales, therefore we combine the two distributions by taking the larger of the two runs in units of ${M_\odot}$yr$^{-1}\,h^3$Mpc$^{-3}$mag$^{-1}$ as shown in [*panel (a)*]{}. The distributions are normalized so that the integration of the combined luminosity PDF for each ${Z_{\rm crit}}$ is unity. Hereafter we discuss the combined luminosity PDFs unless otherwise mentioned.
![ ([*a*]{}) Luminosity PDF of the simulated host galaxies ($z=1.0$) without any metallicity cutoff. The results of the N400L34 and N400L100 run are plotted with double-dotted (black) and double-dot dashed line (orange), respectively. The combined distribution of the two simulations (i.e., the larger of the two distributions) is plotted with solid line (red). ([*b*]{}) Combined luminosity PDFs of the simulated host galaxies with no ${Z_{\rm crit}}$, ${Z_{\rm crit}}/{Z_\odot}= 0.5$, and 0.1 are plotted with the solid (red), dotted (magenta), and dot-dashed line (blue), respectively. ([*c*]{}) Luminosity PDFs of the observed sample of GRB host galaxies and SN host galaxies at $z < 1.2$ (F06) are plotted with dotted (dark blue) and dashed (dark red) histograms, respectively. The arrows indicate the median of the distributions. Colors are only for online version. []{data-label="fig:hostLF"}](figure7.eps)
In [*panel (b)*]{} of Figure \[fig:hostLF\], the three combined luminosity PDFs of the simulated host galaxies with ${Z_{\rm crit}}=$ none, 0.5${Z_\odot}$, and 0.1${Z_\odot}$ are shown. The luminosity PDF shifts to the fainter magnitudes as ${Z_{\rm crit}}$ becomes lower. In the case of ${Z_{\rm crit}}=0.1{Z_\odot}$, the host galaxies are fainter than in the no-${Z_{\rm crit}}$ case by $\sim 1$ magnitude. This result is consistent with the luminosity PDFs of the observed GRB/SN host galaxies at $z<1.2$ (the F06 sample) as shown in [*panel (c)*]{}. The redshift criterion of $z<1.2$ was proposed in F06 to reduce the effect of galaxy evolution on the comparison between the GRB and SN host galaxies. The luminosity PDF of F06 was in the observed frame $V$-band, which is close to the UV in the galaxy’s rest-frame. The median values of each distribution in [*panels (b)*]{} and [*(c)*]{} are indicated by the arrows.
It should be noted that, using a larger observed sample, @Svensson:10a found a smaller difference between the GRB and SN host galaxy luminosity PDFs than that claimed by F06. However, it is difficult to understand the selection effect in the GRB host sample of @Svensson:10a, because a part of it is drawn from the GHostS project database, which consists of publicly available data of GRB host observations done by various groups with differing instruments [@Savaglio:09a]. On the other hand, the F06 sample is a complete sample of host galaxies for all 42 GRBs that have optical afterglow detections ($>3\sigma$), therefore the sampling effect would be smaller in the F06 sample than in the @Svensson:10a sample. Thus we use the GRB host galaxy sample by F06 as a reference data for the discussions in this paper. The 24 out of the 42 GRBs in the F06 sample are at $z<1.2$. We note that the SN host galaxies in the @Svensson:10a sample is significantly fainter than those in the F06 sample, and this is probably one of the causes for the smaller difference between the GRB and SN host galaxies. But the reason for the difference between the SN host samples of F06 and @Svensson:10a is not known.
Even the F06 sample would suffer from the observational bias against dusty host galaxies. We will discuss the selection effect caused by the dust extinction in §\[sec:dustbias\]. We note that the luminosity PDFs studied by F06 and @Svensson:10a are at different wavelengths \[rest-frame UV for F06 versus rest-frame $B$ & $V$-band for @Svensson:10a\], though this cannot explain the difference in their results. The difference in the luminosity PDFs of GRB and SN host galaxies should be larger in the optical than in UV, given that many of observed GRB host galaxies are extremely blue [e.g., @Le-Floch:03a].
Cumulative Luminosity PDF {#sec:cumulative}
-------------------------
We plot the cumulative luminosity PDFs of the simulated and observed host galaxies in Figure \[fig:cumUV\]. The simulated PDFs are shown for the cases of ${Z_{\rm crit}}/{Z_\odot}=$ none, 0.5, 0.1, 0.025, & 0.001 (top to bottom). The observed host galaxy sample at $z<1.2$ of F06 is plotted together as histograms. The shift of luminosity PDF to fainter magnitudes is clearly seen also in the cumulative plot, reproducing the difference in the observed GRB and SN-host galaxy luminosity PDFs. However, the distribution for no-${Z_{\rm crit}}$ case overpredicts the bright-end of the observed SN-host luminosity PDFs, and the ${Z_{\rm crit}}\leq 0.1{Z_\odot}$ case overpredicts the bright-end of the observed distribution of GRB host galaxies. This is expected from the overprediction of the UV LF at the bright-end as we discussed in §\[sec:luminosity\] (see the bottom panel of Fig. \[fig:LF\]). We discuss the effects of this overprediction on our conclusion in §\[sec:bright\].
![ Cumulative luminosity PDFs of the simulated host galaxies is shown together with those of the observed GRB/SN host galaxies of F06 at $z < 1.2$, based on the histograms shown in Fig. \[fig:hostLF\]. The PDF without a metallicity cutoff (no ${Z_{\rm crit}}$) is shown with the solid line (red), and the those with ${Z_{\rm crit}}/{Z_\odot}=$ 0.5, 0.1, 0.025, and 0.001 are represented by the dot-dashed lines (blue) from top to bottom, respectively. Colors are only for online version. []{data-label="fig:cumUV"}](figure8.eps)
Kolmogorov-Smirnov Test {#sec:KStest}
-----------------------
To examine the consistency between the simulated and observed cumulative luminosity PDFs shown in Figure \[fig:cumUV\], we employ the one-sample Kolmogorov-Smirnov (K-S) test, which calculates the probability that a sample distribution is consistent with a given PDF. Here, the sample is the observed luminosity of GRB/SN host galaxies, and the PDFs are the results of our simulation. Figure \[fig:KStest\] shows the K-S probability ${P_{\rm KS}}$ (i.e., one minus the rejection confidence level for consistency) as a function of ${Z_{\rm crit}}$ of simulated host galaxies. The ${P_{\rm KS}}$ takes a value between 0 and 1, and ${P_{\rm KS}}<10^{-4}$ means that the observed sample is dissimilar to the simulated one at a $3\sigma$ level, while a larger ${P_{\rm KS}}$ means a higher probability that the observed sample is derived from a PDF calculated in our simulation. For the SN host galaxies, we find ${P_{\rm KS}}> 0.05$ when ${Z_{\rm crit}}\geq 0.03{Z_\odot}$, and ${P_{\rm KS}}> 0.2$ when ${Z_{\rm crit}}\geq 0.5{Z_\odot}$. On the other hand, for the GRB host galaxies, we find ${P_{\rm KS}}> 0.05$ (0.2) when ${Z_{\rm crit}}/{Z_\odot}\leq 0.1$ (0.03). The KS probabilities are the highest when ${Z_{\rm crit}}/{Z_\odot}=0.005$ and 0.5 for the GRB and SN host galaxies, respectively. This clearly supports the low-metallicity preference for GRB host galaxies.
The luminosity PDF of the simulated host galaxies does not change significantly for ${Z_{\rm crit}}\leq 0.1{Z_\odot}$, and the agreement with the observed PDF is reasonably good for all ${Z_{\rm crit}}$ values for ${Z_{\rm crit}}< 0.1{Z_\odot}$. Therefore we cannot put a lower limit to ${Z_{\rm crit}}$ based on this comparison. However our simulation cannot resolve the galaxies with ${M_\star}<10^8{M_\odot}$, therefore it is possible that the above result suffers from the resolution limit of our simulation. We should be cautious about the results with ${Z_{\rm crit}}< 0.1{Z_\odot}$, but our simulation contains galaxies as faint as ${M_{\rm UV}}\sim -14$ (see the bottom panels of Figure \[fig:LF\]), which are well below the typical luminosity of the simulated host galaxies with ${Z_{\rm crit}}< 0.1{Z_\odot}$.
![ K-S probabilities (i.e., 1.0 minus the rejection confidence level for consistency) of the luminosity PDFs of the observed and the simulated host galaxies is plotted against ${Z_{\rm crit}}$. The symbols are for the SN host galaxies (red empty circles) and GRB host galaxies (blue filled circles). []{data-label="fig:KStest"}](figure9.eps)
Overprediction of UV-bright Galaxies {#sec:bright}
------------------------------------
In § \[sec:luminosity\] and Figure \[fig:LF\], we showed that our N400L100 simulation overpredicts the bright-end of the rest-frame UV LF even after the dust extinction correction. We note however that our simulation agrees very well with the observed rest-frame $B$-band LF, which suggests that either our simulation overpredicts the amount of young stars compared to the older stellar population, or the two sets of observations are inconsistent with each other, or else the Calzetti extinction curve that we adopted is not appropriate for the observed galaxies. The source of this discrepancy is unclear, therefore we focus only on the rest-frame UV LF in this subsection.
To quantify the effect of the bright-end overprediction on our luminosity PDFs, we perform the same K-S test for the following two different test models that modify the shape of the rest-frame UV LF by hand.
$\bullet$ [**Model 1**]{}: We double the gradient of the rest-frame UV LF at the bright-end (${M_{\rm UV}}<-20$) as follows: $$\begin{aligned}
\begin{array}{ll}
M_{\rm UV,model\ 1} = {M_{\rm UV}}+ \Delta{M_{\rm UV}}\\
\Delta{M_{\rm UV}}= \left\{
\begin{array}{ll}
-0.5\times ({M_{\rm UV}}+ 20), & \ \ {M_{\rm UV}}< -20 \\
0, & \ \ {M_{\rm UV}}\geq -20. \\
\end{array} \right.
\end{array} \label{eq:modulation}\end{aligned}$$
The SFR in galaxies with ${M_{\rm UV}}<-20$ is also modified in a consistent manner, motivated by the inference that the overprediction of the bright-end of the UV LF is caused by an overestimation of SFR in massive galaxies [e.g., due to lack of active galactic nucleus feedback in our simulation; @Choi:09c]. The change of SFR in each simulated galaxy is as follows: $$\Delta({\rm log_{10}SFR}) = \Delta({\rm log_{10}}L_{\rm UV}) = -\frac{\Delta{M_{\rm UV}}}{2.5}.
\label{eq:dSFR}$$
$\bullet$ [**Model 2**]{}: We shift the entire UV LF uniformly by $\Delta{M_{\rm UV}}= 1.0$ to fit the observation. In this model, the luminosity PDF will also shift by the same amount of $\Delta{M_{\rm UV}}$, without changing its shape.
Both of these models are ad hoc modifications to fit the observed data points, but they would allow us to evaluate the effects of overprediction of UV LF by our simulation. For comparison purposes, we call the earlier calculations presented in §\[sec:cumulative\] – \[sec:KStest\] the [**Model 0**]{}. We show the LFs of Model 0, 1 & 2 galaxies in Figure \[fig:modLF\]. One sees that the overprediction of the bright-end in Model 0 is removed in Model 1 & 2. The uniform overprediction at any ${M_{\rm UV}}$ is not removed in Model 1, but the uniform overprediction does not affect the luminosity PDF of GRB/SN host galaxies, as we have mentioned in § \[sec:luminosity\].
![ Rest-frame UV luminosity functions of the simulated galaxies in Model 0, 1 & 2 are plotted together with the observation [@Dahlen:07a]. The dust extinction model is included. The original result of our simulation (Model 0, identical to the combined simulation shown in the [*bottom right panel*]{} of Fig.\[fig:LF\]) is shown with the dotted (black) line. The modified LF with the doubled gradient of the bright-end slope (Model 1) and the horizontally shifted LF by $\Delta{M_{\rm UV}}= 1.0$ (Model 2) are represented by solid (red) and dashed (blue) curves, respectively. Vertical dashed line represents ${M_{\rm UV}}= -20$, which is the threshold of the bright-end modification in Model 1. []{data-label="fig:modLF"}](figure10.eps)
![ Same as Fig.\[fig:KStest\], but for the two different models of modified UV LF (Model 1 & 2). The Model 1 & 2 are represented by the squares and triangles, respectively. []{data-label="fig:modKS"}](figure11.eps)
Figure \[fig:modKS\] shows the K-S probabilities of Models 1 & 2 for the observed GRB/SN host galaxies. For the SN host galaxies, the no-${Z_{\rm crit}}$ case has the highest ${P_{\rm KS}}$, and ${P_{\rm KS}}$ decreases with decreasing ${Z_{\rm crit}}$. For the GRB host galaxies, the Model 1 gives the highest probability of ${P_{\rm KS}}\sim 0.6$ at ${Z_{\rm crit}}=0.1\,{Z_\odot}$, and ${P_{\rm KS}}\approx 0.3-0.4$ at ${Z_{\rm crit}}< 0.1\,{Z_\odot}$. In the case of Model 2, the trend is opposite of what we saw in Figure \[fig:KStest\] for ${Z_{\rm crit}}<0.5\,{Z_\odot}$ in the sense that ${P_{\rm KS}}$ increases with increasing ${Z_{\rm crit}}$. But the Model 2 still prefers ${Z_{\rm crit}}= 0.5{Z_\odot}$ to the no-${Z_{\rm crit}}$ case. The Model 2 gives ${P_{\rm KS}}< 0.2$ across all ${Z_{\rm crit}}$ values for the SN host galaxies, which means that it doesn’t agree well with the observations of SN host galaxies compared to the Models 0 and 1.
Discussions {#sec:discussion}
===========
Effect of Dust Extinction in the Host Galaxies
----------------------------------------------
### Observational bias against dusty host galaxies {#sec:dustbias}
In this section we discuss the possibility that the systematic faintness and the low-metallicity of the observed GRB host galaxies is caused by the observational bias against dusty host galaxies, [*not*]{} by the intrinsic nature of GRB progenitors. It is difficult to identify the GRB host galaxy without a detection of an optical afterglow, and the afterglow flux may be reduced significantly by the dust extinction effect. In fact, some GRBs have optical to X-ray afterglow spectral index $\beta_{OX} < 0.5$[^1], which is smaller than the expectation from the standard external shock model of GRB afterglows [e.g., @Jakobsson:04a]. We note that a part of the optically dark GRB could also be caused by the effects other than the dust extinction in their host galaxies, such as the attenuation by neutral hydrogen in the intergalactic medium at $z\gtrsim 6$ [e.g., @Totani:06a; @Nagamine:08a].
![ Same as Fig.\[fig:cumUV\], but here the galaxy is excluded from the sample if the average extinction value of galaxy exceeds $E_{B-V, {\rm crit}}$. We use Model 0 in this figure. []{data-label="fig:EcumUV"}](figure12.eps)
In order to examine the effect of dust extinction in the host galaxy (as opposed to the extinction in the immediate nearby environment of the GRB), here we exclude the galaxy from the sample of simulated host galaxies if $E_{B-V}\geq E_{B-V, {\rm crit}}$, where $E_{B-V}$ is the average extinction of each galaxy computed in §\[sec:modelEbv\]. The GRB event rate in each galaxy is assumed to be proportional to the SFR in the galaxy. Figure \[fig:EcumUV\] shows the cumulative luminosity PDF of simulated host galaxies with various $E_{B-V, {\rm crit}}$. The no-$E_{B-V, {\rm crit}}$ case ([*red solid*]{} line) represents the luminosity PDF of GRB host galaxies when the GRB rate simply traces the SFR, and it is identical to the no-${Z_{\rm crit}}$ case in Figure \[fig:cumUV\]. The luminosity PDF of the simulated host galaxies shifts to the fainter magnitudes with smaller $E_{B-V, {\rm crit}}$, similarly to the ${Z_{\rm crit}}$ dependence shown in Figure \[fig:cumUV\].
Figure \[fig:EKS\] shows the K-S probability versus $E_{B-V, {\rm crit}}$ for the luminosity PDF of the Model 0 galaxies given in Figure \[fig:EcumUV\], together with the results of Model 1 & 2 galaxies. For Model 0 & 1, a reasonable agreement (${P_{\rm KS}}> 0.2$) between the simulated and observed distributions of the SN host galaxies is found for a wide range of critical extinction ($E_{B-V, {\rm crit}} \geq 0.05$), while for the GRB host galaxies the same level of agreement is achieved only when $E_{B-V, {\rm crit}} \leq 0.03$.
![ Same as Figs.\[fig:KStest\] and \[fig:modKS\], but here ${P_{\rm KS}}$ is plotted against $E_{B-V, {\rm crit}}$ instead of ${Z_{\rm crit}}$. []{data-label="fig:EKS"}](figure13.eps)
The above requirement of $E_{B-V, {\rm crit}} \leq 0.03$ for the GRB host galaxies of the Models 0 & 1 does not seem to agree with observations. Some follow up observations of optically dark GRBs suggest that a typical dust extinction to make a GRB optically dark is $A_V \sim 1$ [e.g., @Perley:09a $E_{B-V} \sim 0.3$], which is much larger than the above requirement of $E_{B-V, {\rm crit}}<0.03$. Such large values of critical extinction ($E_{B-V, {\rm crit}} > 0.1$) do not significantly change the luminosity PDF of simulated host galaxies. A small dust extinction such as $E_{B-V} < 0.03$ is unlikely to have a significant impact on the GRB optical afterglow observations.
For the Model 2, the agreement with the observed GRB hosts is good (${P_{\rm KS}}= 0.48$) at $E_{B-V, {\rm crit}}=0.1$, however, ${P_{\rm KS}}> 0.2$ is not obtained for the SN host galaxies with any values of $E_{B-V, {\rm crit}}$. The SFR-weighted luminosity PDF of the Model 2 galaxies are too faint to agree with the observation of SN host galaxies even in the no-$E_{B-V, {\rm crit}}$ case. Hence the difference of luminosity PDFs between the GRB host and SN host galaxies cannot be reproduced by the Model 2.
A caveat here is that the $E_{B-V}$ is computed in our model assuming simplified structure of ISM in each galaxy without considering smaller scale structures such as molecular clouds. Though our model of dust extinction reproduces the observed $B$-band LF of field galaxies, the model may be not representative of dust extinction for lines of sight to GRBs, if small scale structures near GRBs play an important role in the extinction of GRB optical afterglows. Recent observations of optically dark GRBs actually suggest that the effect of dust on small scales may be important [@Rol:07a; @Castro-Tirado:07a; @Perley:09a; @Hashimoto:10a; @Holland:10a]. If the typical extinction in lines of sight to GRBs is greater than the average extinction in its host galaxy by one order of magnitude but yet proportional to the average extinction, the luminosity PDF of GRB host galaxies possibly be reproduced by the observational bias against dusty host galaxies.
In this section, we have considered the possibility that the observational bias against dusty host galaxies is causing the systematic faintness of GRB host galaxies. However, we note that the SN observations possibly suffer more from dust extinction than the optical GRB afterglow observations. In that case, the observational bias against dusty host galaxies would make the SN host galaxy sample fainter than the GRB host galaxies, which is opposite to the observational result. Our simulation does not contradict with the observations even if the SN observations suffer from dust extinction, because the simulated luminosity PDF with $E_{B-V, {\rm crit}} \geq 0.1$ agrees with the SN host observations reasonably well (${P_{\rm KS}}> 0.2$) for the Model 0.
### Fraction of Optically Dark GRBs {#sec:darkfrac}
The physical nature of optically dark GRBs is also an important issue. A recent study by @Cenko:09a reported that the fraction of optically dark GRBs in the entire observed GRB sample is $\sim 50$%, while @Zheng:09a found that the fraction is $\sim 10 - 20$%. Note that these fractions are based on the entire GRB sample that span a wide redshift range, because it is difficult to constrain the redshift of optically dark GRBs.
![ Fraction of optically dark GRBs at $z=1.0$ in our simulation (i.e., the fraction of GRB events in galaxies with $E_{B-V}>E_{B-V, {\rm crit}}$) as a function of $E_{B-V, {\rm crit}}$. The results for three different values of ${Z_{\rm crit}}$ are shown. The observed range of fractions for the GRBs at all redshifts [@Cenko:09a; @Zheng:09a] is indicated with the shade. We use Model 0 in this figure. []{data-label="fig:darkfrac"}](figure14.eps)
In Figure \[fig:darkfrac\], we show the fraction of GRB event rate in galaxies with $E_{B-V} > E_{B-V, {\rm crit}}$ at $z=1.0$ in our simulation for three different values of ${Z_{\rm crit}}$. Assuming that a GRB at $z=1.0$ becomes optically dark when the host galaxy has $E_{B-V} \geq E_{B-V, {\rm crit}} = 0.3$ [$A_V\gtrsim 1,$ @Perley:09a], our simulation predicts that the fraction of optically dark GRBs is $\lesssim$ 10% for all cases of ${Z_{\rm crit}}$. For the no-${Z_{\rm crit}}$ case ([*red solid*]{} line), the dark fraction is close to 10% for $E_{B-V, {\rm crit}} = 0.3$, and is almost consistent with the observational result of @Zheng:09a. However, if ${Z_{\rm crit}}< 0.5\,{Z_\odot}$ (as suggested in §\[sec:result\]), our simulation predicts that the fraction of optically dark GRBs is $< 5$% for $E_{B-V, {\rm crit}} = 0.3$, which is much smaller than the observed fraction.
The small fraction of optically dark GRBs in our simulation compared to the observations can be explained in two different ways. The first possibility is that the optically dark GRBs mainly suffer from the absorption in the optical wavelength due to dust associated with small scale structures, not from the galaxy scale dust as modeled in §\[sec:modelEbv\]. The second possibility is that many of the observed optically dark GRBs originate at different redshifts other than $z\sim 1.0$, although the redshift distribution of GRBs with known redshifts that are typically optically non-dark, peaks at $z\sim 1.0$. The fractional contribution of dusty galaxies to the cosmic SFR would be different at different redshifts. Furthermore, the afterglow of a GRB at $z\gtrsim 6$ would be attenuated by neutral intergalactic medium in the observer-frame optical bands.
It is unclear which of the above two scenarios might be correct at this point. Some observational studies [@Castro-Tirado:07a; @Perley:09a; @Holland:10a] find that the host galaxies of optically dark GRBs have bluer colors than those expected from the dust reddening estimated from the afterglow spectra. On the other hand, the host galaxies that is globally dusty are also found for several optically dark GRBs [@Levesque:10d; @Hashimoto:10a].
Comparison with Previous Works {#sec:previous}
------------------------------
Several studies have been performed to interpret the observed luminosity PDFs of GRB and SN host galaxies. In this section we compare our study with previous results.
@Wolf:07a is one of the important studies in this context. They reproduced the luminosity PDFs of the observed GRB and SN host galaxies based on the empirical relations of galaxy properties at $z \sim 0.7$, such as the luminosity-metallicity relationship. In their study, the model in which the GRBs occur in high metallicity regions ($Z\sim {Z_\odot}$) reproduces the luminosity PDF of GRB host galaxy the best, and the models which requires GRBs to occur only in sub-solar metallicity regions do not reproduce the observations. They assumed a power-law relation for the luminosity-metallicity relation of galaxies following @Kobulnicky:04a. However this relationship is not well constrained by current observations at redshifts of our concern, and in the work of @Kobulnicky:04a, different fitting methods gave different set of best fit parameters for a set of observed data, suggesting large uncertainties in the fit. We also note that the results probably depend on the assumed dispersion around the power-law luminosity-metallicity relation.
@Kocevski:09a studied the stellar mass PDF of GRB host galaxies in a similar way to that of @Wolf:07a and argued for ${Z_{\rm crit}}> 0.5\,{Z_\odot}$. The mass-metallicity relationship of galaxies are better understood than the luminosity-metallicity relation [@Tremonti:04a; @Savaglio:05a]. However, it is difficult to understand the selection effects of the currently available samples of GRB host galaxies with known ${M_\star}$. @Kocevski:09a discussed the observed samples of GRB host galaxies with known ${M_\star}$ collected by @Castro-Ceron:08a and @Savaglio:09a. Both of the samples were collected from publicly available data of GRB host observations. @Castro-Ceron:08a collected all GRB host galaxies with rest-frame $K$-band data including upper limits, while @Savaglio:09a selected their sample requiring multi-band detection of the host galaxies. The observations of GRB host galaxies in the samples are originally done by various groups with differing instrument, and hence the selection effects in the samples are hardly understood. We also note that the constraints on the stellar masses for some of the GRB host galaxies in their sample are not accurate enough to quantitatively discuss the low-metallicity preference.
The approaches of @Wolf:07a and @Kocevski:09a are entirely based on the empirical relations, and are fundamentally different from our approach based on numerical simulations. As discussed above, one difficulty in studying the luminosity PDFs of GRB/SN host galaxies is that properties of galaxies are not observationally well understood at redshifts where the observed sample is available. Numerical simulation approach is powerful in this context, because it gives properties of galaxies such as luminosity, metallicity, and SFR without assuming uncertain relations between them.
Some numerical studies have investigated the luminosity of GRB host galaxies [@Nuza:07a; @Lapi:08a; @Campisi:09a]. However, quantitative comparisons of the predicted luminosity PDF of GRB host galaxies and the observed luminosity PDFs have not been performed in these studies. @Nuza:07a did not reproduce the luminosity PDFs of both GRB and SN hosts, presumably owing to the small box size of their simulation (10$h^{-1}$Mpc). @Lapi:08a and @Campisi:09a have shown that the fainter galaxies have systematically lower metallicities in their models. However, in @Lapi:08a and @Campisi:09a, the luminosity PDFs are calculated simply by selecting galaxies that contain young low-metallicity stars without taking the difference of GRB rate among galaxies into account, therefore the predicted luminosity PDFs cannot be directly compared with the observations of GRB host galaxies.
In this context, our present work is unique in the sense that it can quantitatively reproduce the observations of both GRB and SN host galaxies using a self-consistent cosmological hydrodynamic simulations of galaxy formation. This has been enabled by applying a widely used GRB rate model to our cosmological simulations with large box sizes, which can properly compute the properties of galaxies at $z\sim 1$.
Conclusions {#sec:conclusion}
===========
Using cosmological SPH simulations, we have examined the relation between the metallicity dependence of GRBs and the luminosity PDF of GRB host galaxies. Our results suggest that the observed difference in the UV luminosity PDFs of GRB/SN host galaxies can be explained by the low-metallicity preference of GRBs.
We find that in our simulation, the luminosity PDF of GRB host galaxies agree with the observed one when ${Z_{\rm crit}}< 0.5\,{Z_\odot}$, while the observed luminosity PDF of SN host galaxies can be reproduced without a metallicity dependence of SN rate. The suggested value of ${Z_{\rm crit}}$ for the GRBs by our study is consistent with the suggestion from stellar evolution models [@Yoon:05a; @Woosley:06a], contrary to the results of previous studies [@Wolf:07a; @Kocevski:09a]. Though our simulation suffers from the seeming overprediction of the UV luminosity function at $z=1$ at the bright end, we have explicitly demonstrated that the quantitative agreement between the simulated and observed luminosity PDFs is not a product of this seeming overestimate of the UV-bright galaxies in our simulation. However, we should be cautious about our results when ${Z_{\rm crit}}< 0.1\,{Z_\odot}$, since they may be affected by the resolution limit of our simulation.
We have also discussed the effect of observational bias against the host galaxies of optically dark GRB on the luminosity PDFs. The dust bias causes a shift of luminosity PDFs toward fainter magnitudes, similarly to the effect of the low-metallicity preference of GRBs. However, to obtain an acceptable fit to the observed luminosity PDF of GRB host galaxies, we had to assume that an optical afterglow becomes significantly faint when $E_{B-V}\sim 0.03$ in the host galaxy. Such a small extinction is unlikely to cause a significant effect for the optical follow-up observations of GRBs, and therefore it would be difficult to explain the observed luminosity PDFs of GRB and SN host galaxies only by the overall dust bias of the host galaxies. Nevertheless it is possible that our simplified model of dust extinction is not sufficient to reproduce the dust extinction of GRB optical afterglows.
Assuming that a GRB becomes optically dark when the extinction of its host galaxy is $E_{B-V}\gtrsim 0.3$ as suggested by the observation of @Perley:09a, our simulation predicts that the fraction of optically dark GRBs at $z=1.0$ is $< 5$% for ${Z_{\rm crit}}\leq 0.5\,{Z_\odot}$. This fraction is significantly smaller than what is suggested by the observations of GRBs at all redshifts [@Perley:09a; @Zheng:09a]. This is probably because we did not consider the dust extinction effect on small scales at the GRB site. It is also possible that the GRBs at redshifts other than $z\sim 1$ are important in determining the fraction. We plan to address the effect of small scale dust in the future using higher resolution simulations.
This work was supported by the Grant-in-Aid for the Global COE Program “The Next Generation of Physics, Spun from Universality and Emergence” from the Ministry of Education, Culture, Sports, Science and Technology (MEXT) of Japan. This work was also supported in part by the NSF grant AST-0807491, National Aeronautics and Space Administration under Grant/Cooperative Agreement No. NNX08AE57A issued by the Nevada NASA EPSCoR program, the President’s Infrastructure Award from UNLV, and by the NSF through the TeraGrid resources provided by the Texas Advanced Computing Center. Some numerical simulations and analyses have also been performed on the UNLV Cosmology Cluster. KN and YN are grateful for the hospitality of the Institute for the Physics and Mathematics of the Universe (IPMU), University of Tokyo, where part of this work was done.
[^1]: The optical to X-ray afterglow spectral index $\beta_{OX}$ is defined as $f_{\rm opt} / f_X = (\lambda_{\rm opt} / \lambda_X) ^{\beta_{OX}}$, where $f_{\rm opt}$ and $f_X$ are the optical and X-ray fluxes, respectively.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
author:
- Zhilei Liu
- Le Li
- Yunpeng Wu
- 'Cuicui Zhang[^1]'
bibliography:
- 'egbib.bib'
title: Facial Expression Restoration Based on Improved Graph Convolutional Networks
---
[^1]: Corresponding author.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'In several frustrating systems incommensurate behaviors are often observed. For the $S=1$ bilinear-biquadratic model, we show that the main oscillatory behavior, which is proportional to the free edge spins, is eliminated in the incommensurate subphase, considering the average of triplet and singlet energy spectra under open boundary conditions. In the same way, the $\pi$-mode oscillation is also removed in the commensurate subphase. Moreover, we find that higher order corrections are exponentially decaying from an analysis of small size data.'
author:
- Takahiro Murashima and Kiyohide Nomura
bibliography:
- 'hfm2006.bib'
title: |
Cancellation of oscillatory behaviors\
in incommensurate region
---
Introduction
============
Commensurate-incommensurate (C-IC) transitions are interesting phenomena in frustrating quantum spin systems. In the Haldane gap systems, incommensurabilities have often been regarded as troublesome problems and have rarely been discussed in detail until recently. In inelastic neutron scattering experiments, Xu [*et al*]{}[@Xu2000] have revealed that a quasi-one-dimensional oxide, Y$_{2-x}$Ca$_x$BaNiO$_5$, has an incommensurate double-peaked structure factor. However, an analytical interpretation for incommensurabilities has not been clear.
The spin-1 bilinear-biquadratic model, $${\cal H}=\sum_{i=1}^{N} h_i,\qquad
h_i= {{\bf S}}_i\cdot{{\bf S}}_{i+1}
+\alpha({{\bf S}}_i\cdot{{\bf S}}_{i+1})^2,
\label{eq_Ham}$$ which plays a role of prototype of Y$_{2}$BaNiO$_5$, has the C-IC change point which is corresponding to the Affleck-Kennedy-Lieb-Tasaki (AKLT) point $\alpha=\alpha_{\rm D}=1/3$ [@BXG1995; @SJG1996]. The AKLT point is solvable and has an energy gap above the valence-bond-solid (VBS) ground state[@AKLT1987; @AKLT1988]. The VBS state is recently getting more attention with reference to quantum entanglements[@VMDC2004; @FKR2004; @VC2004; @FKRHB2006].
Analyzing the structure factor among the commensurate and incommensurate subphases, and also the C-IC change point, we have deduced two candidates of the ${\it real}$ structure factor unifying these two regions and the AKLT point [@FS2000; @Nomura2003]. Following the S[ø]{}rensen-Affleck prescription[@SA1994], we have constructed Green functions from the candidates, and compared them with the energy gap numerically obtained under the open boundary conditions[@MN2006]. Thus, we have found that the Green function consists of two elements which have anomalies in upper- or lower-half plane.
In this paper we will show that incommensurate oscillatory behaviors in the incommensurate subphase can be canceled, using triplet and singlet energy spectra. In the same way, we can obtain comparable results in the commensurate case. Moreover we discuss higher order corrections considering small size systems.
Short review on singlet-triplet energy gap
==========================================
The gapped Haldane phase has nonvanishing nonlocal string order[@dNR1989] and effectively free $S=1/2$ spins at the ends of open chains [@Kennedy1990; @HKAHR1990]. These edge spins bring a low-lying excitation, triplet, which is degenerate with the ground state in the thermodynamic limit[@Kennedy1990]. While this excitation is clearly a boundary effect, it is linked to the bulk behavior because of the existence of nonlocal string order[@SJG1996].
According to S[ø]{}rensen and Affleck (SA)[@SA1994], the effective Hamiltonian can be connected with the Green function: $${\cal H}_{\rm eff} = (-1)^N {{{\bf S}'}}_{\rm L}\cdot{{{\bf S}'}}_{\rm R}
\lambda^2 \int \frac{dq d\kappa}{(2\pi)^2}G(q, \kappa)
\exp(i q N) \delta(\kappa), \label{eq_Heff}$$ where ${{{\bf S}'}}_{\rm L}$ and ${{{\bf S}'}}_{\rm R}$ are spin-1/2 operators at the ends of chain. Considering the Clebsch-Gordan coefficient, we obtain singlet and triplet expectation values for edge spins as $$\begin{split}
{\langle}{\rm S}| {{{\bf S}'}}_{\rm L}\cdot{{{\bf S}'}}_{\rm R}|{\rm S}{\rangle}=-3/4, \\
{\langle}{\rm T}| {{{\bf S}'}}_{\rm L}\cdot{{{\bf S}'}}_{\rm R}|{\rm T}{\rangle}=1/4,
\label{eq_CG}
\end{split}$$ where $|{\rm T}{\rangle}=|s^{\rm T}=1, s_z=\pm 1,0{\rangle}$ and $|{\rm S}{\rangle}= |s^{\rm T}=0, s_z=0{\rangle}$[@memo1]. Thus the energy difference between the singlet and triplet states can be described with the Green function: $$\begin{split}
\Delta E_{\rm ST} (N) &\equiv E_{\rm T} - E_{\rm S}\\
&=(-1)^N
\lambda^2 \int \frac{dq}{2\pi}G(q, 0)
\exp(i q N),
\end{split}$$ where $E_{\rm T}$ and $E_{\rm S}$ are the triplet and singlet energies, respectively.
![\[fig\_gap\] Oscillation of the energy gap of edge states. The energy gaps with different chain lengths ($N=9, \cdots, 13$) are plotted as a function of $\alpha$. ](fig1.eps){width="8cm"}
Figure \[fig\_gap\] shows the energy gap behaviors of the model with different chain lengths. Increasing the chain length on some fixed $\alpha$ in the commensurate subphase, we see that the energy gap oscillates between even and odd chains and decreases exponentially fast, while this behavior is not simple in the incommensurate subphase. The even-odd oscillation is modulated by the frustration in the incommensurate subphase.
In our previous study[@MN2006], which has been performed to explain these behaviors, we have found $$\Delta E_{\rm ST}
(N)=
\begin{cases}
(-1)^{N}\widetilde{A}\exp(-\widetilde{m}N)\sin(\sqrt{d}N),
&\quad (\alpha > 1/3)\\
0, & \quad (\alpha=1/3) \\
(-1)^{N}\widetilde{A}\exp(-\widetilde{m}N)\sinh(\sqrt{d}N).
&\quad (\alpha < 1/3)\\
\end{cases}
\label{eq_gap}$$ The parameters, $\widetilde{A}$, $\widetilde{m}$, and $d$, depend on $\alpha - \alpha_{\rm D}$ and have been determined with the nonlinear least-squares fitting method as shown in Ref. [@MN2006]. We have found that the C-IC change is characterized by the following Green function: $$G(q,0) \sim \frac{1}{(q - i\widetilde{m})^2 -d}
+ \frac{1}{(q + i\widetilde{m})^2 -d}.$$
Average of triplet-singlet energies
===================================
Now we define the average of triplet-singlet energies: $${\overline{E}_{\rm TS}}\equiv (E_{\rm S}+3E_{\rm T})/4.$$ From Eqs. and , we expect that we can eliminate the principal term caused by the anomalies of the Green function using this average.
![\[fig\_Ets\] Cancellation of triplet-singlet energies. Triplet and singlet energy spectra, and an average of triplet-singlet energies are plotted as a function of $1/N$ for $1\le N \le 15$ when (a) $\alpha=$ 0.3258 (commensurate case), and (b) 0.3425 (incommensurate case). ](fig2a.eps "fig:"){width="8cm"} ![\[fig\_Ets\] Cancellation of triplet-singlet energies. Triplet and singlet energy spectra, and an average of triplet-singlet energies are plotted as a function of $1/N$ for $1\le N \le 15$ when (a) $\alpha=$ 0.3258 (commensurate case), and (b) 0.3425 (incommensurate case). ](fig2b.eps "fig:"){width="8cm"}
Figure \[fig\_Ets\] shows bare triplet and singlet energies and averages of triplet-singlet energies plotted as a function of $1/N$ for (a) $\alpha=$ 0.3258 (commensurate case), and (b) 0.3425 (incommensurate case). We see that the average of triplet-singlet energies on a fixed $\alpha$ varies linearly with $1/N$ not only in the incommensurate subphase but also in the commensurate subphase.
![\[fig\_fitEts\] Finite size effect in an average of triplet-singlet energies and a least-squares fitting line ($c_0+ c_1/N$) are plotted when (a) $\alpha=0.3258$, (b) $0.3425$. ](fig3a.eps "fig:"){width="8cm"} ![\[fig\_fitEts\] Finite size effect in an average of triplet-singlet energies and a least-squares fitting line ($c_0+ c_1/N$) are plotted when (a) $\alpha=0.3258$, (b) $0.3425$. ](fig3b.eps "fig:"){width="8cm"}
We perform the least-squares fit to the averages of triplet-singlet energies using the following fitting function: $$f(N) = c_0 + c_1 /N. \label{eq_fit}$$ Figure \[fig\_fitEts\] shows the data of triplet-singlet averages and $f(N)$. When (a) $\alpha=0.3258$, each coefficients are obtained as $c_0=-0.68167581 \pm 2 \times 10^{-9}$ and $c_1=-1.1671 \times 10^{-4} \pm 2 \times 10^{-8}$ for $8 \leq N \leq 15$. In the same way, $c_0=-0.6483465831 \pm 2 \times 10^{-10}$ and $c_1=-1.77295 \times 10^{-4} \pm 2 \times 10^{-9}$ when (b) $\alpha=0.3425$. The average of triplet-singlet energies ${\overline{E}_{\rm TS}}/N$ seems to behave high linearly in small $\alpha$ region. In fact, we observe that coefficients of $O(1/N^2)$, $O(1/N^3)$ are very small. Hence, we can say that a higher power of $1/N$ does not appear.
![\[fig\_fit\_c\] Coefficients (a) $c_0$ and (b) $c_1$ in Eq. are plotted near the AKLT point. It appears that $c_0 \propto \alpha -\alpha_{\rm D}$ and $c_1 \propto
(\alpha -\alpha_{\rm D})^2$.](fig4a.eps "fig:"){width="8cm"} ![\[fig\_fit\_c\] Coefficients (a) $c_0$ and (b) $c_1$ in Eq. are plotted near the AKLT point. It appears that $c_0 \propto \alpha -\alpha_{\rm D}$ and $c_1 \propto
(\alpha -\alpha_{\rm D})^2$.](fig4b.eps "fig:"){width="8cm"}
Figure \[fig\_fit\_c\] shows coefficients $c_0$ and $c_1$ in Eq. obtained with the least-squares fitting. The second term in Eq. comes from surface effects. Since the surface effects are caused by the one-dimensionality, they are different from the effect of edge spins. Additionally, the surface effects are perfectly zero, namely $c_1=0$, at the AKLT point[@memo2]. Moreover, $c_1$ always shows negative value except the AKLT point. We see that the fitting parameters $c_0$ and $c_1$ behave approximately as $\alpha - \alpha_{\rm D}$ and $(\alpha-\alpha_{\rm D})^2$, respectively.
So far we have excluded data smaller than $N=8$ since these data differ from Eq. . Then we proceed to study small size corrections from Eq. . For a rough estimate, we consider the difference $$\Delta y (N)= {\overline{E}_{\rm TS}}(N)/N - (c'_0 + c'_1/N) \label{eq_dy}$$ for small $N$, although we use $c'_0$ and $c'_1$, which are determined from the data of $N=14,15$. Figure \[fig\_dy\] (a) shows $\log (|\Delta y|)$ for $1 \le N \le 8$ at $\alpha=0.3258$ in the commensurate region. Since the logarithm of $\Delta y$ decreases linearly with $N$, the difference $\Delta y$ results in $\Delta y \sim \exp(-N/\xi')$. We estimate $\xi'$ as shown in Fig. \[fig\_dy\] (b), and then we find $\xi'\sim \xi/2$ (The correlation length $\xi$ is obtained in the previous study[@MN2006])[@memo3].
 The logarithm of $|\Delta y|$ (Eq. ) are plotted against $N$. The dotted line is drawn for a guide to the eye. (b) Two different correlation lengths and the ratio $\xi'/\xi$ are plotted in the commensurate region.](fig5a.eps "fig:"){width="8cm"} The logarithm of $|\Delta y|$ (Eq. ) are plotted against $N$. The dotted line is drawn for a guide to the eye. (b) Two different correlation lengths and the ratio $\xi'/\xi$ are plotted in the commensurate region.](fig5b.eps "fig:"){width="8cm"}
In the incommensurate region, we see the oscillatory behavior again. We roughly estimate the wave number of this oscillation, and then we find $q'_{\rm IC} \sim 2(q_{\rm
IC}+\pi)$[@memo4]. We will show a detailed calculation in another paper.
Summary
=======
We have shown that the principal oscillatory behaviors, which are proportional to ${{{\bf S}'}}_{\rm L}\cdot{{{\bf S}'}}_{\rm R}$, among the triplet and singlet energy spectra under open boundary conditions cancel out in the commensurate and incommensurate subphases.
We have found that the energy spectra of singlet and triplet states under open boundary conditions consist of the bulk, surface, and edge spin energies: $$E_{\rm C}/N = B + S/N +
{\langle}{\rm C}|{\cal H}_{\rm eff}|{\rm C}{\rangle},
\qquad ({\rm C=\{S, T\}})$$ where $B$ and $S$ are the bulk and surface energies, respectively. Considering small $N$, we have found an exponentially decaying correction term, the correlation length and the incommensurate wave number of which differ by factor 2 from those obtained by the energy gap of edge states. One possibility is that the correction will be $O(({{{\bf S}'}}_{\rm L}\cdot{{{\bf S}'}}_{\rm R})^2)$. Therefore, we will need to improve the SA theory so as to contain such higher terms.
We observe similar results for the $S=1$ next-nearest-neighbor model. Thus the cancellation of triplet-singlet energies can be found in general spin gap systems.
The numerical calculation in this work is based on the program packages TITPACK version 2, developed by Professor H. Nishimori.
This research is partially supported by a Grant-in-Aid for Scientific Research (C), 18540376 (2006), from the Ministry of Education, Science, Sports and Culture of Japan.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'We compare two different approaches for quantization of the Bianchi I model: a reduced phase space quantization, in which the isotropic Misner variables is taken as time, and the Vilenkin proposal, in which a semiclassical approximation is performed for the same variable. We outline the technical and interpretative issues of these two methods and we demonstrate that they provide equivalent results only if the dynamics is essentially dictated by the isotropic matter contribution.'
author:
- 'Leonardo Agostini$^{1}$, Francesco Cianfrani$^{2}$, Giovanni Montani$^{1,3}$'
title: Probabilistic interpretation of the wave function for the Bianchi I model
---
\[sec:introduction\]Introduction
================================
The Wheeler-Dewitt equation [@DeWitt:1967yk; @DeWitt:1967ub; @DeWitt:1967uc], corresponding to the canonical quantization of the gravitational field in the metric approach [@Kuchar1981], is associated to a functional formalism and it becomes a viable theory only in Minisuperspace, where the symmetry restriction reduces the dynamical problem to a finite number of degrees of freedom. The minisuperspace model [@primordial; @cosmology; @book] is a natural arena for the study of Quantum Cosmology, since, as soon as we consider homogeneous Universes (Bianchi models) [@Landau2], the corresponding wave function is taken over a finite number of degrees of freedom, typically the three cosmic scale factors (or some functions of them), determining the evolution of independent spatial directions.
However, the cosmological implementation of the Wheeler-DeWitt equation solves the question concerning the viability of the canonical formalism, but it does not address all the other shortcomings of this approach, like the absence of a dynamical Hilbert space as a general feature and then the issue concerning the predictivity of the considered quantum theory [@primordial; @cosmology; @Blyth:1975is]. Describing the Bianchi models via Misner variables [@Misner:1969hg; @Misner1971], [*i.e.*]{} separating the isotropic volume component of the Universe from the corresponding anisotropy degrees of freedom, we reach a very meaningful representation of the Bianchi model quantum dynamics: a Klein-Gordon equation in the presence of a potential term in the coordinates. Indeed, the minisuper-metric is pseudo-Riemannian and the isotropic Misner variable plays the role of a time-like coordinate, while the anisotropies are space-like variables. What makes puzzling the physical characterization of a generic Bianchi model (especially the type VIII and IX, which are the most general ones allowed by the homogeneity constraint) is the impossibility to separate positive and negative frequency solutions, which is a key-point of the original quantum physics prescription to get a positive defined probability density. This impossibility is due to the time dependence (dependence on the isotropic Misner variable) of the Bianchi model potential (except for the Bianchi type I, where it vanishes), so that the physical interpretation of the solution to the cosmological implementation of the Wheeler-DeWitt equation is de facto forbidden. Actually, such difficulty in constructing a dynamical Hilbert space relies on a more general feature of canonical quantum gravity: the Wheeler-deWitt equation is an equation of Klein-Gordon type, so that it has a conserved current that does not lead to a positive definite inner product. Two different approaches can be pursued to give a meaningful interpretation of the wave function of the Universe in terms of a probabilistic theory: *i*) we can classically solve the Hamiltonian constraint and then quantize the resulting Schrödinger equation [@Misner:1969ae; @Benini:2006xu] (reduced phase space quantization [@Henneaux1992], RPSQ); *ii*) we assume, according to the Vilenkin proposal [@Vilenkin1989], that the isotropic Misner variable approaches a quasi-classical limit, while the anisotropies remain pure quantum degrees of freedom. Even this latter approach leads to a Schrödinger equation, but, as we will see, it avoids the square root non-local Hamiltonian operator emerging from the former. What makes comparable the two approaches is actually the role of the time-like variable, say the internal time of the theory, played by the isotropic Misner variable in the corresponding Schrödinger dynamics.
However, despite its non-local character, the paradigm based on RPSQ is an exact procedure, requiring no WKB approximation on the Universe wave function, which on the contrary is a necessary step in the Vilenkin proposal.
Thus, limiting our attention to the simple case of a Bianchi I cosmology in presence of matter (a time-dependent term), we here address a rigorous comparison of the two quantization methods, in order to determine under which restrictions the Vilenkin adiabatic WKB representation of the Universe volume dynamics becomes predictive.
To this end, we carefully construct and then compare the wave functions of the Bianchi I model in the two cases, so clarifying which restrictions provide equal probabilities on the anisotropy variables. The possibility for a comparison requires that the wave function evolution be described via the same time parameter, [*i.e.*]{} in correspondence to the same functional form of the lapse functions. It is just this overlap of the time variables to impose the most restrictive condition for the validity of the Vilenkin representation, [*i.e.*]{} the spectra of anisotropy momenta must extend over small values with respect to the time-dependent matter term, which is necessary for the quasi-classical limit of the isotropic Misner variable. This restriction is equivalent to say that the adiabatic WKB Vilenkin approach overlaps the quantum dynamics in RPSQ (also dubbed Arnowitt-Deser-Misner reduction of the Hamiltonian problem [@Arnowitt:1962hi]), only if the anisotropy degrees of freedom are small and then the Universe is mainly isotropic. By other words, we fix the impossibility to apply the Vilenkin quantum evolution of the Universe anisotropies near the cosmological singularity, where their values and momenta are arbitrarily large.
Nonetheless, once this strong restriction is fulfilled, we demonstrate how the probabilities (taken on the same domain) for the Universe anisotropies coincide in the two approaches, suggesting that for a quasi-isotropic Universe a robust probabilistic interpretation exists. We conclude observing how, in the absence of a clear mechanism for frequency separation, the procedure of solving the Hamiltonian constraint and then canonically quantizing it remains the only general attempt to the minisuperspace quantum dynamics, although it suffers the aforementioned non-trivial question concerning locality.
Action for Bianchi models
=========================
Let us consider *homogeneous* but not necessarily *isotropic* spacetimes. Following the definition given in [@Wald1984], spatially homogenous spacetimes are those that can be foliated with a family of three-dimensional hypersurfaces $\Sigma_t$ and there exists an isometry of the three-metric $h_{ij}$ that connects any two points of $\Sigma_t$. For such spacetimes, the three-metric can be written as [@Landau2; @primordial; @cosmology]
$$\label{eq:three_metric_homogeneous}
h_{ij}(x^k, t) = h_{ab}(t) e^{(a)}_i(x^k) e^{(b)}_j(x^k),$$
where the vectors $e^{(a)}_i$ constitutes the so-called *frame* and they do not depend on time. Homogeneous spaces have been classified by Bianchi in [@Bianchi1898]. One can write $h_{ab} = (e^{2 \alpha} e^{2 \vb*{\beta}})_{ab}$ where $\beta_{ab}$ is a $3 \times 3$ traceless, symmetric matrix and $\alpha$ is called the isotropic variable, both depending on $t$ only.
It is generically assumed that the supermomentum constraint identically vanishes and one can safely fix $N^i=0$ (see [@Moniz2010] for more details). Furthermore, $\beta$ can be diagonalized and the eigenvalues can be parametrized as follows
$$\label{eq:beta_matrix_diagonal}
\beta_{ab} = diag( {\beta_{+}}+ \sqrt{3} {\beta_{-}}, {\beta_{+}}- \sqrt{3} {\beta_{-}}, - 2 {\beta_{+}}).$$
Making a canonical transformation to $\alpha, {\beta_{+}}, {\beta_{-}}$, their conjugate momenta ${\mathpzc}{p}_{\alpha}, {\mathpzc}{p}_+, {\mathpzc}{p}_-$ are given by $$\begin{aligned}
&{\mathpzc{p}}_{\alpha}=- \frac{3 c^2 \kappa e^{3 \alpha}}{4 \pi G N}\,\dot{\alpha} \\
&{\mathpzc{p}}_+ = \frac{3 c^2 \kappa e^{3 \alpha}}{4 \pi G N}\,\dot{\beta}_+ \\
&{\mathpzc{p}}_-= \frac{3 c^2 \kappa e^{3 \alpha}}{4 \pi G N}\,\dot{\beta}_- \,.\end{aligned}$$
and the Einstein-Hilbert action becomes
$$\label{eq:action_alpha_beta}
\begin{aligned}
S_{EH}=& \int \dd t \Bigg\{ {\mathpzc}{p}_{\alpha} \dot{\alpha} + {\mathpzc}{p}_+ \dot{\beta}_+ + {\mathpzc}{p}_- \dot{\beta}_- - \frac{2 \pi G N e^{-3 \alpha}}{3 c^2 \kappa} \\
&\qty[ - {\mathpzc}{p}_{\alpha}^2 + {\mathpzc}{p}^2_+ + {\mathpzc}{p}^2_- - \frac{3 c^2 \kappa^2 e^{6 \alpha}}{32 \pi^2 G^2} {\tensor}[^{(3)}]{R}{}] \Bigg\}
\end{aligned}$$
where $\kappa = \int \dd^3 x \abs{\det(e^{(a)}_i(x^k))}$. Let us add as matter terms some perfect fluids with pressure $P$ and energy density $\rho$ having the equation of states $P=w\,\rho$. The total action reads as
$$\label{eq:action_alpha_beta_+_matter}
\begin{aligned}
S=& \int \dd t \Bigg\{ {\mathpzc}{p}_{\alpha} \dot{\alpha} + {\mathpzc}{p}_+ \dot{\beta}_+ + {\mathpzc}{p}_- \dot{\beta}_- - \frac{2 \pi G N e^{-3 \alpha}}{3 c^2 \kappa} \\
&\qty[ - {\mathpzc}{p}_{\alpha}^2 + {\mathpzc}{p}^2_+ + {\mathpzc}{p}^2_- - \frac{3 c^2 \kappa^2 e^{6 \alpha}}{32 \pi^2 G^2} {\tensor}[^{(3)}]{R}{} + \mu^2(\alpha)] \Bigg\}
\end{aligned}$$
where the matter contribution is encoded in the term $\mu^2$, reading, according to [@Ryan1975],
$$\label{eq:matter_term}
\begin{aligned}
\mu^2(\alpha)= \sum_w \mu^2_w e^{3(1-w)\alpha},
\end{aligned}$$
and the sum extends over all the fluid components, characterized by different values of $w$, while $\mu^2_w$ are constants.
Let us now consider the easiest case of the Bianchi I model, for which ${\tensor}[^{(3)}]{R}{} = 0$ and we obtain:
$$\label{eq:action_for_bianchi_I}
\begin{aligned}
S=& \int \dd t \Bigg\{ {\mathpzc}{p}_{\alpha} \dot{\alpha} + {\mathpzc}{p}_+ \dot{\beta}_+ + {\mathpzc}{p}_- \dot{\beta}_- - \frac{2 \pi G N e^{-3 \alpha}}{3 c^2 \kappa} \\
&\qty[ - {\mathpzc}{p}_{\alpha}^2 + {\mathpzc}{p}^2_+ + {\mathpzc}{p}^2_- + \mu^2(\alpha)] \Bigg\}.
\end{aligned}$$
Notice that Eq. resembles the action of a relativistic particle moving in a $(\alpha, {\beta_{+}}, {\beta_{-}})$ space with a variable mass. Moreover, it is worth saying that $\kappa$ can be made finite even if we are considering a flat space, closing it with a torus topology or, due to homogeneity, just considering a finite portion of space. From Eq. we can find the classical equations of motion:
\[eq:classical\_equation\_of\_motion\] $$\begin{aligned}
&- \dv{\mu^2(\alpha)}{\alpha} - \frac{3 c^2 \kappa}{2 \pi G N} e^{3 \alpha} \dot{{\mathpzc{p}}}_{\alpha}=0; \label{eq:motion_for_p_alpha}\\
&\dot{{\mathpzc{p}}}_+ = 0, \quad \dot{{\mathpzc{p}}}_- = 0; \label{eq:motion_for_p_+_-}\\
&\dot{\alpha} = - \frac{4 \pi G N e^{-3 \alpha}}{3 c^2 \kappa} {\mathpzc{p}}_{\alpha}; \label{eq:motion_for_alpha}\\
&\dot{\beta}_+ = \frac{4 \pi G N e^{-3 \alpha}}{3 c^2 \kappa} {\mathpzc{p}}_+; \label{eq:motion_for_bp}\\
&\dot{\beta}_- = \frac{4 \pi G N e^{-3\alpha}}{3 c^2 \kappa} {\mathpzc{p}}_-; \label{eq:motion_for_bim}\\
&\mathcal{H} = (- {\mathpzc{p}}^2_{\alpha} + {\mathpzc{p}}_+^2 + {\mathpzc{p}}_-^2) + \mu^2(\alpha)=0. \label{eq:first_class_constraint}\end{aligned}$$
It is worth noticing how momenta ${\mathpzc{p}}_{+}$ and ${\mathpzc{p}}_{-}$ are constants of motion.
RPSQ of Bianchi I
=================
The idea of RPSQ is to reduce the phase-space to the only physical d.o.f. and to choose a time variable *before* quantising the theory. It is worth noticing that we are dealing with $\alpha, {\beta_{+}}$ and ${\beta_{-}}$ and then one variable is to some extent redundant. This is due to the fact that we still have the gauge freedom of the lapse function form, which can be fixed in order to set $\alpha$ as a time-like variable. Similarly to [@Misner1971], we here make the gauge choice $\alpha = t/T := {\mathpzc}{t}$, providing that $\dot{\alpha} > 0$ in order to select the expanding branch of the Universe. This implies ${\mathpzc{p}}_{\alpha} < 0$ since $N$ is positive defined. The constant $T$ can be defined in terms of fundamental constants (in which case it can be chosen proportional to the Planck length), or it can be derived from the specific data of the considered problem (as soon as a proper matter or energy scale is given from the matter fields configuration). The lapse function is then fixed: from we get
$$\label{eq:lapse_function_adm}
N_{RPSQ} ({\mathpzc}{t}) = - \frac{3 c^2 \kappa}{4 \pi G T}\,\frac{e^{3 {\mathpzc}{t}}}{{\mathpzc}{p}_{\alpha}}$$
In order to reach a fully reduced phase-space, we classically solve the first class constraint obtaining
$$\label{eq:solving_firt_class_constraint}
{\mathpzc}{p}_{\alpha} = -\sqrt{{\mathpzc}{p}_+^2 + {\mathpzc}{p}_-^2 + \mu^2(\alpha)}.$$
Then we substitute and into obtaining:
$$\label{eq:rpsq_action}
S_{RPS} = \int \dd {\mathpzc}{t} \qty[ {\mathpzc}{p}_+ \dv{{\beta_{+}}}{{\mathpzc}{t}} + {\mathpzc}{p}_- \dv{{\beta_{-}}}{{\mathpzc}{t}} - \sqrt{{\mathpzc}{p}_+^2 + {\mathpzc}{p}_-^2 + \mu^2({\mathpzc}{t})} ]$$
We have reached a form for the action in which we have only the two physical d.o.f. and a nonlocal Hamiltonian, corresponding to a square root operator. We can therefore proceed with the canonical quantisation by introducing the wave function $\Phi({\beta_{+}},{\beta_{-}},{\mathpzc}{t})=\Phi({\underline}{\beta},{\mathpzc}{t})$ and promoting momenta to derivative operators as follows
$$\label{eq:canonical_quantisation_rps}
{\mathpzc}{p}_+ \rightarrow - i \hbar \pdv{{\beta_{+}}}, \quad {\mathpzc}{p}_- \rightarrow - i \hbar \pdv{{\beta_{-}}}\,,$$
and then we get the following Schr[[$\ddot{\mathrm{o}}$]{}]{}dinger equation:
$$\label{eq:schrodinger_equation_rps}
i \hbar \pdv{{\mathpzc}{t}} \Phi({\underline}{\beta},{\mathpzc}{t}) = \sqrt{- \hbar^2 \qty(\pdv[2]{{\beta_{+}}} + \pdv[2]{{\beta_{-}}}) + \mu^2({\mathpzc}{t})}\, \Phi({\underline}{\beta},{\mathpzc}{t})\,.$$
The equation above resembles the spinless Salpeter equation [@Lammerzahl1993; @Schweber1961], except for the presence of the time-dependent mass term $\mu^2({\mathpzc}{t})$. It is worth noting that for $w\leq 1$ the function $\mu^2({\mathpzc}{t})$ is bounded close to the singularity, corresponding to $\alpha\rightarrow -\infty$, and we denote by $\mu^2_1$ the corresponding value, [*i.e.*]{} $\lim_{{\mathpzc}{t} \to - \infty} \mu^2(t) = \mu_1^2$. Hence, close to the singularity the Schr[[$\ddot{\mathrm{o}}$]{}]{}dinger equation coincides with the Salpeter equation with mass term $\mu_1^2$. Generically, the operator under square root on the right-hand side of is *strongly elliptic* (see [@Lammerzahl1993]), therefore the square root operator is a *pseudo-differential operator*, which can be conveniently analyzed in the base of plane waves $\phi_{{\underline}{p}}({\underline}{\beta}) = A({\underline}{p}) e^{\frac{i}{\hbar} {\underline}{p} \cdot {\underline}{\beta}}$ finding
$$\label{eq:plane_wave_solution}
\sqrt{- \hbar^2 \Delta_{\pm} + \mu^2({\mathpzc}{t})} \phi_{{\underline}{p}}({\underline}{\beta}) = \sqrt{{\underline}{p}^2 + \mu^2({\mathpzc}{t})} \phi_{{\underline}{p}}({\underline}{\beta}),$$
where ${\underline}{p}^2 = p_+^2 + p_-^2$. Hence, the evolution of plane waves is described by
$$\label{eq:evolution_of_phi_transf}
\tilde{\Phi}({\underline}{p},{\mathpzc}{t})= e^{- \frac{i}{\hbar} \int_{{\mathpzc}{t}_0}^{{\mathpzc}{t}} \dd {\mathpzc}{t}' \sqrt{\abs{{\underline}{p}}^2 + \mu^2({\mathpzc}{t})}} \tilde{\Phi}({\underline}{p},{\mathpzc}{t}_0)\,,$$
and via inverse Fourier transform a generic solution can be formally written as
$$\label{eq:evolution_of_phi}
\Phi({\underline}{\beta},{\mathpzc}{t})= e^{- \frac{i}{\hbar} \int_{{\mathpzc}{t}_0}^{{\mathpzc}{t}} \dd {\mathpzc}{t}' \sqrt{-\hbar^2 \Delta_{\pm}^2 + \mu^2({\mathpzc}{t})}} \Phi({\underline}{\beta},{\mathpzc}{t}_0).$$
The main issue of this formulation is the definition of a suitable inner product. A Klein-Gordon-like inner product is an attractive possibility, given the results in [@Schweber1961], but here it would be intrinsically time-dependent, because of the presence of the time-dependent mass term $\mu^2(t)$. Hence, we propose to specify the inner product near the singularity (${\mathpzc}{t} \rightarrow - \infty$), where $\mu^2\rightarrow \mu_1^2$, and to define it generically in a time-independent way, according to [@Mostafazadeh:2002xa], so finding
$$\label{eq:inner_product_sqrt_kg_infty}
\begin{aligned}
\rbraket{\psi}{\phi}({\mathpzc}{t}) =& \int \dd^2 \beta \bigg[ \psi^*({\underline}{\beta},{\mathpzc}{t}) \,\left(\sqrt{-\hbar^2 \Delta_{\pm} + \mu^2_1}\, \phi({\underline}{\beta},{\mathpzc}{t})\right) +\\
&+ \left(\sqrt{-\hbar^2 \Delta_{\pm} + \mu^2_1}\, \psi^*({\underline}{\beta},{\mathpzc}{t})\right) \,\phi({\underline}{\beta},{\mathpzc}{t})\bigg].
\end{aligned}$$
It can be shown that the evolution operator in is *unitary* with respect to such an inner product. This also implies that, if we normalize the wave function in the limit ${\mathpzc}{t} \rightarrow -\infty$, the normalization is preserved for any ${\mathpzc}{t}$.
Semiclassical approximation
===========================
We now adopt the Dirac quantisation method in order to obtain a quantum theory describing the Bianchi I cosmological model, [*i.e.*]{}, instead of classically solving the super-Hamiltonian constraint, we quantise it requiring that it annihilates physical state $\ket{\Psi}$
$$\label{eq:constraint_operator_on_physical_states}
\vu*{\mathcal{H}} \ket{\Psi} = 0.$$
The canonical quantisation procedure can be implemented similarly to the RPSQ case , but in addition we also have to implement $\alpha$ as a multiplicative operator and ${\mathpzc{p}}_{\alpha} \rightarrow -i \hbar \pdv{\alpha}$. Hence, the condition becomes
$$\label{eq:equation_for_psi}
\qty[ \hbar^2 \qty( - \pdv[2]{\alpha} + \pdv[2]{{\beta_{+}}} + \pdv[2]{{\beta_{-}}}) + \mu^2(\alpha)] \Psi({\underline}{\beta},\alpha)=0.$$
Since $\vu*{\mathcal{H}}$ is the generator of gauge transformations, ensures that on a quantum level physical states do not change under gauge transformations, [*i.e.*]{}, the wavefunction does not depend on the label time $t$. Therefore, the time is hidden between the internal variables $\alpha, {\beta_{+}}, {\beta_{-}}$. Focusing on the probabilistic interpretation for the wave function, De Witt [@DeWitt:1967yk] first suggested to define a scalar product from a conserved current $j^{\mu}$, analogously to the Klein-Gordon case. Although a Klein-Gordon-like current can be defined, the corresponding scalar product is not positive defined and no separation of frequencies can be consistently implemented (this is due to a divergence occurring in superspace when the metric determinant vanishes, see for instance [@book]).
A different approach has been proposed by Vilenkin [@Vilenkin1989], who suggested that a proper probabilistic interpretation can be achieved only after the emergence of time via a semiclassical approximation on the wave function $\Psi$. Making the ansatz $$\label{eq:ansatz_semiclassical}
\Psi({\underline}{\beta},\alpha) = A(\alpha) e^{\frac{i}{\hbar} S(\alpha)} \chi({\underline}{\beta}, \alpha),$$ it can be shown that a positive defined probability exists for $\chi$ if the congruence of classical trajectories ([*i.e.*]{} the gradient curves of $S_0$) crosses once and only once the hypersurfaces of constant timelike variable. In particular, $\alpha$ is the time-like variable and the contour lines of $S(\alpha)$ are equal-time hypersurfaces. Inserting the wave function into , a solution is given for
\[eq:eqs\_vilenkin\] $$\begin{aligned}
&-\qty( \dv{S(\alpha)}{\alpha})^2 + \mu^2(\alpha) = 0; \label{eq:eq_for_s_alpha}\\
&\dv{\alpha} \qty[ A^2 (\alpha) \qty(- \dv{S(\alpha)}{\alpha})]=0; \label{eq:classical_continuity_eq} \\
&- i \hbar \dv{S(\alpha)}{\alpha} \pdv{\chi({\underline}{\beta},\alpha)}{\alpha} = - \frac{\hbar^2}{2} \qty( \pdv[2]{{\beta_{+}}} + \pdv[2]{{\beta_{-}}}) \chi({\underline}{\beta},\alpha). \label{eq:approximated_eq_chi}
\end{aligned}$$
In particular, from and one gets
\[eq:solution\] $$\begin{aligned}
&S(\alpha) = \pm \int_{\alpha_0}^{\alpha} \dd \alpha' \sqrt{\mu^2(\alpha')} \\
&A(\alpha) = \frac{\mathcal{K}^2}{\sqrt[4]{\mu^2(\alpha)}}. \end{aligned}$$
where $\mathcal{K}^2$ is an arbitrary integration constant. By introducing the WKB time $t_{WKB}$, for which $\dv{S(\alpha)}{\alpha} \pdv{\alpha} = \pdv{t_{WKB}}$ [@Kiefer1987], becomes the Schr[[$\ddot{\mathrm{o}}$]{}]{}dinger equation describing the motion of a free particle in the $({\beta_{+}},{\beta_{-}})$ plane. Therefore, $\chi$ can be interpreted as the wave function describing a free particle in superspace, in which $\alpha$ is a parametric time-like variable.
By solving the eigenvalue problem for , redefining $\chi({\underline}{\beta},\alpha)$ in order to set $\mathcal{K}^2=1$ and normalizing $\chi({\underline}{\beta},\alpha)$ with respect to the natural scalar product for a free particle, [*i.e.*]{}
$$\label{eq:inner_product_vlienkin}
\braket{\chi(\alpha)}{\chi(\alpha)} = \int \dd^2 \beta | \chi({\underline}{\beta},\alpha) |^2=1,$$
we get the following result
$$\label{eq:final_result_vilenkin}
\begin{aligned}
\Psi(\alpha,{\underline}{\beta}) =& \frac{e^{- \frac{i}{\hbar} \int^{\alpha}_{\alpha_0} \dd \alpha' \sqrt{\mu^2(\alpha)}}}{\sqrt[4]{\mu^2(\alpha)}}\\
& \int_{\mathbb{R}^2} \frac{\dd^2 p}{2 \pi \hbar} e^{- \frac{i}{2\hbar} (p^2_+ + p^2_-) \int^{\alpha}_{\alpha_0} \dd \alpha' \frac{1}{\sqrt{\mu^2(\alpha)} }} \cdot \\
&\cdot e^{\frac{i}{\hbar} {\underline}{p} \cdot {\underline}{\beta}} \tilde{\chi}(\alpha_0, {\underline}{p}).\\
\end{aligned}$$
where the function $\tilde{\chi}(\alpha_0, {\underline}{p})$ determines initial conditions. It is worth saying that two hypotheses lie behind Vilenkin’s semiclassical approximation: a WKB approximation and a Born-Oppenheimer approximation, in which $\alpha$ is the slow variable whereas ${\beta_{+}}$ and ${\beta_{-}}$ are the fast ones. The validity of these approximations implies that
1. we admit a decomposition of the wave function as $\Psi(\alpha,{\underline}{\beta})=e^{\frac{i}{\hbar}\sum_{n=0} (\hbar)^n\,S_{(n)}}$ and the following conditions hold
\[eq\_bianchiI:semiclassical\_conditions\_final\] $$\begin{aligned}
&\bullet \ \abs{\frac{1}{\mu^{3}(\alpha)} \dv{\mu^2(\alpha)}{\alpha}} \ll \frac{4}{\hbar}; \label{eq_bianchiI:first_semiclassical_condition_final}\\
&\bullet \ \hbar \abs{S_2(\alpha)} \ll |S_1(\alpha)| \quad \mbox{and} \quad \hbar \abs{S_2(\alpha)} \ll 1. \label{eq_bianchiI:second_semiclassical_condition_final}\end{aligned}$$
2. the integral over $p$s in extends over those values for which
\[eq\_bianchiI:Born-Oppenheimer\_condition\_final\] $$\begin{aligned}
&\bullet \ (p_+^2+p_-^2) \ne 0; \label{eq_bianchiI:Born-Oppenheimer_non_zero}\\
&\bullet \ (p_+^2 + p_-^2) \ll \mu^2 (\alpha). \label{eq_bianchiI:Born-Oppenheimer_lesser_than_h_m}\end{aligned}$$
From it can be shown that there always exists $\alpha_S$ such that for $\alpha>\alpha_S$ the WKB approximation is satisfied, whereas the BO approximation implies that we need a wave packet $\tilde{\chi}(\alpha_0, {\underline}{p})$ sufficiently peaked near an *ad hoc* value $\bar{{\underline}{p}}=(\bar{{\underline}{p}}_+,\bar{{\underline}{p}}_-)$ for which holds. For simplicity, we choose a Gaussian wave packet. It is worth noting that since the Hamiltonian operator is a function of momentum operators only, no wave function spread occurs in the $p$-representation. This ensures that condition is satisfied for every $\alpha$ if it is satisfied at a certain $\alpha_0$. Thus we get the following final expression for the wave function in semiclassical approximation:
$$\label{eqn_bianchiI:final_wfv_fin}
\begin{aligned}
\Psi(\alpha,{\underline}{\beta}) =& \frac{e^{- \frac{i}{\hbar} \int^{\alpha} \dd \alpha' \sqrt{\mu^2(\alpha)}}}{\sqrt[4]{\mu^2(\alpha)}} \\
& \int_{\mathbb{R}^2} \frac{\dd^2 p}{2 \pi \hbar} \frac{e^{\frac{i}{2\hbar} (p^2_+ + p^2_-) \int^{\alpha} \dd \alpha' \frac{1}{\sqrt{\mu^2(\alpha)} }}}{\sqrt{\pi \sigma_+ \sigma_-}} \cdot \\
&\cdot e^{-\frac{i}{\hbar} {\underline}{p} \cdot ({\underline}{{\mathpzc}{b}}-{\underline}{\beta})} e^{- \frac{(p_+ - \bar{{\underline}{p}}_+)^2}{2 \sigma^2_+}} e^{- \frac{(p_- - \bar{{\underline}{p}}_-)^2}{2 \sigma^2_-}}\,,\\
\end{aligned}$$
where ${\underline}{{\mathpzc}{b}}=({\underline}{{\mathpzc}{b}}_+,{\underline}{{\mathpzc}{b}}_-)$ denotes the point in the $({\beta_{+}},{\beta_{-}})$ plane around which the wave function is initially peaked. For the following analysis, we write here the explicit expression of the lapse function $N_S$, that comes out from the fact that $p_{\alpha} = \dv{S(\alpha)}{\alpha}$ and from equation , namely
$$\label{eq:lapse_semiclassical}
N_S(\alpha) = \frac{3 c \kappa \dot{\alpha} e^{3 \alpha}}{4 \pi G} \frac{1}{\sqrt{\mu^2(\alpha)}}.$$
Comparison
==========
In order to compare the two formulations, we first need to identify the same time variable. It can be done requesting that $\alpha=t/T = {\mathpzc}{t}$ holds also for the semiclassical approximation and asking for the same label time $t$ in the two formulations, *i.e.*, we need the two lapse function and to be the same $$\label{eq_bI:comparison_lapse_w_1_approx}
\frac{3 c \kappa}{4 \pi G T} \frac{e^{3 {\mathpzc}{t}}}{\sqrt{\mathpzc{p}^2_+ + \mathpzc{p}^2_- + \mu^2({\mathpzc}{t})}} = \frac{3 c \kappa}{4 \pi G T} \frac{e^{3 {\mathpzc}{t}}}{\sqrt{\mu^2({\mathpzc}{t})}},$$ This equation is essentially valid only if ${\mathpzc}{p}_+^2 + {\mathpzc}{p}_-^2 \ll \mathcal \mu^2(t)$, but this request coincides with . We now observe that once we promote ${\beta_{+}}$ and ${\beta_{-}}$ to quantum operators, the lapse function $N_{RPSQ}$ becomes itself an operator acting on the wave function [@Arnowitt:1962hi; @Misner:1969ae]. Thus, in order to perform the comparison above, we have to replace it by its expectation value on the wave function, essentially corresponding to its expression calculated on the classical momenta $\bar{{\underline}{p}}$ (this requires, as addressed below, that the wave function is sufficiently peaked around the classical trajectories). However, it is worth noting that for the case of Bianchi I $\bar{{\underline}{p}}$ are classical constants of motion and therefore we can think of them as fixed numbers in the expression of the RPSQ lapse function. Around these fixed values, we can then peak the Universe wave function, in order to perform the comparison .
Our aim is to check whether the functional form of the wave functions obtained in the two considered cases, or their associated probabilities, coincide. In doing this matching, we need to take a range of ${\mathpzc}{t}$ such that the semiclassical approximation is satisfied. Calling ${\mathpzc}{t}_S$ such time, we also impose that ${\mathpzc}{t}_0 \ge {\mathpzc}{t}_S$ and we consider initial conditions providing the same probability to find the Universe in a point $({\underline}{{\mathpzc}{b}},{\underline}{{\mathpzc}{p}})$ in phase space at ${\mathpzc}{t}_0$ in both methods. Hence, by normalizing the wave function with respect to the inner product we get:
$$\label{eq_bianchiI:adm_evolved_gaussian_wave_function}
\begin{aligned}
\Phi(t,{\underline}{\beta}) =& \int_{\mathbb{R}^2} \frac{\dd^2 p}{ (2 \pi \hbar) \sqrt{2 \pi \sigma_+ \sigma_-}} \frac{e^{-\frac{i}{\hbar} \int_{\bar{{\mathpzc}{t}}}^{{\mathpzc}{t}} \dd {\mathpzc}{t}' \sqrt{\mu^2({\mathpzc}{t}') + p_+^2 + p_-^2}}}{\sqrt[4]{\mu_1^2 + p_+^2 + p_-^2}} \\
&e^{\frac{i}{\hbar} {\underline}{p} \cdot ({\underline}{\beta}-{\underline}{{\mathpzc}{b}})} e^{- \frac{({\mathpzc}{p}_+ - p_+)^2}{2 \sigma^2_+}} e^{- \frac{({\mathpzc}{p}_- - p_-)^2}{2 \sigma^2_-}}.
\end{aligned}$$
At this point we implement the BO approximation in , which allow us to write
$$\label{eq_bianchiI:rewriting_approx_adm_wave_packet}
\begin{aligned}
\Phi({\mathpzc}{t},{\underline}{\beta}) \approx& \frac{e^{-\frac{i}{\hbar} \int_{{\mathpzc}{t}_s}^{{\mathpzc}{t}} \dd {\mathpzc}{t}' \sqrt{\mu^2({\mathpzc}{t}')}}}{\sqrt[4]{\mu_1^2}} \int_{\mathbb{R}^2} \frac{\dd^2 p}{(2 \pi \hbar)\sqrt{2 \pi \sigma_+\sigma_-}} \\
&\frac{e^{- \frac{i}{2 \hbar} (p_+^2 + p_-^2) \int_{{\mathpzc}{t}_s}^{{\mathpzc}{t}} \dd t' \frac{1}{\sqrt{\mu^2({\mathpzc}{t}')}}}}{\sqrt[4]{1+ \frac{p_+^2 + p_-^2}{\mu_1^2}}} \\
&e^{\frac{i}{\hbar} {\underline}{p} \cdot ({\underline}{\beta} - {\mathpzc}{{\underline}{b}})} e^{- \frac{({\mathpzc}{p}_+ - p_+)^2}{2 \sigma^2_+}} e^{- \frac{({\mathpzc}{p}_- - p_-)^2}{2 \sigma^2_-}}.
\end{aligned}$$
Comparing with we see two main differences:
- the factor $\qty(1 + \frac{p_+^2 + p_-^2}{\mu_1^2})^{-\frac{1}{4}}$ in is not present in ,
- the factor $(\mu^2_1)^{-\frac{1}{4}}$ in is replaced by $(\mu^2({\mathpzc}{t}))^{-\frac{1}{4}}$ in .
If we inspect the probability of finding the Universe in a region $\mathcal{B}$ of the plane $({\beta_{+}},{\beta_{-}})$, we get the same result for both methods, namely:
$$\label{eq_bianchiI:probability_final}
\begin{aligned}
P(\mathcal{B}) =& \frac{\pi}{\sqrt{\frac{1}{4 \hbar^2} \qty(\int_{{\mathpzc}{t}_s}^{{\mathpzc}{t}} \dd {\mathpzc}{t}' \frac{1}{\sqrt{\mu^2({\mathpzc}{t}')}})^2 + \frac{1}{4 \sigma^4_+}}} \\
&\frac{\pi}{\sqrt{\frac{1}{4 \hbar^2} \qty(\int_{{\mathpzc}{t}_s}^{{\mathpzc}{t}} \dd {\mathpzc}{t}' \frac{1}{\sqrt{\mu^2({\mathpzc}{t}')}})^2 + \frac{1}{4 \sigma^4_-}}} \\
&\int_{\mathcal{B}} \frac{\dd^2 \beta}{(2\pi\hbar)^2 \pi \sigma_+ \sigma_-} e^{- \frac{\qty[\beta_+ - \qty({\mathpzc}{b}_++{\mathpzc}{p}_+ \int_{{\mathpzc}{t}_s}^{{\mathpzc}{t}} \dd {\mathpzc}{t}' \frac{1}{\sqrt{\mu^2(t')}})]^2}{\sigma_+^2 \qty[ \frac{1}{4 \hbar^2}\qty( \int_{{\mathpzc}{t}_s}^{{\mathpzc}{t}} \dd {\mathpzc}{t}' \frac{1}{\sqrt{\mu^2({\mathpzc}{t}')}})^2 + \frac{1}{4 \sigma^4_+} ]}} \\
&e^{- \frac{\qty[\beta_- - \qty({\mathpzc}{b}_- +{\mathpzc}{p}_- \int_{{\mathpzc}{t}_s}^{{\mathpzc}{t}} \dd {\mathpzc}{t}' \frac{1}{\sqrt{\mu^2(t')}})]^2}{\sigma_-^2 \qty[ \frac{1}{4 \hbar^2}\qty( \int_{{\mathpzc}{t}_s}^{{\mathpzc}{t}} \dd {\mathpzc}{t}' \frac{1}{\sqrt{\mu^2({\mathpzc}{t}')}})^2 + \frac{1}{4 \sigma^4_-} ]}}.
\end{aligned}$$
The expectation value of the position operators $\hat{{\underline}{\beta}}$ over the wave packet is
$$\label{eq_bianchiI:mean_position}
\langle \hat{{\underline}{\beta}} \rangle = {\underline}{{\mathpzc}{b}} + {\underline}{{\mathpzc}{p}} \int_{{\mathpzc}{t}_s}^{{\mathpzc}{t}} \dd {\mathpzc}{t}' \frac{1}{\sqrt{\mu^2({\mathpzc}{t}')}}$$
and the variances $\sigma^2_+({\mathpzc}{t})$ and $\sigma_-^2({\mathpzc}{t})$ change in time as follows
\[eq\_bianchiI:width\] $$\begin{aligned}
&\sigma^2_\pm({\mathpzc}{t}) = \frac{\sigma_\pm}{\sqrt{2}} \sqrt{\frac{1}{4 \hbar^2}\qty( \int_{{\mathpzc}{t}_s}^{{\mathpzc}{t}} \dd {\mathpzc}{t}' \frac{1}{\sqrt{\mu^2({\mathpzc}{t}')}})^2 + \frac{1}{4 \sigma^4_\pm}} ; \label{eq_bianchiI:width_+}\,.\end{aligned}$$
The expression coincide with the classical trajectories obtained by solving the equations and for
$$\label{eq_bianchiI:N_to_have_classical_motion}
N({\mathpzc}{t}) = \frac{3 c \kappa}{4 \pi G T} \frac{e^{3 {\mathpzc}{t}}}{\sqrt{\mu^2({\mathpzc}{t})}}\,,$$
while outlines that semiclassical wave packets do not spread in time as the Universe expands in the momenta representation.
Conclusions
===========
The aim of Quantum Cosmology is to provide a quantum description for homogeneous cosmological models. However, several problems arise. A major issue is the problem of time, [*i.e.*]{} the identification of a proper time-like variable avoiding the frozen formalism, which generically leads to inequivalent results whether it is realized before or after quantization (see for instance [@Guven1992]). In this work, we spell out this problem for a Bianchi I model. In particular, we made a comparison between the RPSQ formalism and the Vilenkin proposal. In the former the super-Hamiltonian constraint is classically solved, the isotropic variable is taken as time and only anisotropy degrees of freedom are quantized. In the latter, implementing both WKB and Born-Oppenheimer approximations, a probabilistic interpretation for the wave functions is achieved. The RPSQ is the most straightforward procedure, but it is plagued by the non-locality of the Hamiltonian density and the presence of a time-dependent mass-like term, due to the contribution of matter energy density. These issues can be solved for the Bianchi I model, suggesting a procedure which could be extended to more general cases. The Vilenkin proposal is more manageable, but it cannot be regarded as a fundamental approach, since a probabilistic interpretation is achieved only after performing a semiclassical limit. Hence, making the comparison with RPSQ we tested the viability of the Vilenkin proposal. We outlined how the two approaches provided the same probability densities for finding the Universe in a finite region of the $(\beta_+,\beta_-)$ plane if the spectra of the corresponding momenta extend over sufficiently small values, so that the contribution of anisotropies to the total energy density is negligible with respect to that of matter. In other words, the viability of Vilenkin proposal is restricted to those physical scenarios in which the overall dynamics is dominated by the matter contributions. However, it is worth noting that also the Dirac procedure for solving the constraint is, to some extent, a Born-Oppenheimer approximation on the system dynamics: indeed, selecting the Universe volume as the proper time variable makes it classical ab-initio and its role in the wave function is on a different footing with respect to the gravitational degrees of freedom $\beta_\pm$. The overlap of the probability distributions in the two cases above can then be regarded as the consequence of a similar Born-Oppenheimer approximation on the Dirac constraint, but it must be emphasized that only the reduced phase space approach is intrinsically valid for arbitrary values of the anisotropy variables. On this level, we can conclude that the main achievement of this analysis relies on the constraint we have to impose on the anisotropy variables phase space: the Vilenkin procedure holds only when a “light dynamics” of the anisotropies is considered.
[*Acknowledgments.*]{}— FC is supported by funds provided by the National Science Center under the agreement DEC-2011/02/A/ST2/00294.
[99]{}
B. S. DeWitt, Phys. Rev. [**160**]{}, 1113 (1967).
B. S. DeWitt, Phys. Rev. [**162**]{}, 1195 (1967).
B. S. DeWitt, Phys. Rev. [**162**]{}, 1239 (1967).
K. Kuchar, “Canonical methods of quantisation”, C. Isham, R. Penrose & D. Sciama eds in “Quantum Gravity 2: A Second Oxford Symposium”, Clarendon Press, Oxford, 329–374 (1981).
G. Montani, M.V. Battisti, R. Benini, G. Imponente, “Primordial Cosmology”, (World Scientific 2011)
F. Cianfrani, O.M. Lecian, M. Lulli, G. Montani, “Canonical Quantum Gravity: Fundamentals and Recent Developments”, (World Scientific, 2014)
L.D. Landau and E.M. Lifshitz, “The Classical Theory of Fields”, (Pergamon, Amsterdam 1975)
W. F. Blyth and C. J. Isham, Phys. Rev. D [**11**]{}, 768 (1975).
C. W. Misner, Phys. Rev. Lett. [**22**]{}, 1071 (1969).
C.W. Misner, K.S. Thorne, J.A. Wheeler, “Gravitation”, (Freeman and Co. 1971)
C. W. Misner, Phys. Rev. [**186**]{}, 1319 (1969).
R. Benini and G. Montani, Class. Quant. Grav. [**24**]{}, 387 (2007)
M. Henneaux and C. Teitelboim, “Quantization of Gauge Systems”, (Princeton University Press, 1992)
A. Vilenkin, [*Phys. Rev. D*]{} [**39**]{}, 4, 1116 (1989)
R. L. Arnowitt, S. Deser and C. W. Misner, Gen. Rel. Grav. [**40**]{}, 1997 (2008)
L. Bianchi, [*Mem. Mat. Fis. Soc. Ital. Sci.*]{} [**11**]{}, 267(1898)
P.V. Moniz, “Quantum Cosmology - The Supersymmetric Perspective - Vol. 1: Fundamentals”, (Springer Berlin Heidelberg, 2010)
R.M. Wald, “General Relativity”, (University of Chicago Press, 1984)
M.P. Ryan and L.C. Shepley, “Homogeneous Relativistic Cosmologies”, (Princeton University Press, 1975)
S.S. Schweber, “An Introduction to Relativistic Quantum Mechanics”, (Row, Peterson and company, 1961)
C. L[ä]{}mmerzahl, [*Journal of Mathematical Physics*]{} [**34**]{}, 9, 3918(1993)
A. Mostafazadeh, Class. Quant. Grav. [**20**]{}, 155 (2003)
C. Kiefer, [*Class. Quant. Grav.*]{} [**4**]{}, 5, 1369 (1987)
J. Guven and M.P. Ryan, [*Phys. Rev. D*]{} [**45**]{}, 10, 3559(1992)
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'A novel sentence embedding method built upon semantic subspace analysis, called semantic subspace sentence embedding (S3E), is proposed in this work. Given the fact that word embeddings can capture semantic relationship while semantically similar words tend to form semantic groups in a high-dimensional embedding space, we develop a sentence representation scheme by analyzing semantic subspaces of its constituent words. Specifically, we construct a sentence model from two aspects. First, we represent words that lie in the same semantic group using the intra-group descriptor. Second, we characterize the interaction between multiple semantic groups with the inter-group descriptor. The proposed S3E method is evaluated on both textual similarity tasks and supervised tasks. Experimental results show that it offers comparable or better performance than the state-of-the-art. The complexity of our S3E method is also much lower than other parameterized models.'
author:
-
-
-
-
-
bibliography:
- 'IEEEabrv.bib'
- 'mybibfile.bib'
title: Efficient Sentence Embedding via Semantic Subspace Analysis
---
Introduction {#sec:introduction}
============
Word embedding technique is widely used in natural language processing (NLP) tasks. For example, it improves downstream tasks such as machine translation [@MT], syntactic parsing [@Pars], and text classification [@Shen2018Baseline]. Yet, many NLP applications operate at the sentence level or a longer piece of texts. Although sentence embedding has received a lot of attention recently, encoding a sentence into a fixed-length vector to capture different linguistic properties remains to be a challenge.
Universal sentence embedding aims to compute sentence representation that can be applied to any tasks. It can be categorized into two types: i) parameterized models and ii) non-parameterized models. Parameterized models are mainly based on deep neural networks and demand training in their parameter updates. Inspired by the famous word2vec model [@mikolov2013distributed], the skip-thought model [@kiros2015skip] adopts an encoder-decoder model to predict context sentences in an unsupervised manner. InferSent model [@conneau2017supervised] is trained by high quality supervised data; namely, the Natural Language Inference data. It shows that supervised training objective can outperform unsupervised ones. USE [@USE] combines both supervised and unsupervised objectives and transformer architecture is employed. The STN model [@STN] leverages a multi-tasking framework for sentence embedding to provide better generalizability. With the recent success of deep contextualized word models, SBERT [@reimers-2019-sentence-bert] and SBERT-WK [@SBERT-WK] are proposed to leverage the power of self-supervised learning from large unlabeled corpus. Different parameterized models attempt to capture semantic and syntactic meanings from different aspects. Even though their performance is better as compared with non-parameterized models, parameterized ones are more complex and computationally expensive. Since it is challenging to deploy parameterized models into mobile or terminal devices, finding effective and efficient sentence embedding models are necessary.
![Overview of the proposed S3E method.[]{data-label="S3E"}](figure2.pdf){width="50.00000%"}
Non-parameterized sentence embedding methods rely on high quality word embeddings. The simplest idea is to average individual word embeddings, which already offers a tough-to-beat baseline. By following along this line, several weighted averaging methods have been proposed, including tf-idf, SIF [@arora2017asimple], and GEM [@GEM]. Concatenating vector representations of different resources yields another family of methods. Examples include SCDV [@SCDV] and $p$-mean [@ruckle2018concatenated]. To better capture the sequential information, DCT [@DCT] and EigenSent [@EigenSent] were proposed from a signal processing perspective.
Here, we propose a novel non-parameterized sentence embedding method based on semantic subspace analysis. It is called semantic subspace sentence embedding (S3E) (see Fig. \[S3E\]). The S3E method is motivated by the following observation. Semantically similar words tend to form semantic groups in a high-dimensional embedding space. Thus, we can embed a sentence by analyzing semantic subspaces of its constituent words. Specifically, we use the intra- and inter-group descriptors to represent words in the same semantic group and characterize interactions between multiple semantic groups, respectively.
This work has three main contributions.
1. The proposed S3E method contains three steps: 1) semantic group construction, 2) intra-group descriptor and 3) inter-group descriptor. The algorithms inside each step are flexible and, as a result, previous work can be easily incorporated.
2. To the best of our knowledge, this is the first work that leverages correlations between semantic groups to provide a sentence descriptor. Previous work using the covariance descriptor [@Cov] yields super-high embedding dimension (e.g. 45K dimensions). In contrast, the S3E method can choose the embedding dimension flexibly.
3. The effectiveness of the proposed S3E method in textual similarity and supervised tasks is shown experimentally. Its performance is as competitive as that of very complicated parametrized models.[^1]
Related Previous Work
=====================
Vector of Locally Aggregated Descriptors (VLAD) is a famous algorithm in the image retrieval field. Same with Bag-of-words method, VLAD trains a codebook based on clustering techniques and concatenate the feature within each clusters as the final representation. Recently work called VLAWE (vector of locally-aggregated word embeddings) [@VLAWE], introduce this idea into document representation. However, VLAWE method suffers from high dimensionality problem which is not favored by machine learning models. In this work, a novel clustering method in proposed by taking word frequency into consideration. At the same time, covariance matrix is used to tackle the dimensionality explosion problem of VLAWE method.
Recently, a novel document distance metric called Word Mover’s Distance (WMD) [@kusner2015word] is proposed and achieved good performance in classification tasks. Based on the fact that semantically similar words will have close vector representations, the distance between two sentences are models as the minimal ’travel’ cost for moving the embedded words from one sentence to another. WMD targets on modeling the distance between sentences in the shared word embedding space. It is natural to consider the possibility of computing the sentence representation directly from the word embedding space by semantical distance measures.
There are a few works trying to obtain sentence/document representation based on Word Mover’s Distance. D2KE (distances to kernels and embeddings) and WME (word mover’s embedding) converts the distance measure into positive definite kernels and has better theoretical guarantees. However, both methods are proposed under the assumption that Word Mover’s Distance is a good standard for sentence representation. In our work, we borrow the ’travel’ concept of embedded words in WMD’s method. And use covariance matrix to model the interaction between semantic concepts in a discrete way.
Proposed S3E Method {#sec:approach}
===================
As illustrated in Fig. \[S3E\], the S3E method contains three steps: 1) constructing semantic groups based on word vectors; 2) using the inter-group descriptor to find the subspace representation; and 3) using correlations between semantic groups to yield the covariance descriptor. Those are detailed below.
[**Semantic Group Construction.**]{} Given word $w$ in the vocabulary, $V$, its uni-gram probability and vector are represented by $p(w)$ and $v_w
\in \mathbb{R}^d$, respectively. We assign weights to words based on $p(w)$: $$\label{eq:epsilon}
\mbox{weight}(w) = \frac{\epsilon}{\epsilon+p(w)},$$ where $\epsilon$ is a small pre-selected parameter, which is added to avoid the explosion of the weight when $p(w)$ is too small. Clearly, $0
< \mbox{weight}(w) < 1$. Words are clustered into $K$ groups using the K-means++ algorithm [@k-means++], and weights are incorporated in the clustering process. This is needed since some words of higher frequencies (e.g. *’a’, ’and’, ’the’*) are less discriminative by nature. They should be assigned with lower weights in the semantic group construction process.
[**Intra-group Descriptor.**]{} After constructing semantic groups, we find the centroid of each group by computing the weighted average of word vectors in that group. That is, for the $i_{th}$ group, $G_i$, we learn its representation $g_i$ by $$g_i = \frac{1}{|G_i|}\sum_{w\in G_i} \mbox{weight}(w) v_w,$$ where $|G_i|$ is the number of words in group $G_i$. For sentence $S =
\{w_1, w_2, ..., w_m\}$, we allocate words in $S$ to their semantic groups. To obtain the intra-group descriptor, we compute the cumulative residual between word vectors and their centroid ($g_i$) in the same group. Then, the representation of sentence $S$ in the $i_{th}$ semantic group can be written as $$v_i = \sum_{w\in S\cap G_i} \mbox{weight}(w)(v_w - g_i).$$ If there are $K$ semantic groups in total, we can represent sentence $S$ with the following matrix: $$\label{eq:phi}
\Phi(S) =
\begin{bmatrix}
v_{1}^T \\
v_{2}^T \\
\vdots \\
v_{K}^T
\end{bmatrix}
=
\begin{pmatrix}
v_{11} & \dots & v_{1d} \\
v_{21} & \dots & v_{2d} \\
\vdots & \ddots & \vdots \\
v_{K1} & \dots & v_{Kd}
\end{pmatrix}_{K\times d},$$ where $d$ is the dimension of word embedding.
[**Inter-group Descriptor.**]{} After obtaining the intra-group descriptor, we measure interactions between semantic groups with covariance coefficients. We can interpret $\Phi(S)$ in (\[eq:phi\]), as $d$ observations of $K$-dimensional random variables, and use $u_\Phi\in \mathbb{R}^{K\times 1}$ to denote the mean of each row in $\Phi$. Then, the inter-group covariance matrix can be computed as $$C = [ C_{i,j} ]_{K \times K} = \frac{1}{d}(\Phi-\mu_\Phi)(\Phi-\mu_\Phi)^T \in \mathbb{R}^{K \times K},$$ where $$C_{i,j} = \sigma_{i,j}=\frac{(v_i-\mu_i)^T(v_j-\mu_j)}{d}.$$ is the covariance between groups $i$ and $j$. Thus, matrix $C$ can be written as $$C =
\begin{pmatrix}
\sigma_{1}^2 & \sigma_{12} & \dots & \sigma_{1K} \\
\sigma_{12} & \sigma_{2}^2 & \dots & \sigma_{2K} \\
\vdots & \vdots & \ddots & \vdots \\
\sigma_{1K} & \sigma_{2K} & \dots & \sigma_{K}^2
\end{pmatrix}.$$ Since the covariance matrix $C$ is symmetric, we can vectorize its upper triangular part and use it as the representation for sentence $S$. The Frobenius norm of the original matrix is kept the same with the Euclidean norm of vectorized matrices. This process produces an embedding of dimension $\mathbb{R}^{K(K+1)/2}$. Then, the embedding of sentence $S$ becomes $$\label{eq:sentence_embedding}
v(S) = vect(C) =
\left\{
\begin{array}{ll}
\sqrt{2}\sigma_{ij}, & \text{if}\quad i<j, \\
\sigma_{ii}, & \text{if}\quad i=j.
\end{array}
\right.$$ Finally, the sentence embedding in Eq. (\[eq:sentence\_embedding\]) is $L_2-\text{normalized}$.
[**Complexity**]{}
The semantic group construction process can be pre-computed for efficiency. Our runtime complexity is ($dN+Kd^2$), where $N$ is the length of a sentence, $K$ is the number of semantic groups, and $d$ is the dimension of word embedding in use. Our algorithm is linear with respect to the sentence length. The S3E method is much faster than all parameterized models and most of non-parameterized methods such as [@GEM] where the singular value decomposition is needed during inference. The run time comparison is also discussed in Sec. \[sec:runtime\].
Experiments {#sec:experiments}
===========
We evaluate our method on two sentence embedding evaluation tasks to verify the generalizability of S3E. Semantic textual similarity tasks are used to test the clustering and retrieval property of our sentence embedding. Discriminative power of sentence embedding is evaluated by supervised tasks.
For performance benchmarking, we compare S3E with a series of other methods including parameterized and non-parameterized ones.
1. Non-parameterized Models
1. Avg. GloVe embedding;
2. SIF [@arora2017asimple]: Derived from an improved random-walk model. Consist of two parts: weighted averaging of word vectors and first principal component removal;
3. $p$-means [@ruckle2018concatenated]: Concatenating different word embedding models and different power ratios;
4. DCT [@DCT]: Introduce discrete cosine transform into sentence sequential modeling;
5. VLAWE [@VLAWE]: Introduce VLAD (vector of locally aggregated descriptor) into sentence embedding field;
2. Parameterized Models
1. Skip-thought [@kiros2015skip]: Extend word2vec unsupervised training objectives from word level into sentence level;
2. InferSent [@conneau2017supervised]: Bi-directional LSTM encoder trained on high quality sentence inference data.
3. Sent2Vec [@sent2vec]: Learn n-gram word representation and use average as the sentence representation.
4. FastSent [@fastsent]: An improved Skip-thought model for fast training on large corpus. Simplify the recurrent neural network as bag-of-words representation.
5. ELMo [@ELMO]: Deep contextualized word embedding. Sentence embedding is computed by averaging all LSTM outputs.
6. Avg. BERT embedding [@devlin2018bert]: Average the last layer word representation of BERT model.
7. SBERT-WK [@SBERT-WK]: A fusion method to combine representations across layers of deep contextualized word models.
Textual Similarity Tasks
------------------------
\[STS\]
We evaluate the performance of the S3E method on the SemEval semantic textual similarity tasks from 2012 to 2016, the STS Benchmark and SICK-Relatedness dataset. The goal is to predict the similarity between sentence pairs. The sentence pairs contains labels between 0 to 5, which indicate their semantic relatedness. The Pearson correlation coefficients between prediction and human-labeled similarities are reported as the performance measure. For STS 2012 to 2016 datasets, the similarity prediction is computed using the cosine similarity. For STS Benchmark dataset and SICK-R dataset, they are under supervised setting and aims to predict the probability distribution of relatedness scores. We adopt the same setting with [@tai2015improved] for these two datasets and also report the Pearson correlation coefficient.
The S3E method can be applied to any static word embedding method. Here, we report three of them; namely, GloVe [@GloVe], FastText [@FastText] and L.F.P. [^2]. Word embedding is normalized using [@PVN]. Parameter $\epsilon$ in Eq. (\[eq:epsilon\]) is set to $10^{-3}$ for all experiments. The word frequency, $p(w)$, is estimated from the wiki dataset[^3]. The number of semantic groups, $K$, is chosen from the set $\{ 10,20,30,40,50 \}$ and the best performance is reported.
Experimental results on textual similarity tasks are shown in Table \[STS\], where both non-parameterized and parameterized models are compared. Recent parameterized method SBERT-WK provides the best performance and outperforms other method by a large margin. S3E method using L.F.P word embedding is the second best method in average comparing with both parameterized and non-parameterized methods. As mentioned, our work is compatible with any weight-based methods. With better weighting schemes, the S3E method has a potential to perform even better. As choice of word embedding, L.F.P performs better than FastText and FastText is better than GloVe vector in Table \[STS\], which is consistent with the previous findings [@wang2019evaluating]. Therefore, choosing more powerful word embedding models can be helpful in performance boost.
Supervised Tasks
----------------
\[Supervisedtable\]
The SentEval toolkit[^4] [@conneau2018senteval] is used to evaluate on eight supervised tasks:
1. MR: Sentiment classification on movie reviews.
2. CR: Sentiment classification on product reviews.
3. SUBJ: Subjectivity/objective classification.
4. MPQA: Opinion polarity classification.
5. SST2: Stanford sentiment treebank for sentiment classification.
6. TREC: Question type classification.
7. MRPC: Paraphrase identification.
8. SICK-Entailment: Entailment classification on SICK dataset.
The details for each dataset is also shown in Table \[table:downstream\]. For all tasks, we trained a simple MLP classifier that contain one hidden layer of 50 neurons. It is same as it was done in [@DCT] and only tuned the $L_2$ regularization term on validation sets. The hyper-parameter setting of S3E is kept the same as that in textual similarity tasks. The batch size is set to 64 and Adam optimizer is employed. For MR, CR, SUBJ, MPQA and MRPC datasets, we use the nested 10-fold cross validation. For TREC and SICK-E, we use the cross validation. For SST2 the standard validation is utilized. All experiments are trained with 4 epochs.
Experimental results on supervised tasks are shown in Table \[Supervisedtable\]. The S3E method outperforms all non-parameterized models, including DCT [@DCT], VLAWE [@VLAWE] and $p$-means [@ruckle2018concatenated]. The S3E method adopts a word embedding dimension smaller than $p$-means and VLAWE and also flexible in choosing embedding dimensions. As implemented in other weight-based methods, the S3E method does not consider the order of words in a sentence but splits a sentence into different semantic groups. The S3E method performs the best on the paraphrase identification (MRPC) dataset among all non-parameterized and parameterized methods excluding SBERT-WK. This is attributed to that, when paraphrasing, the order is not important since words are usually swapped. In this context, the correlation between semantic components play an important role in determining the similarity between a pair of sentences and paraphrases.
Comparing with parameterized method, S3E also outperforms a series of them including Skip-thought, FastSent and Sent2Vec. In general, parameterized methods performs better than non-parameterized ones on downstream tasks. The best performance is the recently proposed SBERT-WK method which incorporate a pre-trained deep contextualized word model. However, even though good perform is witnessed, deep models are requiring much more computational resources which makes it hard to integrate into mobile or terminal devices. Therefore, S3E method has its own strength in its efficiency and good performance.
Inference Speed {#sec:runtime}
---------------
We compare the inference speed of S3E with other models including the non-parameterized and parameterized ones. For fair comparison, the batch size is set to 1 and all sentences from STSB datasets are used for evaluation (17256 sentences). All benchmark results can run on CPU[^5] and GPU[^6]. The results are showed in Table \[exp:speed\].
Comparing the other method, S3E is the very efficient in inference speed and this is very important in sentence embedding. Without the acceleration of powerful GPU, when doing comparing tasks of 10,000 sentence pairs, deep contextualized models takes about 1 hour to accomplish, which S3E only requires 13 seconds.
Sensitivity to Cluster Numbers
------------------------------
![Comparing results with different settings on cluster numbers. STSB result is presented in Pearson Correlation Coefficients (%). SICK-E and SST2 are presented in accuracy.[]{data-label="exp:sensitivity"}](sensitivity.pdf){width="\columnwidth"}
We test the sensitivity of S3E to the setting of cluster numbers. The cluster number is set from 5 to 60 with internal of 5 clusters. Results for STS-Benchmark, SICK-Entailment and SST2 dataset are reported. As we can see from Figure \[exp:sensitivity\], performance of S3E is quite robust for different choice of cluster numbers. The performance varies less than 1% in accuracy or correlation.
Discussion
==========
Averaging word embedding provides a simple baseline for sentence embedding. A weighted sum of word embeddings should offer improvement intuitively. Some methods tries to improve averaging are to concatenate word embedding in several forms such as $p$-means [@ruckle2018concatenated] and VLAWE[@VLAWE]. Concatenating word embeddings usually encounters the dimension explosion problem. The number of concatenated components cannot be too large.
Our S3E method is compatible with exiting models and its performance can be further improved by replacing each module with a stronger one. First, we use word weights in constructing semantically similar groups and can incorporate different weighting scheme in our model, such as SIF [@arora2017asimple], GEM [@GEM]. Second, different clustering schemes such as the Gaussian mixture model and dictionary learning can be utilized to construct semantically similar groups [@SCDV; @guptap]. Finally, the intra-group descriptor can be replaced by methods like VLAWE [@VLAWE] and $p$-means [@ruckle2018concatenated]. In inter-group descriptor, correlation between semantic groups can also be modeled in a non-linear way by applying different kernel functions. Another future direction is to add sequential information into current S3E method.
Conclusion
==========
A sentence embedding method based on semantic subspace analysis was proposed. The proposed S3E method has three building modules: semantic group construction, intra-group description and inter-group description. The S3E method can be integrated with many other existing models. It was shown by experimental results that the proposed S3E method offers state-of-the-art performance among non-parameterized models. S3E is outstanding for its effectiveness with low computational complexity.
[^1]: Our code is available at [github](https://github.com/BinWang28/Sentence-Embedding-S3E).
[^2]: concatenated LexVec, FastText and PSL
[^3]: https://dumps.wikimedia.org/
[^4]: https://github.com/facebookresearch/SentEval
[^5]: Intel i7-5930 of 3.50GHz with 12 cores
[^6]: Nvidia GeForce GTX TITAN X
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
author:
- 'S. Rost'
- 'A. Eckart'
- 'T. Ott'
bibliography:
- '7555.bib'
date: 'Received March 27, 2007; accepted March 27, 2008'
title: 'Near-infrared polarization images of the Orion proplyds'
---
Introduction {#sect:intro}
============
The Orion molecular clouds at a distance of about 460 parsecs are the nearest location of ongoing massive star formation. This region also harbors one of the densest young stellar clusters [@1994AJ....108.1382M], the Orion Nebula Cluster (ONC). The core of the ONC is known as the Trapezium cluster, named after the four massive OB stars at its very center. The dominant Trapezium star is $\theta^1$ Ori C, which generates most of the intense ultraviolet radiation field that photo-dissociates and photo-ionizes the surrounding material. The discovery of proto-planetary disks (proplyds) in this region by [@1993ApJ...410..696O] with the Hubble Space Telescope (HST) has provided an unprecedented opportunity to study planet-forming disks around young stars. These objects are surrounded by bright, crescent, and tear-drop shaped ionization fronts (IF) resulting from gas flow off of the disk. At greater distances and shielded from radiation, several proplyds without ionization front have been observed. They are seen in silhouette against the nebular background [@2000AJ....119.2919B].
New insights into the nature of proplyds have been acquired through high-resolution multi-spectral imaging with HST [@1998AJ....116..293B], follow-up imaging, and spectroscopy at a variety of wavelengths from the ground, and intensive theoretical modeling [@1998ApJ...499..758J; @1999ApJ...515..669S e.g.]. As a result, an irradiated disk model has been developed to account for the presently known observed features.
Near infrared (NIR) observations of the Trapezium cluster [@1998AJ....116.1816H; @2004AJ....128.1254L] have shown that most young stars have excess emission in this wavelength regime. This is interpreted as the presence of circumstellar disks. According to the study by [@2000AJ....120.3162L], 80% of the stars in the Trapezium cluster show an infrared excess and are likely to be surrounded by disks.
Currently much activity is devoted to obtain physical properties of the circumstellar dust distribution and these disks. [@2006ApJ...642.1140O] showed that encounters of stars and circumstellar disks in the Trapezium cluster are quite frequent and thus can cause a considerable contribution to the mass loss and the truncation of proplyds. Encounter-triggered planet formation is also possible as an important supplement to triggering by super massive stars [@2003ASPC..287..263B] and supernovae-triggering of the OB-stars [@2005ASPC..341..107H]. The disks of the proplyds close to the Trapezium with a typical radius of 20–80 AU are not resolvable directly. In the Orion nebula the YSO 177-341, also known as HST1 is one of the best-studied proplyds because of its large size and orientation [@1998AJ....116..293B; @1999AJ....118.2350H; @2000RMxAC...9..198H]. The very bright ionization front outshines details of its inner structure.
showed that polarimetry can be used to trace the relative orientation of disks in young binary systems. The polarization structure of a massive young stellar object (YSO) in OMC-1 revealed a surrounding disk/bipolar outflow system [@2005Natur.437..112J]. Near-infrared polarimetry is a proven remedy for mapping hidden effects and structures . In the case of dust around a bright star, only the scattered light from the disk is expected to be polarized, the unpolarized light from the central star is suppressed by imaging polarimetry. Therefore this technique can improve the contrast and the detection-threshold, and it provides a possibility to map structures very close to bright stars without coronographic masks [@2001ApJ...553L.189K; @2006MNRAS.365.1348H]. Polarimetric images can also trace the sources of illumination of the reflection nebula in star-forming regions [@2007PASJ...59..221H].
We carried out a JHK polarimetry survey of selected proplyds in the Trapezium Cluster to study dust and disk parameters in this dense star-forming region. Those disks of the proplyds close to the Trapezium that have a bright IF are so far only seen as IR excesses. With the polarimetric images we want to map the circumstellar dust distribution and extract parameter ranges for the disks to classify the Trapezium proplyds. In this paper we want to present the results from our first polarimetric study of the Orion proplyds. We begin in Sect. 2 with the description of the observational data and summarize in Sect. 3 the basic properties of one of the sources extracted from the polarized data set. In the next section we consider different analytical techniques for a more detailed view of the proplyds properties including modeling of circumstellar dust particles and qualitative comparison to simulated NIR polarimetric images. A summary and an outlook of forthcoming surveys is noted in Sect. 5.
Observations and data reduction {#chap:obs}
===============================
![Near-Infrared (left, KHJ $\equiv$ RGB) image of the giant proplyd 177-341 and its neighbors in comparison to the optical HST image [right, red is , green is , blue is , @1998AJ....116..293B], with north facing upwards and east to the left. The proplyd 178-342 is covered by the ionization front in the optical, but barely visible at the end of the tail.[]{data-label="ifoptn"}](7555fig1){width="45.00000%"}
High-resolution near-infrared polarimetric images were acquired in 2001 with the ESO Adaptive Optics System ADONIS and the SHARP II+ camera at the La Silla 3.6m[^1] telescope . A rotating wire grid polarizer provided polarization angles from $0^\circ$ to $150^\circ$ with a $30^\circ$ increment. The sources are 10 proplyds close to the Trapezium stars, embedded in three pointings with a field-of-view (FOV) of 8.96”$\times$8.96” and a 0.035” plate scale. Images were taken over three nights in the J, H, and K band. No occulting masks were used in these observations due to multiple sources per FOV. The chosen detector integration time for a good signal-to-noise ratio (SNR) and dynamic range was 15 s per polarization angle and FOV. The total time spent on each source was 6750 s, with several measurements per angle in different nights (375 s per source, angle, and band). $\theta^1$ Ori C was used for wavefront sensing of the adaptive optics. The natural seeing varied from 0.5” to 1”. The achieved Strehl ratio in JHK was 6%, 13%, and 20% for the most distant FOV. Off-source sky for each angle and dark frames were taken, as well as frames with unpolarized reference stars.
The data reduction was carried out using the DPUSER software for astronomical image analysis [T. Ott, http://www.mpe.mpg.de/$\sim$ott/dpuser/, see also @1990ASPC...14..336E] and IDL routines. Due to some cooling problems of the detector, some part of the data had to be sorted out by a statistical filter algorithm followed by a manual inspection. The raw images were sky-subtracted and flat field-corrected.
The non-polarized light consists mainly of the central star‘s light and the uncorrected AO residuals, which is the dominant noise source of ground-based AO-observations in the NIR. Polarimetric images are very sensitive to minor deviations of shift-and-add operations. Therefore this reduction step was carefully carried out and manually optimized for every single frame. A minimal number of three frames (spanning more than $90^\circ$ position angle) is required to determine the linear polarization and its position angles. The redundant data set of three orthogonal polarization component pairs increases the SNR and reduces the instrumental polarization. By fitting the oversampled polarization curve with a cosine function, the linear polarization angle and degree maps ($\Pi_L$) have been determined. Polarimetric differential images (PDI) can be obtained using $Q_0=I_{0^\circ}-I_{90^\circ}$ (Stokes Q-parameter), $Q_{30}=I_{30^\circ}-I_{120^\circ}$, and $Q_{60}=I_{60^\circ}-I_{150^\circ}$. In the widely used notation of linear polarized light this converts to $Q=Q_0$, and $$U=\frac{Q_{\phi}-Q\cdot\cos(2\phi)}{\sin(2\phi)}$$ with $\phi=30^\circ$ or $\phi=60^\circ$, and $$\Pi_L=\frac{\sqrt{Q^2+U^2}}{I}, \; \theta=\frac{1}{2} \arctan{\frac{U}{Q}}\;.$$ The Q- and U-parameters are used for better comparison to the simulated images of the model.
Results
=======
Some of the sources show a clearly extended structure in the NIR referring to an illuminated dust envelope. The surrounding nebula emission is not as strong and spatially varying as in the optical. The properties of the targets are listed in Table \[tabprop\], including the IF-sizes and NIR magnitudes, J-H and H-K colors. These were photometrically calibrated with the standard star HD98161 from the 2MASS catalog. This unpolarized star was also used for polarimetric calibration. The aperture polarization of the targets is shown in Table \[tabpol\]. Also the integrated polarization in two annuli is shown for comparison and classification of circumstellar material. The inner region is usually dominated by alignment and subtraction residuals [@2006MNRAS.365.1348H]. The annulus of 0.3”-0.5” refers to the circumstellar dust, which is illuminated by the central star. The outer annulus should be dominated by polarized light from the surrounding reflecting nebula where the source of illumination is unknown.
![Spectral distribution of polarization for J, H, and K bands. In an aperture of 0.1” the light of the central star the scattering by foreground material dominates. In an annulus close to the star the polarization is higher due to the circumstellar envelope and disk. The outer annulus is dominated by the reflection nebula and is highly polarized. The dashed line is the average polarization spectrum, the dotted line is 177-341. []{data-label="jhkplot"}](7555fig2){height="48.00000%"}
The polarization values as a function of the wavelength are plotted in Fig. \[jhkplot\]. Although the integrated polarization degree and angle is highly affected by minor deviations in shift-and-add operations and PSF effects (see Sect. \[simchap\]), the average values give a good overview of the circumstellar polarization distribution. The nebula emission in the outer annuli is polarized from 10% in K-band to 20% in J-band. This is in good agreement with a recent wide-field polarization study of the ONC in the NIR [@2006ApJ...649L..29T].
![Polarimetric differential images (KHJ $\equiv$ RGB) of the sources in the FOV shown in Fig. \[ifoptn\]. Top: Proplyd 177-341 in $Q_0$ (left) and $Q_{60}$ (right). A clear butterfly pattern is visible in all bands. Mid and Bottom: 178-342 and 178-343 show a trefoil that is supposed to be an AO-effect. With this comparison of sources in one FOV, lens- (e.g. astigmatism) and PSF-effects can be excluded as the origin for the results of 177-341.[]{data-label="resrgb"}](7555fig3){width="48.00000%"}
To remove all non-polarized light from the sources, we subtracted orthogonal components from each other. These images of the Stokes Q- or U-parameters show a characteristic trefoil pattern for most of the sources (Fig. \[resrgb\]). The trefoil is a mode in the AO wavefront reconstruction that may contain residual power if the AO-loop is run under variable atmospheric conditions. However, for the giant proplyd 177-341 we find significant butterfly patterns or two lobes, which is the polarization pattern expected for a circumstellar dust envelope.
The sources around 177-341 (Fig. \[resrgb\]) lie in a relatively dark region of the nebula. This is why we concentrate here in particular on the field of the proplyds around 177-341. The NIR-image in Figure \[ifoptn\] shows quite a different view of the proplyds morphology than the known optical images [@1998AJ....116..293B]. First of all the bright ionization front, which covers inner structures of the proplyds in the optical is not visible in the NIR-images. This reveals, for example, a previously unknown source behind the tail of the proplyd 177-341 (Fig. \[ifoptn\], left). If the new source is a proplyd, it would be named 178-342, following the designation system introduced by [@1994ApJ...436..194O]. The observed JHK-colors differ from the neighbor 177-341, 178-342 seems redder (Table \[tabprop\]). This may be due to extinction by the overlying ionization front.
------------- ------------- -------------- ------------------ -------------------------- -------- ------- ------- ------- -------
Designation R.A. DECL. Dist. to IF$^{\mathrm{a}}$ (opt.)
(J2000) (J2000) $\theta^1$ Ori C Tail J H K$_S$ J-H H-K
157-323 05 35 15.71 -05 23 22.59 11.25” $<$0.7” 9.88 9.05 8.93 0.83 0.11
158-323 05 35 15.82 -05 23 22.50 9.75” 1.8–2.5” 9.73 8.83 8.53 0.91 0.3
158-326 05 35 15.83 -05 23 25.62 10.2” 0.85” 12.9 11.34 10.65 1.56 0.69
158-327 05 35 15.79 -05 23 26.61 10.53” 1.0–1.5” 11.68 10.4 9.66 1.28 0.75
163-317 05 35 16.27 -05 23 16.72 6.42” 1.6” 10.14 9.41 9.2 0.72 0.21
166-316 05 35 16.60 -05 23 16.32 7.36” 0.4” 11.4 10.47 10.2 0.93 0.27
167-317 05 35 16.73 -05 23 16.63 7.08” 0.67”$\times$2.1” 10.37 9.41 8.94 0.96 0.47
177-341 05 35 17.67 -05 23 40.96 25.98” 0.8”$\times$3.5” 12.44 11.83 11.53 0.61 0.3
178-342 05 35 17.76 -05 23 42.50 27.2” $\ldots$ 12.93 11.67 11.2 1.26 0.46
178-344 05 35 17.78 -05 23 44.25 29.1” $\ldots$ 12.083 11.28 11.1 0.8 0.17
HD98161 11 17 12.01 -38 00 51.72 $\ldots$ $\ldots$ 6.075 6.019 5.992 0.056 0.027
------------- ------------- -------------- ------------------ -------------------------- -------- ------- ------- ------- -------
[@1998AJ....116..293B]
magnitude errors $\sim0.1$
------------- -------- ------- ----------- ------- ----------- ------- ----------- ------- ----------- ------- ----------- ------- -----------
Designation Filter
157-323 J 18.11 $\pm$0.70 35.5 $\pm$1.0 14.38 $\pm$0.25 45.5 $\pm$0.3 14.89 $\pm$0.37 38.7 $\pm$0.8
H 4.76 $\pm$0.26 84.4 $\pm$5.0 6.41 $\pm$1.96 22.7 $\pm$0.2 11.95 $\pm$0.68 26.0 $\pm$1.0
K$_S$ 3.71 $\pm$1.91 62.1 $\pm$4.7 2.77 $\pm$0.64 -67.2 $\pm$29.6 11.25 $\pm$0.71 -56.7 $\pm$0.9
158-323 J 19.61 $\pm$2.09 21.0 $\pm$2.1 13.33 $\pm$1.65 44.6 $\pm$0.3 12.11 $\pm$0.77 40.3 $\pm$0.5
H 3.58 $\pm$0.18 61.3 $\pm$2.7 5.33 $\pm$2.22 20.8 $\pm$1.8 9.30 $\pm$0.63 23.3 $\pm$0.4
K$_S$ 4.12 $\pm$1.22 24.4 $\pm$6.0 2.72 $\pm$0.92 -60.4 $\pm$31.4 10.31 $\pm$0.37 -51.5 $\pm$0.6
158-326 J 45.32 $\pm$4.52 35.9 $\pm$2.1 32.53 $\pm$0.71 29.9 $\pm$0.7 46.87 $\pm$1.53 29.4 $\pm$0.1
H 9.12 $\pm$2.50 89.2 $\pm$1.6 9.26 $\pm$1.31 29.9 $\pm$0.6 20.06 $\pm$0.74 29.1 $\pm$0.1
K$_S$ 3.08 $\pm$1.17 -76.4 $\pm$10.5 2.61 $\pm$0.46 42.2 $\pm$7.4 4.30 $\pm$0.61 17.6 $\pm$2.0
158-327 J 19.14 $\pm$3.42 32.8 $\pm$0.4 17.27 $\pm$0.93 40.9 $\pm$1.1 18.24 $\pm$0.85 35.0 $\pm$0.2
H 6.94 $\pm$1.69 -78.4 $\pm$3.4 7.95 $\pm$1.25 34.0 $\pm$0.8 11.69 $\pm$0.21 33.7 $\pm$0.3
K$_S$ 3.07 $\pm$1.84 -40.2 $\pm$16.0 3.76 $\pm$0.42 58.1 $\pm$7.4 1.93 $\pm$0.22 -88.7 $\pm$2.3
163-317 J 24.77 $\pm$2.72 34.9 $\pm$1.5 9.92 $\pm$0.44 59.5 $\pm$2.9 13.23 $\pm$1.65 71.0 $\pm$0.3
H 2.89 $\pm$1.66 -29.5 $\pm$23.1 6.41 $\pm$1.91 68.3 $\pm$4.3 12.57 $\pm$0.87 -87.3 $\pm$0.3
K$_S$ 3.96 $\pm$1.61 -79.1 $\pm$5.5 5.10 $\pm$0.98 29.9 $\pm$1.0 3.12 $\pm$0.43 -24.8 $\pm$21.0
166-316 J 25.50 $\pm$1.58 44.3 $\pm$1.3 14.28 $\pm$1.95 54.8 $\pm$1.7 9.75 $\pm$1.33 77.5 $\pm$0.5
H 3.00 $\pm$0.94 36.2 $\pm$8.2 7.46 $\pm$0.71 69.5 $\pm$2.8 5.88 $\pm$0.29 76.4 $\pm$1.1
K$_S$ 5.35 $\pm$2.07 -56.6 $\pm$33.3 2.22 $\pm$1.09 82.4 $\pm$9.8 4.09 $\pm$0.34 -44.2 $\pm$35.1
167-317 J 23.94 $\pm$1.38 30.8 $\pm$1.7 11.70 $\pm$1.25 38.5 $\pm$1.0 15.28 $\pm$2.48 11.9 $\pm$2.3
H 5.86 $\pm$1.48 -10.8 $\pm$1.5 7.84 $\pm$2.30 56.8 $\pm$3.0 13.34 $\pm$6.57 21.4 $\pm$6.4
K$_S$ 1.14 $\pm$0.75 85.6 $\pm$11.5 5.09 $\pm$0.45 68.4 $\pm$2.9 8.26 $\pm$6.54 19.1 $\pm$9.6
177-341 J 5.54 $\pm$1.72 -17.0 $\pm$4.5 9.68 $\pm$1.64 -6.1 $\pm$3.8 20.38 $\pm$2.59 -3.6 $\pm$0.9
H 1.64 $\pm$0.96 36.5 $\pm$4.1 10.27 $\pm$0.80 27.3 $\pm$1.7 19.90 $\pm$2.02 17.4 $\pm$0.5
K$_S$ 3.42 $\pm$0.20 30.3 $\pm$0.3 0.80 $\pm$1.06 -46.3 $\pm$33.0 11.14 $\pm$0.70 -58.0 $\pm$0.4
178-342 J 17.12 $\pm$3.77 50.7 $\pm$0.6 20.48 $\pm$2.03 15.4 $\pm$1.6 25.18 $\pm$2.87 7.5 $\pm$0.2
H 6.30 $\pm$2.07 -41.3 $\pm$35.5 11.02 $\pm$3.93 16.2 $\pm$4.8 23.94 $\pm$1.99 6.4 $\pm$0.6
K$_S$ 11.11 $\pm$3.19 63.5 $\pm$4.4 0.36 $\pm$1.14 87.5 $\pm$30.6 3.14 $\pm$1.08 -48.2 $\pm$34.2
178-343 J 11.08 $\pm$0.91 59.2 $\pm$4.3 17.64 $\pm$2.59 25.8 $\pm$1.6 23.57 $\pm$1.01 15.4 $\pm$0.5
H 11.94 $\pm$3.99 -46.3 $\pm$0.4 11.94 $\pm$3.03 23.5 $\pm$4.2 24.00 $\pm$3.99 8.3 $\pm$0.6
K$_S$ 10.14 $\pm$3.25 87.2 $\pm$0.5 2.35 $\pm$0.73 -39.6 $\pm$25.6 8.27 $\pm$1.35 -62.6 $\pm$0.6
------------- -------- ------- ----------- ------- ----------- ------- ----------- ------- ----------- ------- ----------- ------- -----------
{height="\textwidth"}
In Fig. \[res\] the polarization vectors patterns of 177-341 are shown. Close to the central source in the dust envelope, a centrosymmetric pattern is visible in relatively low polarization in the range of 5-10%. The reflection nebula in the outer regions of the proplyd shows a uniformly oriented polarization pattern with the same direction in all bands. This is almost perpendicular to the direction of $\theta^1$ Ori C, which is most likely the source of illumination of this part of the nebula.
Model and discussion {#model}
====================
Polarization properties {#polprop}
-----------------------
Circumstellar accretion disks mark optically thick regions of the reflection nebula where the polarization patterns clearly depart from centrosymmetry. The vectors in these regions usually show low amplitudes of polarization of about 10% due to multiple scattering [@1988ApJ...326..334B]. In the aperture of 0.35” for J-band to 0.5” for K-band, the centrosymmetric pattern is visible to the left in Fig. \[res\]. A region of shortened and aligned vectors is not clearly resolvable in this view.
The outer regions show higher polarization and an aligned pattern from the illuminated reflection nebula from $\theta^1$ Ori C. The extent of the polarization pattern around 177-341 is determined in the radial plots to the right of Fig. \[res\] as follows. The fraction of polarized light as a function of the distance from the star shows a plateau of low polarization of the circumstellar envelope and is increasing for higher radii due to the nebula polarization. For a perfect centrosymmetry, the profiles along the position angle of the Q- and U-images should resemble a cosine. The amplitude of the cosine fitted to the Q- and U-profiles as a function of the distance then clearly defines the extent. The maximum radius increases from J- to K-band and a small plateau around the maximum extent is visible in all bands.
For a closer look at the spatial proportions, we examined the profile of the Q- and U-images by adding 5-degree bins over the determined radial range. These azimuthal Q-profiles show the characteristic cosine or W-curve for a dust envelope (Fig. \[res\], middle). In our data the profiles are not symmetric around zero as they should be, and this is equivalent to an offset in polarized flux and is related to effects of the PSFs. The long distance of about 26” from 177-341 to the AO guide star $\theta^1$ Ori C (Table \[tabprop\]) results in elongated PSFs for the observations . The direction of elongation is towards $\theta^1$ Ori C (134$^\circ$ P.A.). A shifting of the profiles can be seen, after convolving simulated polarization images with the PSF of the observations (Figs. \[shiftvectors\] and \[elongeffects\], only available electronically); therefore, the profiles of the convolved images of the data and simulations were shifted to be symmetric around zero so as to achieve a result closer to the actual centrosymmetric polarization pattern. This can be seen as a bootstrap calibration of the polarimetric images on the basis of the centrosymmetry of the envelope polarization. This has an impact on the pattern of the surrounding reflection nebula, which is highly polarized, but the total flux is low. The relative changes in the region of centrosymmetry are irrelevant due to the higher flux, while the analysis of the individual Q- and U-images in this study is not affected.
Since the Q- and U-images are not acquired simultaneously, there are deviations due to residual speckles and changes in seeing in addition to the elongated PSFs. These errors are emphasized during the determination of $\Pi_L=\sqrt{Q^2+U^2}$. As a consequence, no clear features are seen in the polarization degree images, and no direct comparison to convolved simulation images is possible. Hence, we tried to deconvolve the observational data in this case. For the PSF estimation the FOV with 177-341 was analyzed with the StarFinder program . Three point-like sources around 177-341 were used to extract a field PSF (Fig. \[psfestim\], only available electronically). For the deconvolution of the individual intensity images $I_\phi$, the iterative Lucy-Richardson [@1974AJ.....79..745L] algorithm was chosen, which is known to be robust and stable even under difficult SNR conditions [@1998ASPC..145..496P]. For the beam reconstruction, the deconvolved $I_\phi$-images were convolved with a Gaussian of 0.25” FWHM, before the polarization degree images were generated. The diffraction limit at K-band is 0.15” for the 3.6m telescope, but under moderate seeing conditions and these low Strehl ratios of the observations (Sect. \[chap:obs\]), an empirical limit for the achieved resolution is 1.5–2 diffraction limit units [@1997SPIE.3126..589B]. Without this smoothing, the deconvolved images are also dominated by high-frequency noise features enhanced progressively with the number of iterations . The algorithm converges fastest at the brightest source structures [@1974AJ.....79..745L]. Therefore choosing too few iterations results in images of undefined resolution and quality. From our extensive experience with the algorithm [@1999ApJ...523..248O], we chose $10^3$ iterations for the case of a single star dominating a single circumstellar structure.
{height="\textwidth"}
For the position angle of the disk, we compared the HST optical images and simulations to the polarization degree maps, obtained from deconvolved intensity images (Fig. \[pa\]). The parameters of the simulation are discussed in Sect. \[simchap\], but here a model with highly polarized envelope is shown to illustrate the disk signature in the envelope signal. In the optical images a small portion of the disk is seen in silhouette around the central star. Almost the same P.A. is found as an elongated region of low polarization in the NIR polarization images and can be quantified to $\sim144^\circ$. The diameter of the disk is about 300AU in this view. It must be stated that the direction of elongation of the PSFs differs only $10^\circ$ from the obtained P.A. and the elongated low polarization region could just be an artefact. But the extent of the region of 300AU clearly exceeds the PSF dimensions at least in the K-band. The P.A. is also confirmed by the following analysis.
There is a noticeable modulation of the amplitudes of the profiles, which can constrain disk parameters if we consider the optically thick accretion disks. In regions where the disk lies in front of the illuminating source from the observer’s point of view the amplitude should be attenuated. The same applies where the disk blocks the propagation of the central light to outer regions of the envelope. This can be associated with the amplitude modulation (AM) technique used for communication and signal transmission. The envelope scattering is the carrier signal, i. e. a cosine with fixed phase in the Q- and U-profiles. Since there is no reference amplitude and we have only two cycle periods (see Fig. \[res\]), only the ratio between the two can be determined as the modulation signal. To investigate these effects, we generated amplitude ratio maps, which are defined in polar coordinates as follows. For every position angle and radius ($r_0$, $\phi_0$) the Q-profile over a range from ($r$ ,$\phi_0-90^\circ$) to ($r$, $\phi_0+90^\circ$) is fitted with a function $f(\phi)=A_0+A\cos{(2\phi-\phi_Q)}$ with free amplitude $A$ and amplitude offset $A_0$, but fixed phase offset $\phi_Q=0^\circ$ and $\phi_U=45^\circ$, respectively. The best-fitting amplitude is divided by the one of the remaining $180^\circ$ (for clarification see Fig. \[res\], middle: dash-dotted lines). The cosine-fitting incorporates the model of the fixed carrier signal amplitude. The individual fitting over the radius $r$ allows arbitrary radial density and polarization profiles for the envelope, but rotational asymmetries like flattened envelopes are not considered. The result has an advantage over the pure Q-/U-profiles, since the attenuation of the signal is determined over the whole position angle, and the SNR is increased under the assumption of centrosymmetric envelope profiles. In principle the achievable spatial resolution is also enhanced by the high sensitivity of the signal to modulations. The results are shown in Figs. \[ratioexplan1\]–\[ratioexplan3\] for models with a flat and a flared disk, respectively. The behavior of the maps for the unconvolved images is as expected and can separate the disk signal from the envelope signal to some extent. A ratio map of a flat disk indicates the inclination angle. In this case the disk silhouette forms an ellipse with semi-major axis $a$ (the disk radius) and semi-minor axis $b$, so the inclination is given by $\xi=\arccos{(b/a)}$. Because the radial extent of the modulation in the amplitude ratio map refers to $b$, the inclination can be determined. However, if the images are convolved with elongated PSFs, the Q- and U-maps perform differently, depending on the P.A. of the disk. For a P.A. of $0^\circ$, the Q-map is distorted, because of the low amplitude of the carrier in the region of the modulation by the disk (Fig. \[ratioexplan3\], left panel, e.g.). The same applies for a P.A. of $\sim 45^\circ$ and the U-map. With a low-amplitude signal, PSF elongation effects become dominant.
-----------------------------------------------------------------------------
{width="48.00000%"} {width="48.00000%"}
-----------------------------------------------------------------------------
-----------------------------------------------------------------------------
{width="48.00000%"} {width="48.00000%"}
-----------------------------------------------------------------------------
-------------------------------------------------------------------------------
{width="48.00000%"} {width="48.00000%"}
-------------------------------------------------------------------------------
![Plot of the test of the inclination of the disk: the amplitudes of the polarization profile (see Fig. \[res\]) are attenuated due to the multiple scattering in the disk. If the disk is inclined, the amplitude of the part with the disk in front is more affected. Therefore the ratio of the amplitudes (“in front” divided by “behind” ) is minimal. Here the maps of the Q- (top) and U-amplitudes ratio in J-, H-, and K-bands (from left to right) are shown in an annulus of 0.1”-0.25”. The position angle for the disk can be located around 144$^\circ$. The radius of the minimum spots ratio along the line perpendicular to the disk can be determined to $\sim$0.12” in Q (H- and K-bands). []{data-label="posan"}](7555fig12){height="50.00000%"}
The maps of the amplitude ratio of the observations in J-, H-, and K-bands are shown in Fig. \[posan\]. In the Q-maps we find extremal ratios in the H- and K-band at a position angle of $144^\circ$ of the disk, so the Q-map should be less affected by elongation effects. The modulation in the U-maps is not significant – note the individual scaling of each band – and is interpreted as PSF- and noise effects. This is clear evidence that the disk is inclined and the portion in front of the star is located in the south west. Moreover, the radius $b$ of the minima can be determined precisely to 0.12” in the H-band and in K-band (Figs. \[posan\] and \[res\] right, dash-dotted line). If a radius of $a=0.35''$ and a flat disk are assumed, the shape of the ellipse is set and the minimum inclination angle is given by $\xi=\arccos{(b/a)}\approx70^\circ$. Considering the height of the (flared) disk, this value is increased. This can be seen in Fig. \[ratioexplan3\], where the flared disk produces a greater extent as the flat disk for the same inclination. This results in an estimated possible inclination range of about 75$^\circ$–85$^\circ$, as the disk has to be inclined according to the signature in the ratio maps. The values of maximum ratios increase from the J- to the K-band. This can be interpreted as a stronger relative attenuation of the disk at higher wavelength. Since the polarization is lower in K-band, the total attenuation might be constant.
Modeling {#simchap}
--------
There are two main processes for producing NIR/optical polarization: (i) scattering and (ii) differential absorption by non-spherical dust grains with the short axis primarily aligned along the local magnetic field, referred to as dichroic extinction [@1988ApJ...326..334B]. In the case of a circumstellar envelope and disk, the first is relevant. To interpret the polarization data, we used a 3D radiation transfer code, based on the Monte Carlo method [@1992ApJ...395..529W; @1993ApJ...402..605W]. The model parameters of the presented simulations, which illustrate the signature of the disk in the polarization images (Figs. \[pa\]–\[posan\]), are described in the following. In addition different models for grain size distributions and the impacts on the polarization images are discussed as an attempt to explain the increasing polarization pattern from the J- to the K-band.
The standard models for low-mass star formation include a slowly rotating, isothermal dense molecular cloud that collapses. The material with high angular momentum forms the circumstellar disk, while the one with lower angular momentum falls into the proto star. We adopted the model of an infalling proto-stellar cloud described by [@1984ApJ...286..529T], often referred to as the TSC-model. Our configuration includes both a circumstellar envelope and a disk. We skipped a detailed modeling of bipolar cavities, since we do not see any clear evidence in the data. As mentioned in [@1993ApJ...402..605W], an infalling envelope can easily produce the specific polarization pattern of reflection nebulae, which are often interpreted as scattering from a disk. As a consequence we expect a highly inclined disk in the case of 177-341 as suggested by the optical images (Fig. \[pa\]), although we see a round, nearly centrosymmetric pattern in the polarimetric images. The critical parameter for reproducing the pattern is the mass infall rate in the TSC-model. As the envelope size is limited to about 300AU in our images (Fig. \[res\]), the centrifugal radius is determined to 150AU in the model.
------ -- ---------- ---------- ------ ------------------ -- ---------- ---------- ------ ------------------ -- ---------- ---------- ------ ------------------
Band $\kappa$ $\omega$ $g$ $P_{\mbox{max}}$ $\kappa$ $\omega$ $g$ $P_{\mbox{max}}$ $\kappa$ $\omega$ $g$ $P_{\mbox{max}}$
J 65 0.42 0.16 0.81 63 0.46 0.32 0.58 48 0.60 0.34 0.58
H 38 0.33 0.06 0.91 38 0.42 0.29 0.59 33 0.56 0.31 0.59
K 20 0.21 0.03 0.94 22 0.36 0.25 0.60 20 0.50 0.27 0.60
------ -- ---------- ---------- ------ ------------------ -- ---------- ---------- ------ ------------------ -- ---------- ---------- ------ ------------------
The unit of $\kappa$ is \[cm$^2$g$^{-1}$\].
The dependency of polarization intensity on the wavelength refers to the effective grain size in the dust envelope, which scatter and polarize the emission from the central star . The polarization spectrum and peak polarization also depend strongly on the albedo $\omega$, the opacity $\kappa$, and the asymmetry parameter $g$ of the grains. The latter defines the relation between forward scattering and isotropic scattering. For the grain size distribution and parameters, we adopted the models from [@1977ApJ...217..425M], hereafter MRN and [@1994ApJ...422..164K], KMH. Table \[graintab\] gives an overview of the differences between the two commonly used models in the NIR (data from [@1994ApJ...422..164K] and [@1997ApJ...485..703W]). In MRN, grain sizes are limited to $0.25\mu m$, which is small compared to the wavelength of NIR-light, so the albedo and opacity drop for increasing wavelength. The KMH-model uses grain parameters obtained from fits to the interstellar extinction law. Local variations in the interstellar extinction curve are parameterized with the ratio of total to selective extinction $R_V\equiv A(V)/E(B-V)$. While the standard KMH deals with the diffuse interstellar medium ($R_V=3.1$), the alternative (here denoted as KMH$_{DC}$) is calculated for a dense cloud region with $R_V=5.3$. [@1981ApJ...244..483M] stated that (i) an increasing $R_V$ implies an increasing size of particles, and (ii) $\lambda_{\mbox{max}}$, the wavelength of maximum polarization, is larger for regions with $R_V$ larger than the average $3.1$. The peculiar extinction of $\theta^1$ Ori C is specified with $R_V=5.5$. From the fit to the modified extinction curve in KMH$_{DC}$, the $R_V=5.3$ distribution contains significantly fewer small- and intermediate-sized particles and a modest increase in larger sizes [@1994ApJ...422..164K]. This implies both a higher albedo and slightly greater $g$ and therefore an increased forward scattering. The extinction is less at short wavelengths and flattened towards higher wavelengths. We also tested a modified model with different grain properties only for the disks region. If grain growth in the disk is achieved, the scattering properties here should differ from the envelope. We adopted calculations from [@1990ApJ...349..107P] for large grains with a $\sim$4 times higher opacity and higher albedo, but also implemented a wavelength dependency, so that the scattering in K-band is increased. With high opacities and thus a high degree of multiple scattering, photons could be reprocessed by the disk more easily. Therefore more photons would be available in the outer regions of the envelope, so the scattering toward the observer could be enhanced.
The parameters for the envelope and the disk are chosen as determined from the data in the previous Section and listed in Table \[tabmc\]. This includes the envelope and disk size, and the inclination and P.A. of the disk. The mass infall rate for the envelope was probed in a wide range of $10^{-9}$–$10^{-6}$ $M_\odot y^{-1}$ to vary the visible extent of the polarization pattern. A value of about $10^{-9}$ gives the best results for the average radius of 0.48”. Such low infall rates would refer to a low-mass proto star in a transition between late Class I (embedded in envelopes) and Class II (with T Tauri disks) [@2003ApJ...598.1079W].
Parameter Value
------------------------------- ---------------------------------
Number of photons 10.000.000
Stellar radius $R_S$ 1.5$R_\odot$
Disk mass $M$ 0.1$M_\odot$
Outer disk radius 150AU
Inner disk radius 4$R_S$
Inclination angle $75^\circ$
Position angle $144^\circ$
Scale height of disk $h$ 0.001$R_S$
Inner envelope radius 10$R_S$
Outer envelope radius 150AU
Mass infall rate for envelope $1\cdot 10^{-9} M_\odot y^{-1}$
: Parameters used in Monte Carlo simulation.[]{data-label="tabmc"}
The spatial dimensions are derived from the observational data as described in Sect. \[polprop\].
The simulation images are convolved with the PSFs from the observational data to compare the results. The convolution with the elongated PSFs causes a change in the polarization properties: an offset of polarization of the position angle in the direction of the elongation is now visible (Fig. \[shiftvectors\], only available electronically). This concurs with a shift in the Q- and U-profiles as seen in the centrosymmetric regions of the data (Fig. \[res\] and Sect. \[polprop\]). The amount of offset depends on the specific model parameters, particularly the densities of disk and envelope and the disk inclination.
![Results from simulations with different grain models in comparison to the observational data. The polarization extent in arcsec as a function of wavelength is shown (solid line, determined as in Fig. \[res\]). The standard models MRN (dotted line) and KMH (dashed line) show a reversed spectral dependency. KMH$_{DC}$ (dash-dotted lines) for a dense cloud is flattened. []{data-label="simres"}](7555fig13){height="45.00000%"}
Discussion
----------
[@1999AJ....118.2350H] derived an inclination angle of the disk of 177-341 of 75$^\circ$–85$^\circ$ by fitting photo-evaporation model data to emission line profiles of spectroscopic data. [@2002ApJ...578..897R] modeled the SED for a photo-ionized proplyd and found that the disk emission is bright in the infrared and comparable to the central star. They also propose a highly inclined disk configuration for 177-341 by comparison to 10$\mu$m-images. [@2000AJ....119.2919B] interpreted HST -images of 177-341 as a disk seen in silhouette against the bright IF. They propose a P.A. of $105^\circ$, although the silhouette suggests a P.A. of about $135^\circ$ to $145^\circ$ because of the compact source about 0.5” east of this proplyd, which is likely to be an ultra compact Herbig-Haro object (177-341b in Fig. \[ifoptn\]). performed a deep analysis on the HST data and determined the disk size from the IF diameter. For a distance of 0.01–0.3 pc to $\theta^1$ Ori C the FUV (far-ultraviolet) photons dominate the dissociation process and the disk radius is typically $r_d \gtrsim r_{IF}$ [@1998ApJ...499..758J; @1999ApJ...515..669S]. In the case of 177-341 the diameter of the IF is $\sim$0.8" or 350 AU.
Considering these constraints on the disk and envelope configuration we find good agreement of the parameters with our polarimetric analysis. The compliant diameter and P.A. justify the deconvolution process and the introduced polarization amplitudes ratio maps. The inclination range obtained is also consistent with previous estimates. In the special case of a polarized envelope and a circumstellar disk, polarimetric differential images can be interpreted with this technique even under poor SNR and AO conditions. With the obtained values for the envelope size and the size, P.A. and inclination of the disk the simulations show similar polarization images.
According to a common model of the eroding disk mechanism [@2003ASPC..287..263B], first the outer parts of the disk are being destroyed. One way for disks to survive the photo-evaporation by UV-radiation is for the dust grains to grow to sizes with radii $\gtrsim1$cm. Evidence of grain growth in circumstellar disks in Orion has previously been found, because the outer portions of the giant disk of 114-426 contains grains $\gtrsim2\mu$m [@2003ApJ...587L.109S]. The efficiency of grain growth is predicted to be highest in the center of the disk where the highest densities and temperatures are found and photo-evaporation does not operate efficiently [@2001Sci...292.1686T].
However our comparison of different dust models cannot explain the peculiar spectral polarization of 177-341. The models for dense clouds with high $R_V$ result in improved scattering at longer wavelengths, but do not reproduce the increasing polarization pattern extent from the J- to the K-band (Fig. \[simres\]). Also a simple model (KMH$_{DC}+$) of large grains in the disk shows only a moderate increase in the extent. As a conclusion, advanced modeling is required including significantly different dust properties. In particular the opacity distribution within the envelope might be worth a detailed investigation in future model improvements. [@2008ApJ] recently showed discrepancies between the extent of dust continuum emission and molecular gas emission, suggesting a different shape for the edge of the disk. An alternative disk model with a tapered edge agrees more with observations than the commonly used truncated power-law model.
Summary and conclusions
=======================
![Schematic overview of our model and results for the proplyds 177-341 and 178-342. 177-341 has an inclined disk and an illuminated dust envelope. The diameter, inclination, and P.A. of the disk and the size of the envelope are drawn according to the results of the polarimetric analysis. []{data-label="cart"}](7555fig14){width="45.00000%"}
We have presented polarized emission maps of the circumstellar dust distribution of the giant proplyd 177-341 and found a complex morphology including both a circumstellar envelope and an accretion disk. The size of the disk was estimated with the diameter of the IF and confirmed in the polarization degree images. We extracted the inclination of the disk from a detailed study of the polarization profiles in the Stokes images. For the P.A. of the disk, we presented evidence of part of the disk seen in silhouette in the optical images, and we find matching position angles in the polarization amplitude ratio and in the deconvolved polarization-degree images due to multiple scattering in the disk. An overview of the results of the envelope and disk configuration is given in Fig. \[cart\]. The size of the dust envelope is obtained from the polarimetric differential images (Fig. \[res\]) and increases from the J- (blue) to the K-band (red). In the H-band it is comparable to the diameter of the ionization fronts head in the optical (Fig. \[ifoptn\], right). 178-342 has no resolvable dust distribution, as the images in Fig. \[resrgb\] show and modulates the shape of the IF of 177-341 in the optical (Fig. \[ifoptn\]). The wavelength dependency of polarization intensity and the extent of the polarization pattern cannot be reproduced in the models by large grains in the circumstellar material. The comparison of common dust models shows evidence of the contribution of larger grain to polarization and scattering effects, but another mechanism must be present to explain the observational results.
We showed with our 4 m-class data that the technique of ground-based polarimetric differential imaging is suitable for mapping the dust emission very close to the central star (here about 0.1”). This is the case for the largest and brightest proplyds. Other proplyds close to the Trapezium might be worth observing with an 8 m-class telescope. The proplyds 158-326 and 166-316 are the most suited candidates, as these show extended emission and a modulated trefoil pattern in the polarimetric differential images. With the introduced PDI analysis techniques, we are able to constrain disk parameters, such as inclination, P.A., and size even when the disk structure itself is not resolvable.
This work was supported in part by the *Deutsche Forschungsgemeinschaft (DFG)* via grant SFB 494.
[^1]: ESO program-ID: 066.C-0219
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
author:
- 'P.M. Marrese [^1]'
- 'S. Marinoni'
- 'M. Fabrizio'
- 'G. Altavilla'
bibliography:
- 'marrese.bib'
date: 'Received ; accepted'
subtitle: Algorithms and results
title: '*Gaia* Data Release 2. Cross-match with external catalogues '
---
[Although the *Gaia* catalogue on its own is a very powerful tool, it is the combination of this high-accuracy archive with other archives that will truly open up amazing possibilities for astronomical research. The advanced interoperation of archives is based on cross-matching, leaving the user with the feeling of working with one single data archive. The data retrieval should work not only across data archives but also across wavelength domains. The first step for a seamless access to the data is the computation of the cross-match between *Gaia* and external surveys.]{} [We describe the adopted algorithms and results of the pre-computed cross-match of the *Gaia* Data Release 2 (DR2) catalogue with dense surveys (Pan-STARRS1 DR1, 2MASS, SDSS DR9, GSC 2.3, URAT-1, allWISE, PPMXL, and APASS DR9) and sparse catalogues (Hipparcos2, Tycho-2, and RAVE 5).]{} [A new algorithm is developed specifically for sparse catalogues. Improvements and changes with respect to the algorithm adopted for DR1 are described in detail.]{} [The outputs of the cross-match are part of the official *Gaia* DR2 catalogue. The global analysis of the cross-match results is also presented. ]{}
Introduction {#intro}
============
The *Gaia* satellite allows determining high-accuracy positions for $\sim$1.7 billion sources and parallaxes and proper motions for $\sim$1.3 billion sources observed all-sky down to magnitude G$\sim$20.7. Compared to the first intermediate *Gaia* Data Release (DR1, see @Brown2016 for a summary of the astrometric, photometric, and survey properties, and @Prusti2016 for the scientific goals of the mission), the second intermediate *Gaia* Data Release [@Brown2018] provides 48% additional sources, parallaxes, and proper motions with an unprecedented accuracy for 77% of all observed sources, which are complemented by a precise and homogeneous multi-band photometry and a large radial velocity survey for more than 7000000 sources with G magnitude in the $4-13$ range. Astrophysical parameters for $\sim$160 million sources, data on more than 500000 variable stars, and $\sim$14000 solar system objects are also available in DR2[^2].
The main goal of adding a pre-computed cross-match to *Gaia* DR2 data is complementing *Gaia* with existing astrophysical quantities (that are widely used by the scientific community). This allows the full exploitation of the scientific potential of *Gaia* .
The general principles of the adopted cross-match algorithm are given and discussed in @Marrese2017 (hereafter Paper I). We here briefly recall that any cross-match algorithm is a trade-off between multiple requisites, and a fraction of mismatched and/or missed objects is always present. Our aim is to define and implement a cross-match algorithm that on one hand should be general enough to be exploited for different scientific cases, and on the other should have complete results that can later be filtered to better fullfil a specific scientific problem. We tried to find a reasonable compromise between the completeness and correctness requirements, which implies that we needed to avoid adding too many spurious matches.
In Sections \[sec:gen\] and \[sec:details\] we describe the general principle and the details of the cross-match algorithms defined for *Gaia* DR2, respectively. Section \[sec:extcat\] contains the list of the external catalogues that we matched with *Gaia* DR2 data and a short description for the newly added catalogues, together with some issues or caveats that are relevant to the cross-match. In Sections \[sec:output\] and \[sec:results\] we describe and discuss the cross-match output content and the results. Finally, Appendix \[sec:app\] contains a discussion of the effective angular resolution of external catalogues and its influence on the cross-match.
*Gaia* pre-computed cross-match: general principles\[sec:gen\]
==============================================================
Following the same approach as in Paper I, we define the cross-match algorithm according to the scientific problem we are faced with. Since the cross-match results with external catalogues are part of the official *Gaia* DR2 and are integrated in the *Gaia* catalogue access environments, it is fundamental to match *Gaia* with each external survey separately and independently, in a consistent and homogeneous manner. We therefore created links between different surveys through the *Gaia* catalogue, which is all-sky and has the highest angular resolution. *Gaia* is thus at the centre of our cross-match schema, as depicted in Figure \[figure:XMscheme\].
![*Gaia* DR2 cross-match schema: for large dense surveys (orange arrows), *Gaia* is the leading catalogue, while for sparse catalogues (yellow arrows), the external catalogue leads. []{data-label="figure:XMscheme"}](figure1.png){width="0.92\linewidth"}
For multi-catalogue searches, the catalogue specific matches to *Gaia* that are common to different surveys can be selected using multiple joins.
The external catalogues to be matched with *Gaia* DR2 are all obtained in the optical/near-IR wavelength region (with the exception of allWISE, which extends in the medium-IR domain), are general surveys not restricted to a specific class of objects, and have an angular resolution lower than *Gaia*, as was the case for *Gaia* DR1. However, in contrast to the case of the cross-match of *Gaia* DR1, the external catalogues to be matched with *Gaia* DR2 are not sufficiently homogeneous among themselves for the exact same algorithm to be used for all of them. We therefore broadly separated the external catalogues into two different groups: large dense surveys, and sparse catalogues, and we defined two slightly different algorithms for the two groups. External catalogues are here defined as dense surveys when it is possible to define a precise (i.e. based on a reasonable number of objects) and accurate (i.e. local) density around the majority of their objects. The two algorithms we define are not symmetric, and for the dense surveys, we use *Gaia* as the leading catalogue, while for sparse catalogues, we use *Gaia* as the second catalogue.
The cross-match algorithms we use in DR2 are quite similar to the algorithm that was successfully used in DR1, however we could take advantage of the enormous increase in the number of sources with proper motions and parallaxes with respect to *Gaia* DR1, and we ameliorated the algorithm in many respects: *a)* use of the full five-parameter covariance matrix, *b)* improved density definition, *c)* source-by-source definition of the initial search radius, which allows matching high proper motion stars, and *d)* definition of the proper motion threshold to be used for *Gaia* sources with no proper motions based on a trade-off between completeness and correctness.
Similarly to what was done for DR1, in the *Gaia* DR2 cross-match algorithms, we have not defined any special treatment for binary stars so far. The binary stars that may represent a problem for the cross-match are physically related sources with an additional motion that is due to multiplicity, which can displace their positions enough to prevent them from matching. As a general principle, when we knew of an effect that influences astrometry (and thus the cross-match results) but there was no indication in *Gaia* data how strongly this would affect a specific source, we added a systematic to all affected sources, as we did when we broadened the position errors of *Gaia* sources without proper motions (see Subsection \[sub:epochdiff\]). However, when an effect influences the astrometry of a specific subsample of sources (such as binaries), but there is no information on which sources and how strong the influence is in the *Gaia* data, we assumed a more cautious attitude and only added a caveat stating that the effect was not taken into account.
We repeat here some basic definitions that are still valid in DR2, but can also be found in Paper I. A good neighbour for a given object in the leading catalogue is a nearby object in the second catalogue whose position is compatible within position errors with the target. We assume that when a good neighbour is found, it is the counterpart. When more than one good neighbour is found, the best neighbour (i.e. the most probable counterpart according to the figure of merit we define, see Section \[sec:details\]) is chosen among the good neighbours. Also for *Gaia* DR2, we produced two separate cross-match outputs: a BestNeighbour table, which lists the leading catalogue matched objects with their best neighbour, and a Neighbourhood table, which includes all good neighbours for each matched object (see Section \[sec:output\] for a detailed output description).
For dense surveys, the higher angular resolution of *Gaia* requires a many-to-one algorithm: therefore the algorithm we used is not symmetric and more than one *Gaia* object can have the same best neighbour in a given dense survey. Two or more *Gaia* objects with the same best neighbour are called mates. True mates are objects that are resolved by *Gaia*, but are not resolved by the external survey. For sparse catalogues (such as Hipparcos2, Tycho-2, and RAVE 5), where the external catalogue is the leading catalogue, a one-to-one match is forced and mates are not allowed. Additional good neighbours in *Gaia* for each sparse catalogue source can be found in the Neighbourhood output table.
The cross-match algorithms used for *Gaia* DR2 are positional and evaluate the second catalogue environment, like for DR1. However, for DR2, we exploit the full five-parameters covariance matrix calculated for the *Gaia* astrometric solution (@Lindegren2018, @Mignard2018) when it is available (i.e. 77% of sources).
Accounting for epoch differences\[sub:epochdiff\]
-------------------------------------------------
Cross-match algorithms are based on the comparison of source positions in different surveys. Surveys can have been obtained at different epochs, which can be decades apart, and sources often move appreciably in the meantime, therefore it is important to take the source motion into account.
![Cumulative distribution of the angular distance of correct best matches obtained with the broadening method using two different proper motion thresholds (50 mas/yr and 0 mas/yr) when compared to the proper motion propagation of positions method (red curve indicated by *All* label).[]{data-label="figure:PMthresh"}](figure2.pdf){width="0.96\linewidth"}
\[table:PMthresh\]
----------- ------------------ -------------------- ------------------ ---------------- -------------------------- ------------------------
Catalogue $N$ Best matches Proper motion $N$ Best matches % Best matches $N$ Correct best matches % Correct best matches
pos prop method threshold (mas/yr) broad method broad method broad method broad method
2MASS PSC 424265005 50 432794791 102.01 422382563 99.56
424265005 0 408841264 96.36 407427704 96.03
GSC 2.3 727460368 50 731416596 100.54 725395647 99.72
727460368 0 725035070 99.67 723152510 99.41
----------- ------------------ -------------------- ------------------ ---------------- -------------------------- ------------------------
\[table:Density\]
Catalogue $Radius_{\mathrm{max}}$ % sources % sources % sources % sources % sources
----------------- ------------------------- ---------------- -------------------- -------------------- --------------------- -----------------
(arcsec) $N_{stars}<10$ $10<=N_{stars}<30$ $30<=N_{stars}<50$ $50<=N_{stars}<100$ $N_{stars}=100$
*Gaia* DR2 300 0.000001 0.0002 0.11 7.24 92.65
Pan-STARRS1 DR1 120 0.0004 0.0326 0.21 30.58 69.18
GSC 2.3 480 0.0001 0.000007 0.0003 0.22 99.78
PPMXL 480 0.000001 0.000006 0.000026 0.51 99.48
SDSS DR9 600 0.000006 0.00004 0.00005 0.006 99.99
URAT-1 480 0.001 0.05 0.156 13.31 86.48
2MASS PSC 600 0 0 0.007 2.52 97.48
allWISE 480 0 0 0.000008 0.0068 99.99
APASS DR9 600 0.0061 1.52 6.27 28.81 63.40
In order to do so, we moved the *Gaia* objects to the individual epoch of the possible matches in the external catalogues using the algorithm provided in the Hipparcos and Tycho Catalogue documentation (@Hipparcos). While this algorithm requires the use of all six parameters, $\alpha$ (Right Ascension), $\delta$ (Declination), $\pi$ (parallax), $\mu_{\alpha*}$ (proper motion in $\alpha\cos\delta$), $\mu_{\delta}$ (proper motion in $\delta$), and $V_{\mathrm{R}}$ (radial velocity), $V_{\mathrm{R}}$ is not included in the published astrometric solution in *Gaia* DR2. Nonetheless, according to @Lindegren2018, and in particular their Section 3, $V_{\mathrm{R}}$ is relevant only for very few sources (53).
For the fraction of *Gaia* sources for which only a position (i.e. 2 parameters) astrometric solution is available, we applied the broadening method described in Paper I for the sake of completeness. We thus defined a proper motion threshold that is common to all sources and all external catalogues for homogeneity and consistency reasons. In DR1, the adopted threshold (200 mas/yr) was chosen by evaluating the distribution of known high proper motion stars. Instead, while we are aware that the peak of the total proper motion distribution for Gaia sources is $\sim$6 mas/yr, and with the aim of also recovering high proper motion stars in the subsample of *Gaia* sources with two-parameter astrometric solutions, we decided for DR2 to derive from the data which was the most appropriate proper motion threshold to use. We therefore considered the subsample of *Gaia* sources with a five-parameter astrometric solution, and we compared the cross-match results obtained using the position propagation method on one hand and the broadening method on the other. In this evaluation, we assumed *a)* that the subsample of *Gaia* sources without available proper motions has the same proper motion distribution as the subsample with measured proper motions, and *b)* that the result obtained using the position propagation is correct. We conducted different tests with the broadening method in order to determine a proper motion threshold that allowed maximising the number of correctly recovered matches and minimising the addition of spurious matches. The tests were performed on all catalogues using different thresholds. The number of sources recovered in the cross-match output for a given external catalogue depends on the combination of the typical epoch difference between the external catalogue and *Gaia* DR2 and the typical size of the position errors of the external catalogue. The larger the epoch difference and the smaller the position errors, the larger the number of recovered sources and thus the more relevant the position error broadening. On the other hand, the denser the external catalogue, the larger the number of added spurious matches.
Table \[table:PMthresh\] and Figure \[figure:PMthresh\] illustrate the method we used and show the comparison of the cross-match results for 2MASS PSC and GSC 2.3 between the position propagation method, the broadening method with the adopted 50 mas/yr threshold, and the method without position propagation (i.e. broadening threshold 0 mas/yr). In order to describe how the position error broadening method works, we chose two catalogues: 2MASS, for which the method gives a good improvement in the number of matched sources, and GSC 2.3, for which the improvement is less relevant. In the case of 2MASS, the typical epoch difference is $\sim$15 years, which combined with a typical *Gaia* total proper motion of 6 mas/yr, implies a $\sim$0.09 arcsec displacement. This displacement must be compared with the 2MASS position errors, which for most of the sources, are smaller than 0.1 arcsec. In the case of GSC 2.3, instead, the typical epoch difference with *Gaia* is $\sim$25 years, which implies a displacement due to proper motions of about 0.15 arcsec. This displacement is small compared with the 0.3$-$0.4 arcsec values of the typical GSC 2.3 position errors.
Therefore, the adopted proper motion threshold for DR2 is 50 mas/yr. This is our best compromise between completeness and the quantity of spurious matches added to the cross-match.
Environment\[subsec:environment\]
---------------------------------
As discussed in Paper I, the cross-match is not only a source-to-source but also a local problem, thus the figure of merit used to evaluate the good neighbours [and to choose the best neighbour among them]{} should also take into account the local surface density of the second catalogue. The density is thus included in the adopted figure of merit (FoM, see Subsection \[subsec:fom\]), and its precision (which depends on the number of sources used to obtain it) has an important influence on the FoM precision. Ideally, $\sim$100 sources are required to evaluate the FoM with a good precision, while $\sim$30 sources are still acceptable. The radius needed to obtain the minimum number of sources is instead a measure of the accuracy of the density and consequently of the FoM. A more local determination is indeed more accurate, especially in dense fields, where there are density variations on small scales and where the FoM is more important as it is used to select the best neighbour among an higher number of good neighbours.
For *Gaia* DR2, the local density was pre-calculated around each second catalogue source and was fed to the cross-match algorithm. We used a *K*-nearest method that aims to determine the radius at which the 100th nearby source is found. We also set a maximum radius to search for nearby sources that depends on the catalogue number of sources weighted by its sky coverage. The reason we defined a maximum radius is that we consider an accurate (i.e. local) density more important than a precise density, but computation performances were also taken into account. When the algorithm reached the maximum radius threshold, the corresponding star number was used to calculate the density, even if it was lower than 100. Table \[table:Density\] allows determining for each catalogue, including *Gaia*, the fraction of sources with a sub-optimal density determination.
The density determination is improved for DR2 with respect to DR1. However, it is a compromise just like many other details of the cross-match algorithm described in this paper, specifically, a compromise between accuracy and precision.
*Gaia* pre-computed cross-match: details\[sec:details\]
=======================================================
{width="0.96\linewidth"}
\[table:ExtProp\]
----------------- ------------- ------------------------- ---------------------- ------------------------------- ------------------------- -------------
Catalogue $N$ Sources $PosErr_{\mathrm{max}}$ Effective resolution $\Delta Epoch_{\mathrm{max}}$ $SysErr_{\mathrm{max}}$ Survey type
(arcsec) (arcsec)
*Gaia* DR2 1692919135 0.4
Pan-STARRS1 DR1 2264263282 1.0 $\sim$1.1 18.02 0.18 Dense
GSC 2.3 945592683 1.6 $\sim$2 62.79 0.63 Dense
PPMXL 910468688 1.342 $\sim$2 15.5 0.155 Dense
SDSS DR9 469029929 10.0 $\sim$0.7 16.79 0.17 Dense
URAT-1 228276482 0.429 $\sim$2.5 3.189 0.03 Dense
2MASS PSC 470992970 1.21 $\sim$2.5 17.29 0.173 Dense
allWISE 747634026 35.944 6.1, 6.8, 7.4,12.0 5.47 0.055 Dense
APASS DR9 61176401 2.359 $\sim $5 3.5 0.035 Dense
Hipparcos2 117955 0.1684 $\sim $0.3 24.25 0.2425 Sparse
Tycho-2 2539913 0.254 $\sim$0.8 24.275 0.2475 Sparse
RAVE 5 457555 0.6 3.5 15.5 0.155 Sparse
----------------- ------------- ------------------------- ---------------------- ------------------------------- ------------------------- -------------
We recall here the basic details of the cross-match algorithm and outline the differences with DR1. As described in Paper I, the cross-match algorithm uses a plane-sweep technique that requires the catalogues to be sorted by declination, implies the definition of an active list of objects in the second catalogue for each leading catalogue object, and allows reading the input data only once, which speeds up the cross-match computation (@Power2005, @Abel2004, @Devereux2004, @Power2004). We used the same filter and refine technique as in DR1, but the first filter is now defined on an object-by-object basis (i.e. it is different for each target within a given leading catalogue), rather than being fixed for a given pair of leading and second catalogues, and it is only used to select candidate good neighbours and not to calculate the density on-the-fly. The second filter is used to select good neighbours among the candidates. The selection of the best neighbour among good neighbours is based on the same FoM as described in Paper I. A normal distribution for position errors is still assumed, and the position error ellipses are projected on the tangent plane.
Initial search radius (first filter)
------------------------------------
In the following, subscript *L* stands for leading catalogue and subscript *S* stands for second catalogue. The definition of the initial search radius ($R_{I}$) depends on the position in the cross-match algorithm of the *Gaia* catalogue as leading (i.e. dense surveys cross-match) or second catalogue (i.e. sparse catalogues cross-match).
$R_{I}$ is computed around each leading catalogue object as $$R_{I} = H_{\gamma} \cdot PosErr_{\mathrm{L,max}} + \left(\frac{PM \cdot \Delta Epoch_{\mathrm{L,max}}}{1000} \right),$$ where $H_{\gamma} = 5$ corresponds to a confidence level $\gamma$ of 0.9999994267; $PosErr_{\mathrm{L,max}} $ is the combined position error for each L source with the maximum position error in the S catalogue; $\Delta Epoch_{\mathrm{L,max}}$ is the maximum reference epoch difference between the L source and the S catalogue; and $PM$ is the proper motion considered. The definition of $PM$ is different in different cases:
- [proper motion of the L source if *Gaia* is the leading catalogue and the L source has a five-parameter astrometric solution;]{}
- [proper motion threshold if *Gaia* is the leading catalogue and the L source has a two-parameter astrometric solution;]{}
- [maximum of the *Gaia* catalogue proper motions if *Gaia* is the second catalogue.]{}
The combined position error $PosErr_{\mathrm{L,max}} $ is now defined as $$\begin{split}
PosErr_{\mathrm{L,max}} =~&\max[ RAerr_{L},DECerr_{L}] + \\&\max[ \max(RAerr_{S}),\max(DECerr_{S}) ]
\end{split},$$ where $RAerr$ and $DECerr$ are the uncertainties in Right Ascension and Declination. The maximum epoch difference between the L source and the S catalogue being matched is defined as $$\begin{split}
\Delta Epoch&_{\mathrm{L,max}} = \\& \max \Big[~|\max(refEpoch_{L})-\min(refEpoch_{S})|~,\\& |\min(refEpoch_{L})-\max(refEpoch_{S})|~ \Big]
\end{split}.$$ In the above equations, $R_{I}$ is in arcsec, $PosErr_{\mathrm{L,max}}$ in arcsec, $PM$ is in mas yr$^{-1}$, and $refEpoch$ is in years.
Broadening of position errors\[subsec:broad\]
---------------------------------------------
While a detailed discussion of the broadening method is available in Paper I, we repeat here for clarity the equations defining the position error broadening: $$\begin{split}
&\sigma_{x_{G'}} = \sigma_{x_{G}} + SysErr_{x} = \sigma_{x_{G}} + PM \cdot \Delta Epoch/5 \\&
\sigma_{y_{G'}} = \sigma_{y_{G}} + SysErr_{y} = \sigma_{y_{G}} + PM \cdot \Delta Epoch/5
\end{split}
,$$ where G stands for *Gaia*. In DR2 we always broadened the *Gaia* position errors when a five-parameter astrometric solution was not available, regardless of whether *Gaia* was the leading or second catalogue.
Good neighbour selection (second filter)
----------------------------------------
In order to define the second filter, it was necessary to convolve the leading and second catalogue position errors. We refer to @pineau and to Paper I for a detailed definition and derivation of the position error convolution ellipse.
The second filter is based on the Mahalanobis normalised distance $K_{\gamma}$ (see Equation 9 in Paper I). $K^{2}_{\gamma}$ has a $\chi^2$ distribution with two degrees of freedom, and its adopted value corresponds to a value of the confidence level $\gamma$ of 0.9999994267, which in 1D is equivalent to 5$\sigma$. Good neighbours are defined as neighbours that fall within the ellipse defined by the confidence level $\gamma$. The second filter is thus defined as $$\label{eq:secondfilter}
\frac{d}{\sigma_{x_{C}} \sqrt{ 1 - \rho_{C}^{2}} } \leq K_{\gamma},\\$$ where $d$ is the angular distance, $\sigma_{x_{C}}$ is the convolution ellipse error in the direction from the leading catalogue object to the possible counterpart, and $ \rho_{C}$ is the correlation between $\sigma_{x_{C}}$ and $\sigma_{y_{C}}$. The high-confidence level was chosen in order to improve the completeness of the cross-match.
Best neighbour selection: figure of merit\[subsec:fom\]
-------------------------------------------------------
The FoM we used to select the best neighbour among the good neighbours evaluates the ratio between two opposite models/hypotheses: the counterpart candidate is a match or it is found by chance. The FoM depends on the angular distance and the position errors, on the epoch difference, and on the local surface density of the second catalogue. For each of the good neighbours, we computed the FoM and the derived score, described in detail in Paper I. The score is listed in the Neighbourhood output table. The best neighbour is defined as the good neighbour with the highest score value.
External catalogue characteristics\[sec:extcat\]
================================================
Following is the list of external catalogues that were cross-matched with the *Gaia* DR2 catalogue and had already been matched with DR1:
- [GSC 2.3 (@GSC2.3)]{}
- [PPMXL (@PPMX [@PPMXL])]{}
- [SDSS DR9 primary objects (@SDSS9 [@SDSS12])]{}
- [URAT-1 (@URAT1) ]{}
- [2MASS PSC (@2MASS)]{}
- [allWISE (@WISE [@allWISE])]{}
Following is the list of the new external catalogues that were cross-matched with *Gaia* DR2:
- [Pan-STARRS1 DR1 (@panstarrs1 [@panstarrs1b; @panstarrs1c; @panstarrs1d; @panstarrs1e; @panstarrs1f])]{}
- [APASS DR9 (@apass9)]{}
- [Hipparcos2 (@Hipparcos [@Hipparcos2])]{}
- [Tycho-2 (@Hipparcos [@Tycho2])]{}
- [RAVE 5 (@rave5 [@RAVEON])]{}
The main properties to consider when matching the external catalogues with *Gaia* are *a)* the effective angular resolution, *b)* the astrometric accuracy, *c)* the celestial reference frame, either HCRF[^3] or Gaia-CRF2 [@Mignard2018], *d)* how the catalogue is tied to the International Celestial Reference System (ICRS), *e)* the coordinate epochs, *f)* the need of propagating astrometric errors when the catalogue proper motions are available and positions are given at epoch J2000.0, but errors on positions are given at mean epoch, and *g)* the known issues and caveats. It is also important to take into account how the external catalogue properties compare to the corresponding *Gaia* catalogue properties.
Table \[table:ExtProp\] lists the *Gaia* DR2 and external catalogues properties relevant to the cross-match. The effective angular resolution values reported in Table \[table:ExtProp\] were derived from the external catalogue reference papers or their on-line documentation. In some cases, the authors directly report the value of the effective angular resolution, in others, they list related quantities such as seeing, pixel scale, and the full width at half maximum of the point spread function (PSF FWHM), which can be used to derive the effective angular resolution. In Subsection \[subsec:resultangres\] we describe the effects of the difference in effective angular resolution. Appendix \[sec:app\] compares the effective angular resolution values reported in Table \[table:ExtProp\] with the actual content of the external catalogues. In some cases, the fraction of suspected duplicates is relevant. Figures \[fig:FigAngResa\] and \[fig:FigAngResb\] are useful to understand some details of the cross-match results (see Section \[sec:results\]).
Figure \[figure:all\] shows the sky coverage and the surface density distribution for *Gaia* DR2 and the external catalogues that are newly matched with *Gaia*. The corresponding figures for the external catalogues that had been matched before with *Gaia* DR1 can be found in Paper I. The surface density is calculated by counting the number of sources in each pixel obtained using a HEALPix tessellation: for dense surveys, we adopted a resolution of $N_{\mathrm{side}}=2^{8}$ , which has 786432 pixels with a constant area of $\Omega\sim188.89~$arcmin$^{2}$, while for sparse catalogues, we adopted a resolution of $N_{\mathrm{side}}=2^{6}$ , which has 49152 pixels with a constant area of $\Omega\sim0.8$ degree$^{2}$.
The external catalogue quantities used by the cross-match computations described in this study are positions, position errors, position error correlation (if available), and coordinate epochs. Different surveys may have a different definition of some of these quantities and/or use different units. The external catalogue input quantities were thus homogenised in order to simplify the cross-match calculations.
In the following we briefly describe the newly added external catalogues together with some caveats and known issues that are relevant for the cross-match computations. For catalogues that had been cross-matched with DR1, we describe some issues that were not apparent in DR1, but are relevant for the DR2 cross-match.
Pan-STARRS1 DR1
---------------
The Panoramic Survey Telescope and Rapid Response System (Pan-STARRS) is a system for wide-field astronomical imaging developed and operated by the Institute for Astronomy at the University of Hawaii. Pan-STARRS1 (PS1) is the first part of Pan-STARRS to be completed and is the basis for Data Release 1 (DR1). The PS1 survey used a 1.8 meter telescope and its 1.4 gigapixel camera to image the sky in five broadband filters (*g*, *r*, *i*, *z*, *y*). The version of the catalogue we used for cross-match computation is a filtered subsample of the $10\,723\,304\,629$ entries that are listed in the original ObjectThin table.\
We used only ObjectThin and MeanObject[^4] tables to extract what we needed. This means that objects that are detected only in stack images are not included. The main reason for avoiding objects detected in stack images (for cross-match purposes) is that their astrometry is not as good as the mean object astrometry, as stated in the Pan-STARRS1 DR1 documentation: “The stack positions (raStack, decStack) have considerably larger systematic astrometric errors than the mean epoch positions (raMean, decMean)”. The astrometry for the MeanObject positions uses *Gaia* DR1 as a reference catalogue, while the stack positions use 2MASS as a reference catalogue.
In detail, we filtered out all objects where
- [nDetections = 1;]{}
- [no good-quality data in Pan-STARRS, objInfoFlag 33554432 not set;]{}
- [mean astrometry could not be measured, objInfoFlag 524288 set;]{}
- [stack position used for mean astrometry, objInfoFlag 1048576 set;]{}
- [error on all magnitudes equal to 0 or to -999;]{}
- [all magnitudes set to -999;]{}
- [error on RA or DEC greater than 1 arcsec.]{}
The number of objects in the Pan-STARRS1 DR1 version used for cross-match is $2\,264\,263\,282$.
APASS DR9
---------
The AAVSO Photometric All Sky Survey (APASS) is obtained in five photometric bands: *B*, *V*, *g’*, *r’* and *i’*, and the observed targets cover the magnitude range $10<V<17$. APASS data are obtained with dual bore-sighted 20cm telescopes, designed to obtain two bandpasses of information simultaneously, from two sites near Weed, New Mexico, in the Northern Hemisphere and at CTIO in the Southern Hemisphere. The APASS DR9 contains approximately 62 million stars covering about 99% of the sky. The APASS project is being completed and DR9 is not a final release. According to the APASS documentation, there are some issues in the catalogue that should be taken into account when cross-matching it:
- The APASS team does not provide star IDs until the final product and suggests that stars are identified by their RA and DEC.
- There are a number of duplicate entries. These appear to be caused by the merging process, where poor astrometry in one field may cause two seed centroids to form for a single object.
- There are a number of entries with 0.000 errors.
- Centroiding in crowded fields is very poor; blends cause photometric as well as astrometric errors.
- There are saturated stars in the catalogue, and the APASS team suggests to avoid using sources brighter than $V=7$.
The issues described above are reflected in the quality of the cross-match results.
Given the lack of an identifier provided by authors, and because the Vizier TAP service[^5] is the only available resource for bulk download, we used the CDS *recno* as identifier, although we are aware that the record number assigned by the VizieR team should not normally be used for identification.
RAVE 5
------
The RAdial Velocity Experiment (RAVE) is a multi-fiber spectroscopic astronomical survey of stars in the Milky Way using the 1.2m UK Schmidt Telescope of the Australian Astronomical Observatory (AAO). RAVE contains multiple observations of the same source, which are identified with the same RAVEID and a different RAVE\_OBS\_ID. The number of entries in the catalogue is $520\,701$, while the number of distinct sources is $457\,555$. For cross-match calculations we used the distinct sources.
Hipparcos2
----------
Hipparcos2 is a new improved reduction of the astrometric data produced by the Hipparcos mission. The astrometric accuracies are much better (up to a factor of 4) than in the original catalogue.
Tycho-2
-------
The Tycho-2 catalogue is an astrometric reference catalogue containing positions, proper motions, and two-colour photometric data for the 2.5 million brightest stars in the sky. The Tycho-2 positions and magnitudes are based on precisely the same observations as the original Tycho catalogue collected by the star mapper of the ESA Hipparcos satellite, but Tycho-2 is much larger and slightly more precise, owing to a more advanced reduction technique. Components of double stars with separations down to 0.8 arcsec are included.
There are $109\,445$ sources in Tycho-2 without an astrometric solution. These objects are indicated by [*pFlag*]{}=X, where [*pFlag*]{} is the mean position flag. For these objects we used the observed Tycho-2 values for coordinates, coordinate errors, and reference epoch.
There are $13\,098$ sources in Tycho-2 for which [*pFlag*]{}=P. These objects are binaries (actually for 82 of them one of the two components is missing in the sample flagged with [*pFlag*]{}=P) and have different source Ids, but identical astrometry since the photocentre is used for the astrometric solution, which includes proper motions.
SDSS DR9
--------
A detailed description of the astrometric SDSS calibration is given in @SDSSQA, and a summary is provided in the on-line documentation[^6]. The *r* photometric CCDs serve as the astrometric reference CCDs for the SDSS. That is, the positions for SDSS objects are based on the *r* centroids and calibrations. The *r* CCDs are calibrated by matching bright stars detected by SDSS with the UCAC astrometric reference catalogues. The SDSS collaboration implemented an astrometry quality assurance (QA) system in order to identify errors in the SDSS imaging astrometry and provided a summary file[^7] containing all information about the SDSS field astrometry QA, including offsets from each of the reference catalogues. The astrometry QA summary file is available for download[^8]. The method we used to include the results of the above analysis in the cross-match algorithm is described in Section \[sec:results\].
Cross-match output\[sec:output\]
================================
The cross-match output consists of two separate tables: ** includes the best matches (selected as the good neighbour with the highest value of the score), while ** includes all the good neighbours (selected using the second filter, see Equation \[eq:secondfilter\]). The cross-match output datamodels are described in Tables \[table:BestD\] and \[table:NeigD\]. The content and some statistics of the ** and ** output tables for each external catalogue are summarised in Tables \[table:Be\] and \[table:Ne\].
Field name Short description
--------------------------- --------------------------------------------------------------
SourceId *Gaia* source identifier
OriginalExtSourceId Original External Catalogue source identifier
AngularDistance Haversine angular distance (arcsec)
NumberOfMates Number of mates in *Gaia* catalogue
NumberOfNeighbours Number of good neighbours in the second catalogue
BestNeighbourMultiplicity Number of neighbours with same probability as best neighbour
GaiaAstrometricParams Number of *Gaia* astrometric
parameters used
: BestNeighbour output table content.[]{data-label="table:BestD"}
Field name Short description
----------------------- -----------------------------------------------
SourceId *Gaia* source identifier
OriginalExtSourceId Original External Catalogue source identifier
AngularDistance Haversine angular distance (arcsec)
Score Figure of Merit
GaiaAstrometricParams Number of *Gaia* astrometric
parameters used
: Neighbourhood output table content.[]{data-label="table:NeigD"}
\[table:Be\]
----------------- -------- ---- ------- ----- ------- ----- -------- -- --
Catalogue
Pan-STARRS1 DR1 5.23 6 98.56 13 99.42 1 0
GSC 2.3 8.96 16 96.99 25 77.65 16 120344
PPMXL 4.02 8 89.35 14 87.78 2 2
SDSS DR9 52.10 6 99.49 80 99.77 3 2
URAT-1 2.12 3 99.99 3 99.84 1 0
2MASS 5.01 3 99.78 11 94.25 2 10
allWISE 181.15 3 99.99 24 98.39 1 0
APASS DR9 11.75 56 86.53 59 58.74 2 8
Hipparcos2 1.67 2 99.40 N/A N/A N/A 0
Tycho-2 1.94 3 99.51 N/A N/A N/A 0
RAVE 5 3.21 11 89.91 N/A N/A N/A 0
----------------- -------- ---- ------- ----- ------- ----- -------- -- --
\[table:Ne\]
----------------- -------- ------------- -------------- -- --
Catalogue
Pan-STARRS1 DR1 5.25 0.000000599 21.436935748
GSC 2.3 8.96 0.000002516 21.673957559
PPMXL 4.02 0.000004754 17.426431196
SDSS DR9 52.42 0.000000076 17.312649378
URAT-1 2.12 0.000088545 18.779691209
2MASS 5.01 0.000013321 13.538666301
allWISE 181.15 0.000000676 15.578377337
APASS DR9 11.75 0.000002632 16.908079945
Hipparcos2 1.74 0.000086919 20.415553214
Tycho-2 1.95 0.000029288 16.896918580
RAVE 5 3.44 0.000001223 8.424990229
----------------- -------- ------------- -------------- -- --
\[table:Stat\]
----------------- ------------ ----------- ------- ------- -----------
Catalogue
Pan-STARRS1 DR1 2264263282 810359898 80.49 35.68 816314072
GSC 2.3 945592683 870899123 51.44 80.96 884748168
PPMXL 910468688 716220357 42.31 73.50 757738601
SDSS DR9 469029929 113718207 64.46 24.22 114011744
URAT-1 228276482 188071510 27.32 82.32 188071646
2MASS 470992970 450688227 26.62 92.91 451193296
allWISE 747634026 300207917 17.73 39.83 300209602
APASS DR9 61176401 75018791 4.43 90.66 81278312
Catalogue
Hipparcos2 117955 83034 0.005 70.39 83283
Tycho-2 2539913 2475900 0.15 97.47 2482025
RAVE 5 457555 450587 0.027 98.48 474824
----------------- ------------ ----------- ------- ------- -----------
Results\[sec:results\]
======================
The cross-match results are part of the official *Gaia* DR2 release and are available at the ESA Gaia Archive[^9] and at Partner Data Centres Archives[^10]. The cross-match results are described in Table \[table:Stat\] and in Figs. \[fig:FigResulta\]- \[fig:FigAngDista\].
Given the size of the catalogues involved in this cross-match study, the analysis of the results can be performed only on general grounds, certainly not on an object-by-object basis. The aim of the following analysis is thus to give users information on the global characteristics of the cross-match results for a given catalogue, that is, sky and magnitude distributions of matched sources, distribution of angular distance of matched pairs, and fraction of matched sources.
In particular, the surface density maps displayed in the left column of Figure \[fig:FigResulta\] show the fraction of matched *Gaia* sources, while the maps in the right column show the fraction of matched external catalogue sources. These maps, combined with the corresponding maps available in Figure \[figure:all\] or in Paper I, allow the spatial analysis of the cross-match results. In the case of GSC 2.3 and PPMXL, the cross-match with the duplicated sources located at the plate borders results in an over-density of matched *Gaia* sources that is clearly visible in Figure \[fig:FigResulta\] as a square pattern.
The histograms in Figure \[fig:FigHist\] show the magnitude distribution of the matched external catalogue sources compared with the distribution of the full catalogue. Figure \[fig:FigHist\] can thus be used to assess the fraction of matched and missed external catalogue sources as a function of magnitude.
The angular distance distributions shown of cross-matched pairs in Figure \[fig:FigAngDista\] can be used to evaluate the global agreement of the external catalogue astrometry with *Gaia*. They can be used to retrieve information about the angular distance at which the bulk of the matched pairs are found (blue histograms) and about the angular distance within which all the matched sources are found (cumulative red curves). In addition, they also show no indication of the Poisson tail that is always present in cone search results: one of the advantages of a cross-match over a cone search is indeed that the search radius is defined on a pair-by-pair basis and is not fixed for all pairs. For example, even if in Figs. \[fig:FigResulta\] and \[fig:FigHist\] APASS DR9 and 2MASS cross-match results show similar behaviours, it is instead clear from Figure \[fig:FigAngDista\] that the 2MASS positions are in much better agreement with *Gaia* than the APASS DR9 positions. It is important to note that the histograms in Figure \[fig:FigAngDista\] are not a direct indication of the astromentric quality of the external catalogues. The main reason is that the histograms show only the matched sources, while astrometric issues in a catalogue often prevent the match of a fraction of the potential counterparts, leaving only the sources with good astrometry. In this study, by “good astrometry” we mean not only accurate positions, but also a careful evaluation of the position errors and the inclusion of systematics in position errors. The cross-match algorithms that require counterparts to be compatible within position errors easily highlight when position errors are underestimated. In the following we illustrate and discuss some specific features and characteristics of the cross-match results.
Effect of effective angular resolution differences on XM results\[subsec:resultangres\]
---------------------------------------------------------------------------------------
The comparison between the effective angular resolution of *Gaia* and of the external catalogues is very important for the cross-match. The higher *Gaia* angular resolution (which will improve with the forthcoming releases) implies that *Gaia* will frequently resolve sources that are unresolved in the external catalogue. The larger the difference in effective resolution between *Gaia* and the external catalogue, the more common the resolved objects. For this reason, since *Gaia* DR1 we chose a many-to-one algorithm for dense surveys and defined the mates, which are two or more *Gaia* sources with the same best neighbour in the external catalogue. The external catalogue sources that are the counterpart of two or more *Gaia* sources are thus very likely sources that are resolved in *Gaia*. The chances that mates correspond to a resolved object are obviously higher when all mates have proper motions available, and thus their positions are reliably propagated to the external catalogue epoch. A more subtle effect arises when the photocentre of the unresolved external catalogue source is too far from the corresponding two or more *Gaia* counterparts to allow a match within position errors. For these cases, a complex dedicated treatment is required. While the released *Gaia* DR2 data, and in particular the availability of accurate five-parameter astrometric solutions, allow addressing the angular resolution difference effects on cross-match, the solution is not trivial and requires carefully planned tests and a thorough analysis. A detailed treatment of this effect will be included in the cross-match of *Gaia* DR3 and subsequent releases.
Hipparcos2\[subsec:hipissue\]
-----------------------------
While we expect to find *Gaia* counterparts for most of the Hipparcos2 sources, with the exception of the brightest ones, the cross-match results include only $\text{about two-thirds}$ of them. This means that according to the adopted cross-match algorithm, only $\text{about two-thirds}$ of the Hipparcos2 objects have a *Gaia* counterpart that is compatible within the position errors (i.e. have at least one good neighbour). Hence the Hipparcos2 cross-match results clearly show an issue that needs to be investigated.
Around each Hipparcos2 object, we calculated a cone search with a fixed radius of 1 arcsec, which propagates the *Gaia* positions to Hipparcos2 epoch exactly in the same way as in the cross-match algorithm described in this paper. Then we selected the nearest neighbour. The cone search is thus consistent with the cross-match and allows us to make a direct comparison of the angular distance distributions obtained with the two methods that were used to determine possible counterparts. We defined two subsamples of Hipparcos2 sources. The first includes the cross-matched sources, and the second the additional associations that were found using the cone search. We then tried to identify a characteristic (either in Hipparcos2 or in *Gaia*) that could be used to separate the two samples and thus to understand the nature of the considered issue. The two samples are indistinguishable in terms of size of astrometric errors (see @Lindegren2018, Appendix A, and @Arenou2018 Subsection 4.6, for a detailed discussion of *Gaia* astrometric errors), magnitude or colour distribution, sky distribution, and many other quantities listed in the Hipparcos2 and *Gaia* catalogues. The only parameters that seem on average to allow separating the two samples are parameters related to the *Gaia* astrometric solution quality, for example the *astrometricGofAl* (goodness-of-fit statistics of the astrometric solution for the source in the along-scan direction).
The top panel of Figure \[Fig:HipIssue\] shows the angular distance distribution of the cross-matched sample (red histogram) and the sample of additional sources added with the cone search (blue histogram). The panel clearly shows that the blue sample Hipparcos2 sources are found at larger angular distances from their Gaia counterparts than the red ones. The blue sample associations are found at an average angular distance of 75.4 mas, while the red sample sources are found at an average angular distance of 13.8 mas. The middle panel of Figure \[Fig:HipIssue\] shows the *astrometricGofAl* distribution of the cross-matched sample and the sample of associations added with the cone search, but only for *Gaia* sources with a five-parameter astrometric solution. The bottom panel of Figure \[Fig:HipIssue\] shows the sky distribution of the *astrometricGofAl* averaged over healpix obtained with an HEALPix tessellation with resolution $N_{\mathrm{side}}=2^{8}$ for the *Gaia* catalogue sources with a five-parameter astrometric solution. The *astrometricGofAl* sky distribution allows a comparison between the values of the *astrometricGofAl* for the two samples with values of the *Gaia* DR2 catalogue.
The adopted cross-match algorithm does not account for the effects that arise because Hipparcos2 and *Gaia* DR2 have different reference frames, HCRF and Gaia-CRF2, respectively. According to Subsection 5.1 of @Lindegren2018, the global alignment of Gaia-CRF2 evaluated by the frame orientation parameters \[$\epsilon_{X},\epsilon_{Y},\epsilon_{Z}$\] at J2015.5 is constrained within $\pm$0.02 mas per axis for faint sources, and there is no indication of a misalignment larger than $\pm$0.3 mas per axis at the bright end. The Hipparcos2 misalignment at epoch J1991.25 is $\pm$0.6 mas per axis. Concerning the spin of the reference frame relative to the quasars, @Lindegren2018 confirmed that the faint reference frame of Gaia DR2 is globally non-rotating to within $\pm$0.02 mas/yr in all three axes. However, using a subsample of the Hipparcos2 sources present in TGAS (Tycho-Gaia Astrometric Solution, the subsample of *Gaia* DR1 sources with a five-parameter astrometric solution), the authors suggested that the bright (G$\lesssim$12) reference frame of Gaia DR2 has a significant ($\sim$0.15 mas/yr) spin relative to the fainter quasars. According to them, the most reasonable explanation is systematics in the Gaia DR2 proper motions of the bright sources. The effects of either the combination of HCRF and Gaia-CRF2 misalignments or the inertial spin of the Gaia DR2 proper motion system are too small when compared to the bulk of angular distances between Hipparcos2 sources associated with the cone search and their Gaia counterparts. These effects therefore cannot account for the bulk of the missing Hipparcos2 matches.
The cross-match is particularly critical when two catalogues with such small positional uncertainties are combined. The explanation for the one-third of Hipparcos2 sources without a *Gaia* counterpart compatible within position errors seems to reside in non-optimal astrometric solutions for part of the Hipparcos2 sources, as a result of astrometric perturbations that are probably caused by multiplicity, variability, and/or peculiarities.
Since we do have the a priori knowledge that we should match almost all the Hipparcos2 sources, we decided to add the result of the 1 arcsec cone search described above to this paper and make it available to users for download[^11]. The table contains three columns: the *Gaia* and Hipparcos2 identifiers, and the angular distance (in arcsec) for each nearest associated source. Table \[table:cone\] contains the first ten entries of the cone search results.
\[table:cone\]
--------------------- -------- -----------------------
5188150893900488576 48752 0.002032030044478262
5764614467999340032 71348 0.015010173408366224
5188178214189131008 42708 0.0033109008787438705
5764662880870489728 78866 0.006914112668101691
4611734916632361600 22645 0.002173230360375837
5188197627441445632 54065 0.24059879382091026
4611782058193541248 3560 0.0056529407401837315
6341351575677860992 90987 0.11695414581743536
5188247891443554688 40104 0.2276057244922324
6341181494973204096 104382 0.021732003615886262
--------------------- -------- -----------------------
: Sample of the cone search results described in Subsection \[subsec:hipissue\]
Tycho-2
-------
As described in Section \[sec:extcat\], for the cross-match computations we preferentially used the Tycho-2 set of coordinates propagated to epoch J2000.0. For a fraction of binary sources resolved by Tycho-2, however, the photocentre of the binary was used to obtain the astrometric solution (and thus the binary components have the same coordinates). Since Tycho-2 binaries have separations larger than $\sim$0.8 arcsec, they should also be resolved by *Gaia*, even if both components are not always present in the *Gaia* DR2 catalogue. In these cases, which involve $13\,098$ sources, the cross-match results are very poor, and we matched only 3744 sources. In these cases, both components are included in the cross-match output and will obviously both have the same *Gaia* counterpart(s). This problem will be addressed for DR3, when we will use the Tycho-2 observed positions, which are given separately for different components and allow a greatly improved number of binary matches.
GSC 2.3 and PPMXL
-----------------
The GSC 2.3 and PPMXL catalogues can be considered similar since they are both based on the same photographic plates, but PPMXL has a composite nature (see the Introduction of @PPMXL). The PPMXL coordinates that are available in the original catalogue were propagated to J2000.0. For cross-match purposes, we computed the position errors at J2000.0 using the position errors at mean epoch available in PPMXL. Nevertheless, the GSC 2.3 position errors are typically four times larger than the propagated PPMXL errors. According to @GSC2.3, GSC 2.3 position errors should be considered conservative estimates of the uncertainties. The epoch difference between *Gaia* and GSC 2.3 is $\sim$25 years on average, and it is 15.5 years for PPMXL.
Given the above, the cross-match results are quite different for the two catalogues. First of all, $\sim$81% of GSC 2.3 sources and only 73.5% of PPMXL sources have a *Gaia* counterpart. Figure \[fig:FigAngDista\] shows that when *Gaia* counterparts are found, PPMXL sources are closer than GSC 2.3 sources, but this does not mean that the PPMXL astrometry is better than GSC 2.3 astrometry. The longer the time interval for which a given *Gaia* source is propagated, the larger is the possible misplacement, due to proper motion uncertainties. This explains in part why GSC 2.3 counterparts are found at larger distances than PPMXL counterparts. On the other hand, since PPMXL positions are propagated to J2000.0 using PPMXL proper motions, when they are not accurate, the *Gaia* counterparts are less easy to find. The PPMXL small position errors also contribute to the counterpart matching failures, while the larger GSC2.3 position errors allow us to find counterparts at larger distances. The net effect is that *Gaia* counterparts are fewer but closer in PPMXL and more numerous but at larger distances in GSC2.3.
The similarity of the two catalogues instead accounts for the similar issue with duplicates at plate edges and for the similar secondary feature shown in Figure \[fig:FigAngDista\], roughly between 0.8 and 1.8 arcsec, which is due to the presence of mates in cases when one of the two different *Gaia* sources that share the same best neighbour in the external catalogue is much closer than the other. The duplicate issue is more evident for PPMXL, while the described secondary feature is more distinguishable for GSC 2.3, see Figure \[figure:all\].
2MASS and allWISE
-----------------
Both 2MASS and allWISE have good astrometry and show no strong indications of an issue with duplicated sources. In particular, 2MASS does not show signatures of astrometric problems or position error underestimation in the maps included in Figure \[fig:FigResulta\]. As detailed in Table \[table:Be\], the 5.75% of the *Gaia* sources that match a 2MASS source have a mate (i.e. are resolved in *Gaia*), this means that the 2MASS cross-match probably already includes most of the *Gaia* resolved objects and will not benefit much from the more detailed treatment foreseen for DR3 (see Subsection \[subsec:resultangres\]).
Of the external catalogues included in this study, allWISE is the farthest in the infrared and has the lowest angular resolution. Another characteristic of allWISE is that the Galaxy is less prominent in its surface density distribution (see Figure 2 in Paper I), meaning that the surface density distribution variations are lower. These three characteristics explain why the fraction of matched Gaia sources and the fraction of matched allWISE sources are both small. Given its low angular resolution and the relatively small position errors, allWISE will probably appreciably benefit from the more detailed treatment of resolved *Gaia* sources, but it will not dramatically increase the number of matches.
SDSS DR9 and Pan-STARRS1 DR1
----------------------------
SDSS DR9 and Pan-STARRS1 DR1 are similar as both are deeper than *Gaia*, are observed in the same photometric system, and have comparable angular resolutions ($\sim$0.7 and $\sim$1.1 arcsec respectively). However, the Pan-STARRS1 DR1 position errors are definitely smaller than those of SDSS DR9. In the case of Pan-STARRS1 DR1, 90% of the objects have position errors smaller than $\sim$120 mas, while in the case of SDSS DR9, the position errors of the 90% of sources are smaller than $\sim$250 mas.
After the first attempts to cross-match *Gaia* DR2 with SDSS DR9, we realised that we were matching too few objects compared to what we obtained for DR1. We thus decided to use the astrometry QA summary file described in Section \[sec:extcat\], and in particular, the listed astrometric differences in RA and DEC with respect to UCAC-3, which are average differences within a given field. We computed the standard deviation considering all fields obtaining $\sim$50 mas in RA and $\sim$70 mas in DEC, and we thus applied a systematic that is common to all SDSS sources, ameliorating the general cross-match. For *Gaia* DR3, we will apply for each SDSS source the systematic of the corresponding field, and we will use SDSS DR13, which has a new improved photometric calibration with respect to DR9.
Figure \[fig:FigAngDista\] shows that the angular distance distribution of matched sources is narrower for Pan-STARRS1 DR1 than for SDSS DR9 and that the peak is closer to zero in the case of Pan-STARRS1 DR1. However, the maps in panels a) and g) of Figure \[fig:FigResulta\] clearly show a different pattern for SDSS DR9 and Pan-STARRS1 DR1. In the case of SDSS DR9, most *Gaia* sources are matched as expected, given SDSS DR9 has an higher photometric depth. In the case of Pan-STARRS1 DR1, in contrast, a fraction ($\sim$30%) of *Gaia* sources at high Galactic latitudes are not matched, even though these sources were observed by Pan-STARRS1. An in-depth analysis of the characteristics of matched and not matched Pan-STARRS1 DR1 sources shows that the cause might be an issue with the Pan-STARRS1 DR1 astrometric calibration at high Galactic latitudes, which is highlighted by the small position errors. The Pan-STARRS1 DR1 astrometric calibration is described in @panstarrs1e. It should also be noted that the position error broadening method used in the cross-match algorithm described in this paper (see Subsection \[sub:epochdiff\]) implies that when the unknown proper motion of a given *Gaia* source is small, it is somewhat easier to find a match for that source compared with the *Gaia* sources with a full five-parameter astrometric solution. As a direct consequence, the cross-match result for Pan-STARRS1 DR1 includes the fainter *Gaia* sources, which constitute the bulk of *Gaia* sources without proper motions. For both SDSS DR9 and Pan-STARRS1 DR1, the histograms included in Figure \[fig:FigHist\] show that the cross-matched sample correctly does not include the faint sources, which are not observed by *Gaia*.
URAT 1
------
URAT 1 is shallower than *Gaia*, has a larger effective angular resolution, and the position errors for most sources are smaller than 25 mas. The cross-match results summarised in Tables \[table:Be\] and \[table:Stat\] indicate that $\sim$82% of URAT 1 sources have a counterpart in *Gaia* and that in most cases, a single *Gaia* object is matched to a given URAT 1 source (i.e. most matched *Gaia* sources do not have mates). The angular distance distribution in Figure \[fig:FigAngDista\] shows that the peak is very close to zero ($\sim$0.027 arcsec), but panel j) of Figure \[fig:FigResulta\] shows that regardless of the shallowness, not all URAT 1 sources have a *Gaia* counterpart. This can be due to either small astrometric issues, or most probably, to the fact that position errors are underestimated to some degree. The 2MASS survey is also shallower than *Gaia,* and the map that shows the fraction of its sources that matched *Gaia* (panel l) of Figure \[fig:FigResulta\]) can be compared with the corresponding URAT 1 map. It is clear that the astrometry for 2MASS agrees better with *Gaia*, as confirmed by the total of 2MASS matched sources ($\sim$93%).
APASS DR9
---------
The analysis of APASS DR9 cross-match results clearly shows that this survey is affected by various issues, such as the anomalously low steepness of the the cumulative percentage angular distance of matched pairs shown in Figure \[fig:FigAngDista\]. Even though the algorithm found a match for more than 90% of the APASS sources with *Gaia*, a significant fraction of them consists of duplicated sources. This assertion is supported by the angular distance distribution of the nearest neighbours shown in Figure \[fig:FigAngResa\], where it is evident that a large portion of APASS sources ($\sim$8.7 million) have at least one neighbour located at a smaller spatial scale than the angular resolution of the survey ($\sim$5 arcsec). In contrast to GSC 2.3, PPMXL, and Pan-STARRS1 DR1, in the case of APASS the presence of duplicates is a more general problem and they are present not only at tile edges. Hence, it is important to recall that DR9 is not the final release of the APASS project, and the cross-match results should be used with particular caution.
RAVE 5
------
When we analysed the angular resolution of the external catalogues, we found 5633 pairs of sources and 13 triplets of sources in RAVE 5, which, while having different RAVEIDs, seem to be the same sources and are found at distances closer than 3.0 arcsec from each other (and which can be easily found using a cone search). Since RAVE 5 is a sparse catalogue and the cross-match algorithm we use for sparse catalogues forces a one-to-one match, only one of the sources belonging to a given pair or triplet is matched with a *Gaia* source.
{width="0.96\linewidth"}
{width="0.96\linewidth"}
{width="0.96\linewidth"}
![continued.[]{data-label="fig:FigAngDistb"}](figure6b.pdf){width="0.96\linewidth"}
Conclusions\[sec:final\]
========================
![Issues encountered when cross-matching Hipparcos2 with *Gaia*. For a detailed description and explanation of the results shown in this figure, we refer to the main text (Subsection \[subsec:hipissue\]).[]{data-label="Fig:HipIssue"}](figure7.pdf){width="1.0\linewidth"}
We presented the algorithms we developed for the official cross-match of the high-accuracy *Gaia* DR2 astrometric data with eight large dense surveys and three sparse catalogues. The defined algorithms are positional and are able to fully exploit the enormous number of *Gaia* sources with accurate proper motions and parallax measurements using the full five-parameters astrometric covariance matrix on an object-by-object basis. In addition, we included an improved definition of the surface density of observed objects for each catalogue, which allows a better evaluation of the local environment.
The external catalogues and cross-match results were also described. In particular, we analysed the global behaviour of the cross-match results by evaluating their sky distribution, statistical indicators, magnitude, and angular distance distributions. More importantly, we tried to supply scientists, both in the output tables and in the analysis performed in this paper, with all the means to verify the quality of the cross-match results and to understand whether this cross-match is appropriate for their scientific needs.
The excellent data provided by the *Gaia* DR2, and in particular the proper motions, substantially improve the quality of *Gaia* counterparts that are found in external catalogues. The high accuracy of the current *Gaia* data gives a strong drive and powerful tools for understanding and quantifying known complex issues (such as resolution effects, and the presence of astrometric binaries and of duplicated sources) that influence the cross-match results and require non-trivial solutions. The issues will be tackled in the forthcoming *Gaia* data releases.
It is a pleasure to acknowledge the anonymous referee for their contribution to the improvement of the content and the readability of our manuscript. We would like to acknowledge the financial support of INAF (Istituto Nazionale di Astrofisica), Osservatorio Astronomico di Roma, ASI (Agenzia Spaziale Italiana) under contract to INAF: ASI 2014-049-R.0 dedicated to SSDC.
This work has made use of data from the European Space Agency (ESA) mission [*Gaia*]{} (<https://www.cosmos.esa.int/gaia>), processed by the [*Gaia*]{} Data Processing and Analysis Consortium (DPAC, <https://www.cosmos.esa.int/web/gaia/dpac/consortium>). Funding for the DPAC has been provided by national institutions, in particular the institutions participating in the [*Gaia*]{} Multilateral Agreement. We would like to thank G. Fanari, D. Bastieri, B. Goldman, R.A. Power, C. Babusiaux, F. Arenou, C. Fabricius, L. Inno, C. Bailer-Jones, and T. Zwitter for very useful discussions, suggestions, help, and support.
Effective angular resolution\[sec:app\]
=======================================
The effective angular resolving power (or resolution) that results from combining a telescope and its detector is the smallest angle between close objects that can be seen to be separate. The effective angular resolution can be ill defined in astronomy for various reasons: it depends on the brightness difference between two objects, and for ground-based surveys, it is influenced by seeing.
Considering catalogues rather than images, additional considerations must be taken into account. Close sources may have less accurate astrometry and photometry as a result of the disturbing presence of the other nearby source and may be preferentially filtered out from catalogue releases. In addition, when the sky is observed at several different epochs or when fields of view overlap, different observations of the same source may not be recognised as such and duplicated entries are introduced in the catalogue.
The effective angular resolution of the external catalogue is important in order to recognise and correctly match the sources that are resolved in *Gaia* but not in the external catalogue (see Subsection \[subsec:resultangres\]). The comparison between the effective angular resolution values reported in Table \[table:ExtProp\] and the separation distribution in the catalogues is important to evaluate their consistency. In addition, the analysis of the effective angular resolution allows flagging suspected duplicates in the external catalogue, which also can hamper the cross-match results.
In Figure \[fig:FigAngResa\] we show the results of a search of the nearest neighbour (neglecting additional neighbours except for the nearest) around each object in a given dense survey. We used a fixed radius of 5 arcsec for most surveys, with the exception of Pan-STARRS1 DR1, for which, given its size, we used a 3 arcsec radius, and allWISE and APASS DR9, for which, given their resolution, we used 12 and 10 arcsec, respectively. Figure \[fig:FigAngResb\] shows instead density maps (obtained using a HEALPix tessellation with resolution $N_{\mathrm{side}}=2^{8}$) of the number of sources with a nearest neighbour within the search radius defined above.
The histograms in Figure \[fig:FigAngResa\] show the real distribution of nearest neighbours (i.e. the source separation distribution) in the catalogues and should be compared with the effective angular resolutions (dotted vertical lines) listed in Table \[table:ExtProp\]. Nearest neighbours much closer than the marked angular resolution are most probably duplicated sources, while nearest neighbours at distances smaller than but close to the marked angular resolution are still possibly truly distinct sources. The maps shown in Figure \[fig:FigAngResb\] allow evaluating whether the sky distribution of sources with a close nearest neighbour is correlated with known Galactic features or if they are instead related to the survey observation methods.
In Figure \[fig:FigAngResa\], the expected histogram shape of a well-cleaned catalogue is similar to the shapes of SDSS DR9, URAT-1, allWISE, or 2MASS. The initial rise in the Pan-STARRS1 DR1 distance distribution indicates duplicated sources. In Figure \[fig:FigAngResb\] the fields observed by Pan-STARRS1 DR1 with higher source counts are clearly visible, together with some issues at the borders of the hexagonal gigapixel camera tiles. The over-densities are also clearly distinguishable in Figure \[figure:all\]. The very small peak, visible at distances close to zero, in GSC 2.3 indicates the Tycho-2 and SKY2000 (@sky) sources. These sources, which were added to complement GSC 2.3 at the bright end, cannot be considered duplicates. Conversely, the second GSC 2.3 peak (around 0.25 arcsec) indicates the presence of duplicated sources that are mainly present at tile edges (see Figure \[fig:FigAngResb\]). For PPMXL no analysis (or flagging of suspected duplicates) is possible because of the composite nature of the catalogue and because the original observed coordinates are not present in the catalogue, which includes only a set of positions propagated to J2000.0. Nevertheless, it is clear from the map in Figure \[fig:FigAngResb\], but also from Figure 2, panel d) in Paper I, that there is a relevant issue with duplicates. The reverse shape of the APASS DR9 histogram implies that the catalogue is largely affected by duplicated sources: $\sim$14.3% of the objects have at least one neighbour that is located too closely. The map in Figure \[fig:FigAngResb\] also shows several issues in completeness and duplicates.
This type of analysis will be the base for the planned further cross-match developments when we will deal with the possibly duplicated sources, and we will address the issues related to different angular resolutions between *Gaia* and the external catalogues in a more complete way.
{width="0.96\linewidth"}
[^1]:
[^2]: A more exhaustive overview of the mission and DR2 details can be found at [ https://www.cosmos.esa.int/web/gaia/dr2-papers]( https://www.cosmos.esa.int/web/gaia/dr2-papers)
[^3]: Hipparcos Celestial Reference Frame
[^4]: A description of the original ObjectThin and MeanObjects tables can be found at: <https://outerspace.stsci.edu/display/PANSTARRS/PS1+Database+object+and+detection+tables>
[^5]: <http://tapvizier.u-strasbg.fr/>
[^6]: <http://www.sdss.org/dr12/algorithms/astrometry/>
[^7]: <https://data.sdss.org/datamodel/files/BOSS_PHOTOOBJ/astromqa/astromQAFields.html>
[^8]: <http://data.sdss3.org/sas/dr9/boss/photoObj/astromqa/astromQAFields.fits>
[^9]: <https://gea.esac.esa.int/archive/>
[^10]: Space Science Data Center - ASI (<http://gaiaportal.ssdc.asi.it/>), Leibniz Institut für Astrophysik Potsdam - AIP (<https://gaia.aip.de/>), Astronomisches Rechen-Institut (<http://gaia.ari.uni-heidelberg.de/>)
[^11]: <https://www.cosmos.esa.int/web/gaia/dr2-known-issues>
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
author:
- |
Sheng Bau[^1]\
University of Natal, Pietermaritzburg, South Africa\
E-mail: [email protected]\
title: The Decycling Number of Graphs
---
\[section\] \[section\] \[section\] \[section\] \[section\] \[section\] \[section\]
For a graph $G$ and $S\subset V(G)$, if $G - S$ is acyclic, then $S$ is said to be a decycling set of $G$. The size of a smallest decycling set of $G$ is called the decycling number of $G$. The purpose of this paper is a comprehensive review of recent results and several open problems on this graph parameter. Results to be reviewed include recent work on decycling numbers of cubes, grids and snakes(?). A structural description of graphs with a fixed decycling number based on connectivity is also presented. Graphs with small decycling numbers are characterized.
The Decycling Number of Graphs
==============================
The minimum number of edges whose removal eliminates all cycles in a given graph has been known as the [*cycle rank*]{} of the graph, and this parameter has a simple expression. That is, if $G$ is a graph with $p$ vertices, $q$ edges and $\omega$ components, then the cycle rank (or the Betti number) of $G$ is $b(G) = q - p +\omega$ ([@harary], Chapter 4). The corresponding problem for removal of vertices does not have a simple solution. This latter question is difficult even for some simply defined graphs.
Let $G = (V, E)$ be a graph. If $S\subseteq V(G)$ and $G - S$ is acyclic, then $S$ is said to be a [*decycling set*]{} of $G$. The smallest size of a decycling set of $G$ is said to be the [*decycling number*]{} of $G$ and is denoted by $\nabla (G)$. A decycling set of this cardinality is said to be a [*minimum decycling set*]{}. Determining the decycling number of a graph is equivalent to finding the greatest order of an induced forest and the sum of the two numbers equals the order of the graph. It was shown in [@karp] that determining the decycling number of an arbitrary graph is $NP$-complete (see Problem 7 on the feedback node set in the main theorem of [@karp], which asks for a set $S\subseteq V$ of minimum cardinality in a digraph $G$ such that every directed cycle of $G$ contains a member of $S$.).
Clearly, $\nabla (G) = 0$ if and only if $G$ is a forest, and $\nabla (G) = 1$ if and only if $G$ has at least one cycle and a vertex is on all of its cycles. It is also easy to see that $\nabla (K_p) = p - 2$ and $\nabla (K_{r, s}) = r - 1$ if $r\leq s$. This is easily extendable to all complete multipartite graphs. For the Petersen graph $P$, $\nabla (P) = 3$.
We now cite some basic results from [@beinekev].
[([@beinekev], Lemma 1.1)]{} Let $G$ be a connected graph with $p$ vertices, $q$ edges and degrees $d_1, d_2,\ldots , d_p$ in non-increasing order. If $\nabla (G) = s$, then $$\sum_{i = 1}^s (d_i - 1)\geq q - p + 1.$$
[([@beinekev], Corollary 1.2)]{} If $G$ is a connected graph with $p$ vertices, $q$ edges and maximum degree $\Delta$, then $$\nabla (G)\geq\frac{q - p + 1}{\Delta - 1}.$$
For graphs regular of degree $r$, one may wonder whether there is a constant $c$ such that $$\nabla (G)\leq\dfrac{q - p + 1}{r - 1} + c?$$ This is not to be the case, even for cubic graphs (graphs that are regular of degree $3$). Let $G$ be any cubic graph of order $2n$. Replace each vertex of $G$ with a triangle and denote the resulting graph by $H$. Then $|H| = 6n$ and $\nabla (H)\geq 2n$. Thus $$\nabla (H) -\dfrac{q - p + 1}{2}\geq 2n -\dfrac{3n + 1}{2}
\geq 2n -\dfrac{3n}{2} = \dfrac{n}{2}.$$
Which cubic graphs $G$ with $|G| = 2n$ satisfy $\nabla (G) = \lceil\frac{n + 1}{2}\rceil$ ?
Which cubic planar graphs $G$ with $|G| = 2n$ satisfy $\nabla (G) = \lceil\frac{n + 1}{2}\rceil$ ?
Let $S\subseteq V(G)$ and define $$\sigma (S) =\sum_{v\in S} d(v),\quad \epsilon (S) = |E(G|_S)|$$ and denote by $\omega (G)$ the number of components of $G$. Define the [*outlay*]{} of $S$ to be $$\theta (S) =\sigma (S) - |S| -\epsilon (S) -\omega (G - S) + 1.$$
[([@beinekev], Lemma 1.3)]{} Let $G$ be a connected graph with $p$ vertices and $q$ edges. If $S$ is a decycling set of $G$, then $$\theta (S) = q - p + 1.$$
[([@beinekev], Theorem 1.4) ]{} If $G$ and $H$ are homeomorphic graphs then $\nabla (G) = \nabla (H)$.
Denote by $\alpha (G)$ and $\beta (G)$ the independence and the covering numbers of $G$ respectively. Then these two parameters are related by the equality $\alpha (G) +\beta (G) = |G|$.
[([@beinekev], Theorem 1.5)]{} For any nonnull graph $G$, $\nabla (G)\leq\beta (G) - 1$.
Let $G$ and $H$ be two graphs. Then the [*cartesian product*]{} $G\times H$ of $G$ and $H$ is defined by assigning $$V(G\times H) = V(G)\times V(H),$$ $$E(G\times H) = \{\{(x, y), (x', y')\} : [x = x'\wedge yy'\in E(H)]\vee [y = y'\wedge xx'\in E(G)]\}.$$
[([@beinekev], Theorem 1.8)]{} For any graph $G$, $$2\nabla (G)\leq\nabla (K_2\times G)\leq\nabla (G) +\beta (G).$$
The equalities in Lemma 1.5 are satisfied by some graph of each order. For example, if $G = K^c_p$, then $\nabla (G) =\nabla (K_2\times G) = 0$ and both equalities hold. Also, for the equality to the lower bound, if $p\geq 2$ then $\nabla (K_2\times K_p) = 2p - 4 = 2\nabla (K_p)$. The path of order $p$ gives equality to the upper bound.
Cubes
=====
As we have remarked in the previous section, that the determination of the decycling number of an arbitrary graph is $NP$-complete [@harary]. However, results on the decycling number of several classes of simply defined graphs have been obtained ([@baubdlv], [@baubv] and [@beinekev]).
In [@beinekev], upper and lower bounds for the decycling numbers of cubes and grids were obtained, and in [@baubv], an exact formula for the decycling numbers of snakes was given. The results in [@baubdlv] and [@beinekev] will be reviewed in this section.
The $n$-dimensional cube (or $n$-cube) $Q_n$ can be defined recursively: $Q_1 = K_2$ and $Q_n = K_2\times Q_{n-1}$. An equivalent formulation, as the graph having the $2^n$ $n$-tuples of $0$’s and $1$’s as vertices with two vertices adjacent if they differ in exactly one position, gives a coordinatization of the cube. The following result of [@beinekev] gives a lower bound on $\nabla (Q_n)$.
[([@beinekev], Lemma 2.1)]{} Let $n\geq 2$. Then
$(1)$ $\nabla (Q_n)\geq 2\nabla (Q_{n-1})$.
$(2)$ $\nabla (Q_n)\geq 2^{n-1} -\dfrac{2^{n-1} - 1}{n-1}$.
For $n\leq 8$, [@beinekev] computed $\nabla (Q_n)$ exactly.
$n$ 1 2 3 4 5 6 7 8
---------------- --- --- --- --- ---- ---- ---- -----
$\nabla (Q_n)$ 0 1 3 6 14 28 56 112
One of the main results of [@beinekev] is the computation of upper and lower bounds of the decycling numbers of $n$-cubes for $9\leq n\leq 13$. The result is:
Cubes Lower Bound for $\nabla$ Upper Bound for $\nabla$
---------- -------------------------- --------------------------
$Q_9$ 224 312
$Q_{10}$ 448 606
$Q_{11}$ 896 1184
$Q_{12}$ 1792 2224
$Q_{13}$ 3584 4680
These results were improved in [@baubdlv].
[([@baubdlv], Lemma 3.1)]{} For any bipartite graph $G$ with partite sets of cardinality $r$ and $s$ with $r\leq s$, $\nabla (G)\leq r - 1$.
Since the cartesian product of two bipartite graphs is a bipartite graph, $Q_n$ is bipartite. With this observation, the above upper bounds can be lowered a little. For example, they are $255, 511, 1023, 2047$ and $4095$ respectively. Applying Lemma 2.1 (2), one can lift the lower bound a little as well. That is, these lower bounds can be lifted to $225, 456, 922, 1862$ and $3755$ respectively. However, one can still go a little further.
[([@baubdlv], Lemma 3.2)]{} If $e$ and $f$ are two adjacent edges of the $n$-cube $Q_n$, then there is a unique $4$-cycle containing $\{e, f\}$.
[([@baubdlv], Corollary 3.1)]{}
$(1)$ Every edge $uv$ of $Q_n$ is contained in precisely $n - 1$ $4$-cycles;
$(2)$ If $n\geq 3$, then $Q_n$ has precisely $n(n - 1)2^{n-3}$ $4$-cycles.
Denote by $\rho (u, v)$ the distance between points $u$ and $v$. Let $x_0\in Q_n$ and define $$V_k(Q_n, x_0) = \{x\in Q_n : \rho (x, x_0) = k\}.$$ Then there is a nice connection between sizes of the sets $V_k(Q_n, x_0)$ and the binomial coefficients.
[([@baubdlv], Theorem 3.1)]{} $$|V_k(Q_n, x_0)| = {n\choose k}, \quad 0\leq k\leq n.$$
Let the two partite sets of $Q_n$ be denoted by $X_n$ and $Y_n$. Then $|X_n| = |Y_n| = 2^{n-1}$. If $x\in X_n$, then the induced subgraph $G|_{ \{x\}\cup N(x)}$ is a star $S(x)$ of order $n + 1$ centered at $x$. Call a vertex of degree $1$ of a tree a [*leaf*]{}.
[([@baubdlv], Theorem 3.2)]{} For $n\geq 2$, let $x, x'\in X_n$. If $S(x)$ and $S(x')$ are stars then either $S(x)\cap S(x')=\emptyset$ or $S(x)$ and $S(x')$ have precisely two leaves in common.
[([@baubdlv], Theorem 4.1)]{}
$(1)$ $225\leq\nabla (Q_9)\leq 237;$
$(2)$ $456\leq\nabla (Q_{10})\leq 493;$
$(3)$ $922\leq\nabla (Q_{11})\leq 1005;$
$(4)$ $1862\leq\nabla (Q_{12})\leq 2029;$
$(5)$ $3755\leq\nabla (Q_{13})\leq 4077$.
To obtain the upper bounds given in this theorem, a decyclicng set of the given cardinality is to be exhibited in each case and Lemma 2.1 (2) is to be applied. The reader is referred to the elaborate proof of this theorem in [@baubdlv].
Grids
=====
Another class of graphs for which the decycling number has been studied to some precision are the grid graphs $P_m\times P_n$. A standard notation corresponding to matrix notation is to be adopted for convenience. Thus the $i$th vertex in the $j$th copy of $P_m$ will be denoted $v_{i,j}$.
If $S$ is a set of vertices in $P_m\times P_n$, then $S(j)$ will denote the vertices of $S$ in the $j$th column, and put $S(j, k) = S(j)\cup S(j+1)\cup\cdots\cup S(k)$. Let $N(j) = |S(j)|$ and $N(j, k) = |S(j, k)|$.
The following results are obtained in [@beinekev].
[([@beinekev], Corollary 3.3)]{} If $m, b\geq 3$, then $$\nabla (P_m\times P_n)\geq\left\lfloor\frac{mn - m - n + 2}{3}\right\rfloor.$$
[([@beinekev], Theorem 5.1)]{} For $n\geq 4$,
$(1)$ $\nabla (P_2\times P_n) = \displaystyle\left\lfloor\frac{n}{2}\displaystyle\right\rfloor ;$
$(2)$ $\nabla (P_3\times P_n) = \displaystyle\left\lfloor\frac{3n}{4}\displaystyle\right\rfloor ;$
$(3)$ $\nabla (P_4\times P_n) = n ;$
$(4)$ $\nabla (P_5\times P_n) = \displaystyle\left\lfloor\frac{3n}{2}\displaystyle\right\rfloor -\displaystyle\left\lfloor\frac{n}{8}\displaystyle\right\rfloor - 1;$
$(5)$ $\nabla (P_6\times P_n) = \displaystyle\left\lfloor\frac{5n}{3}\displaystyle\right\rfloor ;$
$(6)$ $\nabla (P_7\times P_n) = 2n - 1 .$
[([@beinekev], Theorem 5.3)]{} Let $m = 6q + r$ and $n = 6s + t$ with $1\leq r, t\leq 6$. Then $$\nabla (P_m\times P_n)\leq\min\left\{q(2n - 1) +\nabla (P_r\times P_n),
s(2m - 1) +\nabla (P_t\times P_m)\right\}.$$
[([@beinekev], Theorem 5.4)]{} For $m, n > 2$, $$\nabla (P_m\times P_n) = \frac{mn}{3} + O(m + n).$$
[([@beinekev], Theorem 5.5)]{} Suppose that $n\equiv 0\quad {\rm (mod 2)}$ and $m = 3r + 1$. Then $$\nabla (P_m\times P_n) = rn - r + 1.$$ \[the:gridsd\]
[([@baubdlv], Theorem 5.3)]{} If $S$ is a minimum decycling set of $P_m\times P_n$ with $$\nabla (P_m\times P_n) =\left\lceil\frac{mn - m - n + 2}{3}\right\rceil$$ and $$T = \{v_{ij} : i = 2, 4, \cdots , 3m - 2; j = 2, 4, \cdots , 2n - 2\}$$ then $S\cup T$ is a minimum decycling set of $P_{2m-1}\times P_{2n-1}$.
[([@baubdlv], Corollary 5.1)]{} For any positive integers $r$ and $s$ $$\nabla (P_{6r+1}\times P_{4r-1}) = 8rs - 4r + 1.$$
This theorem covers some cases other than that covered by Theorem \[the:gridsd\].
The problem of determining the decycling numbers of the remaining cases of the grid graphs is open. For the cartesian products $C_m\times C_n$, the following problem is also open.
$\nabla (C_m\times C_n) = ?$
Snakes
======
In this section, a chordless cycle is referred to as a [*cell*]{}. A [*snake*]{} can be defined recursively as follows. A snake with two cells consists of two cycles with one common edge, one of the two cells will be designated the [*head*]{} and the other the [*tail*]{}. A snake with $n+1$ cells is obtained from a snake with $n$ cells by identifying an edge of a new cell with an edge of the tail of the old snake that lies on no other cell. The tail of the new snake is the new cell, and the head remains the same. The [*length*]{} of a snake is the number of cells in it.
Determining a minimum decycling set for a snake is algorithmically straightforward. Given a snake $G$, let $v$ be a vertex on the head with largest possible degree. Put $v$ in the decycling set, then delete it along with all vertices which lie only on cells that contain $v$. What remains is either a shorter snake or a single cell. Either repeating this process on the snake which remains (where the new head is the cell originally adjacent to the old head) or choosing any vertex from a single cell clearly results in a decycling set $S$ of $G$. That $S$ has the minimum possible order follows from the fact that each vertex in $S$ is on some cell that has none of its other vertices in $S$. Thus, $G$ has a set of $|S|$ vertex disjoint cycles. Hence $\nabla (G)\geq |S|$.
Let $G$ be a snake. A [*major pair*]{} is a pair of vertices of degree $3$ such that the edge joining them lies on two cells. A [*minor pair*]{} is a pair of vertices of degree $3$ in a cell which contains exactly two vertices of degree $4$. A minor pair will be said to lie [*between*]{} the two vertices of degree $4$. A vertex of degree at least $4$ is called a [*major*]{} vertex.
Note that adding a new tail cell to an existing snake increases the degree of two of its vertices of the old tail by $1$ each. Since the new tail cell can be incident with at most one vertex of degree at least $3$ in the old snake, its addition either creates a new major vertex, adds one to the degree of an old major vertex, or creates a major pair. This idea gives a natural order, from head to tail, $(u_1, u_2, \cdots , u_s)$ to the set of major vertices, major pairs and minor pairs. Define the [*name*]{} of a snake $G$ to be the sequence $(n_1, n_2,\cdots , n_s)$ where $n_i$ is $3$ if $u_i$ represents a major pair, the degree of the vertex if it represents a major vertex, and $2$ if it represents a minor pair (where $u_{i-1}$ and $u_{i+1}$ are the vertices of degree $4$ that $u_i$ lies between).
With this definition, there may be several snakes with the same name even if the cells are of uniform length. It is easily seen that every finite sequence of integers greater than or equal to $3$ is the name of some snake.
Given a snake $G$ and its name $N(G) = (n_1, n_2,\cdots , n_s)$, define the [*nickname*]{} $C(G)$ of $G$ (as a subset of $\{1, 2,\cdots , s\}$) as follows. (1) $1, s\in C(G)$.
\(2) Assume that for $i<s - 1$ it has been determined whether or not each of $1, 2,\cdots i$ is in $C(G)$. Then
\(i) If $n_{i+1}\geq 6$, then $i+1\in C(G)$;
\(ii) If $n_{i+1} = 5$, then $i+1\in C(G)$ if and only if $i\not\in C(G)$;
\(iii) If $n_{i+1} = 4$ then $i+1\in C(G)$ if and only if either $n_i\geq 5$ and $i\not\in C(G)$, or $n_i = 4$ and $i, i-1\not\in C(G)$.
With this definition, it can be shown that the decycling number of a snake whose cells are all $4$-cycles is the cardinality of its nickname. From this result the following theorem follows.
[([@baubv], Theorem 2.1)]{} Let $G$ be a snake with nickname $C(G)$. Then $\nabla (G) = |C(G)|$.
A [*subsnake*]{} of a snake $G$ is a subgraph of $G$ that is itself a snake. A [*straight segment*]{} of a snake whose cells are all squares is a subsnake in which the vertices of each of the shared edges is a major pair. A [*maximal*]{} straight segment $T$ of a square-celled snake $G$ is a straight segment of $G$ such that for each cell $s\not\in T$, $T\cup s$ is not a straight segment of $G$. A quare-celled snake $G$ is said to be [*nonsingular*]{} if each its maximal straight segment has at least three cells; otherwise it is said to be [*singular*]{}. The [*segment sequence*]{} of a square-celled snake $G$ is the sequence of lengths of maximal straight segments of $G$ ordered from head to tail.
[([@baubv], Theorem 3.1)]{} If $(d_1, d_2,\cdots , d_k)$ is the segment sequence of a nonsingular snake $G$, then $$\nabla (G) =\sum\limits^k_{i=1}\left\lceil\frac{d_i}{2}\right\rceil - k + 1.$$
The decycling number of a singular snake is certainly related to that of a nonsingular one by means of a certain transformation (surgery) of the snake. The decycling numbers of snakes with cell size not equal to $4$ are related to those of snakes with cell size $4$ by means of simple transformations. It is therefore possible to consider the decycling problem with restriction to square-celled snakes only.
A snake with triangular cells is a special type of triagulation of a polygon, namely that in which every triangle contains at least one edge of the polygon. This raises the question of decycling triangulations of polygons, or equivalently, the maximal outerplanar graphs. In general, this seems to be considerably more complicated than decycling snakes. At present, we content ourselves with bounds.
[([@baubv], Theorem 3.4)]{} If $G$ is a maximal outerplanar graph of order $n$, then $$1\leq\nabla (G)\leq\left\lfloor\frac{n}{3}\right\rfloor.$$
Even an algorithm similar to that described at the beginning of this section is not known for the computation of the decycling number of a triangulation of a polygon. An interesting open problem is to determine the decycling number of outerplanar graphs.
Is there a fast algorithm for computing the decycling numbers of (maximal) outerplanar graphs?
Determine the decycling numbers of $2$-dimensional trees.
Fixed Decycling Numbers
=======================
This section is a simple note on the dependency of decycling number of a graph on its connectivity number. In particular, we consider the almost trivial question of determining all graphs with decycling number $2$ or $3$.
Let $H$ and $J$ be graphs, $S\subseteq V(H)$ and $T\subseteq V(J)$ with $|S| = |T|$. Let $f : S\rightarrow T$ be a bijection. An [*identification*]{} of $H$ and $J$ [*via*]{} $f$ is any graph $G$ denoted $H\circ_f J$ obtained by identifying $H$ and $J$ via $f$ and maintaining the adjacencies in the rest of $H$ and $J$. As for the adjacencies in $S$ and $T = f(S)$, keep all the edges $E(H|_S)\cup E(J|_T)$. Formally, for each $x\in S$, identify $x$ and $f(x)$, thus embedding $S$ onto $T = f(S)$. Now for $G = H\circ_f J$, $$V(H\circ_f J) = V(H)\cup V(J)\backslash T,$$ $$E(H\circ_f J) = E(H)\cup E(J).$$
Let $S$ be a $($minimum$)$ decycling set of $H$ and $T$ be a $($minimum$)$ decycling set of $J$. If $|S| = |T|$ and $f : S\longrightarrow T$ is a bijection, then $S$ is a $($minimum$)$ decycling set of $H\circ_f J$.
[*Proof*]{}: Since $S$ and $f(S)$ decycle $H$ and $J$, $S$ is a decycling set of $H\circ_f J$. $H\circ_f J - S$ is a forest that is the union of forests $H - S$ and $J - S$. If $S$ is a mimimum decycling set of $H$ and $T$ is a minimum decycling set of $J$, then $S$ is a minimum decycling set of $H\circ_f J$. To see this, let $S'$ be any minumum decycling set of $H\circ_f J$ such that $|S'|<|S|$. Then since $\nabla (H) = |S|$, $S'|_H$ cannot decycle $H$. Let $C$ be a cycle of $H - S'|_H$. Then $C$ must be a cycle of $H\circ_f J - S'$ since $H - S'|_H$ is an induced subgraph of $H\circ_f J - S'$.
Denote by $\kappa (G)$ the connectivity of $G$.
If $\nabla (G) = k$ then $\kappa (G)\leq k + 1$.
[*Proof*]{}: Let $S$ be a decycling set of $G$ with cardinality $k =\nabla (G)$. If $\kappa (G)\geq k + 2$, then $G - S$ is $2$-connected, and hence $G - S$ is not a forest. This contradicts the choice of $S$. Hence $\kappa (G)\leq k + 1$.
Let $G$ be a $(k + 1)$-connected graph with $\nabla (G) = k$. Since $G$ is not $(k + 2)$-connected, $\kappa (G) = k + 1$. Let $S$ be a minimum decycling set of $G$. Then $G - S$ is a connected acyclic graph, i.e., a tree. Thus $G$ can be obtained by joining a set $S$ to a tree of order at least $k$ in a way so as to make $G$ a $(k + 1)$-connected graph.
Let $\nabla (G) = k$. Then $\kappa (G) = k + 1$ if and only if there is a tree $T$ and a set $S =\{x : x\not\in T\}$ with $|S| = k$ such that $S$ and $T$ are embedded in $G$ and ??.
[*Proof*]{}: Let $\kappa (G) = k +1$ and $S$ be a minimum decycling set of $G$. Since $G$ is $(k + 1)$-connected, $G - S$ is $k$-connected. Since $S$ is decycling set, $G - S$ is acyclic. Hence $G - S$ is a tree. Therefore, $G$ is as described.
On the other hand, if $G$ is a $(k + 1)$-connected graph as described, then $G - S$ is a tree and $\kappa (G) = k + 1$.
Inductive construction of all graphs with $\nabla (G) = k$?
We now determine the graphs with decycling number $2$ and $3$.
Applications
============
The direct application of systems of paths in $n$-cube in coding theory is well known (see Chapter ?? of [@macwilliamss]), while a maximal such system of paths is what is left after deletion of a minimum decycling set. Thus the study of decycling number of cubes, and other particular classes of graphs in general, lends itself to coding theory—a more readily applied branch of mathematics.
S. Bau, L.W. Beineke, G-M. Du, Z-S. Liu and R.C. Vandell, Decycling cubes and grids, [*Utilitas Math.*]{}, 1(2000), 10-18. S. Bau, L.W. Beineke and R.C. Vandell, Decycling snakes, [*Congr. Numer.*]{}, 134(1998), 79-87. L.W. Beineke and R.C. Vandell, Decycling graphs, [*J. Graph Theory*]{}, 25(1996), 59-77. F. Harary, Graph Theory, Academic Press, New York 1967. R.M. Karp, Reducibility among combinatorial problems, Complexity of Computer Computations (R.E. Miller, J.W. Thatcher, ed.), Plenum Press, New York-London (1972), 85-103. F.J. MacWilliams and N.J.A. Sloane, The Theory of Error-Correcting Codes, North-Holland, Amsterdam 1978.
[^1]: Work supported in part by University of Natal Research Fund URF No. 4509.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'We consider an error model for quantum computing that consists of “contagious quantum germs" that can infect every output qubit when at least one input qubit is infected. Once a germ actively causes error, it continues to cause error indefinitely for every qubit it infects, with arbitrary quantum entanglement and correlation. Although this error model looks much worse than quasi-independent error, we show that it reduces to quasi-independent error with the technique of quantum teleportation. The construction, which was previously described by Knill, is that every quantum circuit can be converted to a mixed circuit with bounded quantum depth. We also consider the restriction of bounded quantum depth from the point of view of quantum complexity classes.'
author:
- Gil Kalai
- Greg Kuperberg
title: Contagious error sources would need time travel to prevent quantum computation
---
=1
[^1]
[^2]
Introduction
============
Is quantum computation realistic even in principle? If we accept quantum mechanics (more precisely, quantum probability), then at the theoretical level this question is usually interpreted as the fault tolerance problem: Can a quantum computer still work if all of its gates and qubits are noisy? There are by now various fault-tolerance theorems for quantum computation, which establish that reliable quantum computation is indeed possible in principle assuming that the noise present in different qubits or gates is quasi-independent [@KLZ:resilient; @AB:constant; @Kitaev:imperfect], and is below some threshold error rate. This threshold is called the fault tolerance constant. Thus, any remaining doubt that quantum computation is possible in principle reduces to one of three possibilities:
1.
: Quantum probability is not exactly true.
2.
: The fault tolerance constant is unattainable.
3.
: The quasi-independence assumption is too optimistic.
In this note, we will consider noise models with a relaxed version of the quasi-independence assumption, namely *contagious noise*. It seems possible that each qubit in a quantum computer might not just be noisy, but carry with it a noise source, a contagious “bug", that spreads to all of the output qubits of each quantum gate. Each bug could get worse over time. Worse still, the descendants of the bug could be correlated and thus violate the quasi-independence assumption. If a quantum gate has two different bugs among its inputs, the bugs might also interact and make new bugs. Such possibilities come to mind given that one of the first bugs in the history of modern computing was an actual bug, a small moth [@Hopper:bug]. That bug was no longer interacting with anything other than the relay switch where it had died. A “bug" can also mean a germ; at least in biological computers, germs can both replicate and affect data.
More realistically, contagious error is related to some forms of leakage error, where what was the state of a qubit leaves the qubit Hilbert space and enters a larger Hilbert space. Knill has noted that leakage error is implicitly solved by teleportation [@Knill:scalable], which is also the method that we will use. Leakage error is generally thought of as a measured error; if it is measured, it amount to a qubit erasure and the qubit can be reset. However, before it is measured, leakage error can be contagious, since the effect of a quantum gate is undefined for leaked states.
In this article, we propose a mutual generalization of contagious germs and leakage error which we call *contagious quantum germs*. If a qubit has a Hilbert space ${\mathcal{H}}_Q \cong {\mathbb{C}}^2$, then we attach to it another Hilbert space ${\mathcal{H}}_G$, the germ state space, so that its total state space is ${\mathcal{H}}_Q
{\otimes}{\mathcal{H}}_G$. At each time step, each qubit interacts with its germ and its germ evolves. At each gate, all of the germs of the input qubits interact in some way to make the germs of the output qubits. The only restrictions are that each fresh qubit is created with an independent germ, that the effect of a germ on its qubit is bounded above for the first few steps of its life, and that classical bits do not carry germs. Note that quantum germs not only strictly generalize leakage error, but can also do things that would be peculiar for leakage error. For instance, once activated, they can spread insidiously with no possibility of being detected, and then hit hard everywhere that they have spread. We have no argument that all types of quantum germs are realistic, only that they include various possibly realistic noise models as special cases.
With a constant overhead factor, every quantum circuit can be re-encoded so that noise from contagious quantum germs becomes quasi-independent. \[th:main\]
As a corollary, we can then apply the standard fault-tolerance theorem to conclude that quantum computation is possible in our model with polylogarithmic overhead.
A more detailed version of [Theorem \[th:main\]]{} is stated as [Theorem \[th:rig\]]{} and [Corollary \[c:replace\]]{}, using the formalism defined in [Section \[s:rig\]]{}.
A closely related result is the “one-way" quantum fault tolerance proposal of Raussendorf, Harrington, and Goyal [@RHG:oneway; @RHG:cluster]. This is a specific encoded computation with a high fault-tolerance constant, and which is one-way in the sense that after an initial encoded state is created, all of the computation is carried out by (adaptive) one-qubit measurements. Since in addition all of the parity checks of their code have bounded weight, their circuits automatically have bounded depth. Our [Theorem \[th:main\]]{} is both weaker and more general than the RHG construction. Without obtaining any new bound on the fault tolerance constant, it implies that any method of quantum fault tolerance whatsoever can be re-encoded in bounded quantum depth.
The idea of our proof of [Theorem \[th:main\]]{} is not original. The basic construction uses quantum teleportation [@BBCJPW:teleport]; it might first have been published by Knill [@Knill:postsel]. We state the construction in terms of mixed quantum-classical circuits.
With a constant overhead factor, every quantum circuit (or mixed classical-quantum circuit) can be re-encoded as a mixed classical-quantum circuit with bounded quantum depth. \[th:knill\]
The surprising property of a mixed circuit with quantum teleportation is that, if we orient all of the qubit edges forward in time, then the qubit subgraph only needs to be weakly connected from its input to its output in order to transmit quantum information. (Recall that directed graph is *strongly* connected from $a$ to $b$ if it has a directed path from $a$ to $b$; and *weakly* connected if it merely has some connecting path with no restriction on the orientations of the edges.) In a sense, quantum information can travel backward in time, as long as classical information transports the result forward in time. Indeed, quantum teleportation is a reasonable description of *any* mixed circuit that transmits quantum information via a weakly connected quantum subgraph; in this sense, quantum information that travels backwards in time is just a slogan for teleportation. Theorems \[th:main\] and \[th:knill\] then say that if we imagine that qubits (but not bits) are infected with contagious germs, then the germs would have to travel backward in time to prevent fault-tolerant quantum computation.
Given that we can loosely interpret a teleportation circuit as allowing quantum information to travel backwards in time, the question arises whether quantum germs could travel backwards in time in the same sense of teleportation. The answer is no, unless either:
1.
: Germs can infect classical bits.
2.
: Germs can travel on paths that are not in the circuit.
Given the many different ways that classical bits are encoded and communicated in practice, the first possibility seems implausible or at least avoidable. The second possibility certainly can happens in the sense that noise may not be limited to a quantum circuit in example physical implementations. In this case, the noise does not need to travel backwards in time either. However, if correlated noise can spread anywhere with no causal restriction other than that it travels forward in time, then it is well-known that either classical or quantum fault tolerance is impossible.
The authors were led to consider the constructions considered here by alternative error models proposed by the first author in which errors are convolved, or smoothed, in time [@Kalai:adversarial]. Our basic observation led the first author to change his model to require convolution both forward and backward in time [@Kalai:fail]. This leads to issues regarding causality that we will not discuss here.
If a mixed circuit is not even weakly quantumly connected, then it is equivalent to a model in quantum information theory known as “local operations and classical communication" [@NC:qcqi §12.5]. In particular, it is immediate that LOCC is weaker than full quantum communication, since it leaves no way to create quantum entanglement between weakly connected components of the circuit, and therefore no way to violate Bell-type inequalities.
To understand our hypothesis and our conclusion, it is important to distinguish between sources of error and erroneous qubits (or bits). An erroneous qubit is one whose state is different from what is intended in a quantum algorithm. If the intended state is a pure state ${|\psi\rangle}$ (or a density operator $\rho$), then the actual state might be some other state ${|\psi'\rangle}$ (or a density operator $\rho'$). Erroneous states propagate through gates and through quantum teleportation. In computer science in general, this is called *error propagation* and it is why computers need classical or quantum error correction. In fact, an error can propagate through a mixed path of classical and quantum edges in a mixed circuit; in particular, an error can propagate through the classical bits used in quantum teleportation. However, while errors can be corrected, by our rules error-causing germs cannot be removed. (Or, they can only be removed indirectly with teleportation.)
We will prove [Theorem \[th:knill\]]{} in [Section \[s:knill\]]{} and [Theorem \[th:main\]]{} in [Section \[s:main\]]{}. Finally, in [Section \[s:complexity\]]{}, we give a complexity theory interpretation of [Theorem \[th:knill\]]{}. One way to limit the power of quantum computation is to only allow bounded-time layers of it in between classical computation layers. (We do not mean pseudo-classical operators that are quantum but in the computational basis. Rather we mean classical data processing revealed to the environment.) We remark that if this is done asynchronously, then [Theorem \[th:knill\]]{} implies that the resulting polynomial complexity class is exactly ${\mathsf{BQP}}$.
The authors would like to thank Dorit Aharonov, Andrew Childs, Joe Fitzsimons, Eleanor Rieffel, and especially Daniel Gottesman for useful conversations and corrections.
Rigorous definitions {#s:rig}
====================
Mixed circuits {#s:mixed}
--------------
We consider the circuit model of computation. As usual, a circuit is a kind of acyclic, directed graph with labelled vertices which are called gates. In defining classical circuits carefully enough to generalize them to quantum circuits, we have to count bit copying as a gate with 1 input and 2 outputs. Also, every type of circuit that we consider in this paper can be assumed to be in a uniform circuit family, created by a classical Turing machine or similar.
The circuits of interest to us have two kinds of circuit edges, classical or bit edges, and quantum or qubit edges. In order to understand what a general mixed gate can do with a combination of bit and qubit inputs and outputs, we can consider the *hybrid quantum memory* model [@Kuperberg:memory]. In practice, we can simplify the definition of a mixed circuit to the following types of gates:
1.
: Deterministic classical gates acting on bits.
2.
: A measurement gate that converts a qubit to a bit.
3.
: Unitary quantum gates that may have classical control bits.
4.
: A gate that creates a fresh qubit in the state ${|0\rangle}$.
Since our circuits are uniformly generated, and for other reasons, we also want a finite set of quantum gates that densely generate unitary groups acting on qubits, such as the Hadamard and Toffoli gates. However, rather than proving something for every gate set, we interpret [Theorem \[th:main\]]{} as saying that there exists a set of gates such that the result holds with a constant overhead factor for those gates. (Changing gate sets requires the Solovay-Kitaev theorem, which has polylogarithmic overhead; we do not know that the stringent constant overhead factor can be satisfied by every universal gate set.) The only standard gates that we need for quantum teleportation ([Figure \[f:teleport\]]{}), which is the only idea we need to prove [Theorem \[th:knill\]]{}, are Hadamard and CNOT gates, and 1-qubit Pauli gates with classical control bits.
[Figure \[f:teleport\]]{} has an example of a mixed circuit, with the qubit edges in red and the bit edges in blue. In general a circuit has a *total depth*, which is the length of its longest directed path; and a *quantum depth*, which is the length of its longest directed path following only qubit edges. The graph of a mixed circuit has a *quantum subgraph* consisting only of its qubit edges. (The specific gates used in the circuit are defined in [Section \[s:knill\]]{}.)
[f:teleport]{}[The teleportation circuit $T$, with qubit edges in red and bit edges in blue. The output is connected to the input by a path of qubit edges, but not by an *oriented* path]{}
(-5,1) – (-4,1); (-3,1) – (-2,1);
(0,0) – (1,0); (1,0) – (6,0);
(0,1) – (1,1); (1,1) – (3,1); (3,1) – (4,1); (4,1) – (6,1); (6,1) – (7,1); (7,1) – (8.5,1);
(-1,2) – (3,2); (3,2) – (4,2); (4,2) – (7,2);
(-4,.5) rectangle (-3,1.5); (-3.5,1) node [$T$]{}; (-1.5,1) node [$=$]{}; (-.25,-.25) rectangle (.25,.25); (0,0) node [$0$]{}; (-.25,.75) rectangle (.25,1.25); (0,1) node [$0$]{}; (.75,-.25) rectangle(1.5,1.25); (1.125,.5) node [$U^{-1}$]{}; (2.75,.75) rectangle(3.25,2.25); (3,1.5) node [$U$]{}; (3.75,.75) rectangle (4.25,1.25); (4,1) node [$m$]{}; (3.75,1.75) rectangle (4.25,2.25); (4,2) node [$m$]{}; (5.75,-.25) rectangle(6.25,1.25); (6,.5) node [$cZ$]{}; (6.75,.75) rectangle(7.25,2.25); (7,1.5) node [$cX$]{};
Quantum germs {#s:germs}
-------------
As usual $U(n)$ is the group of unitary $n \times n$ matrices; let $M(n)$ be the vector space of all $n \times n$ matrices. If ${\mathcal{H}}$ is a Hilbert space, then we let $U({\mathcal{H}})$ be the corresponding abstract unitary group, and we let $M({\mathcal{H}})$ be the abstract space of all operators on ${\mathcal{H}}$. The algebra $M({\mathcal{H}})$ comes with an operator norm; by definition $$||A|| = \sup_{{\langle\psi|\psi\rangle} = 1} {\langle\psi|A|\psi\rangle}.$$ If ${\mathcal{H}}$ is infinite-dimensional, then technically we take $M({\mathcal{H}})$ to be the bounded operators, meaning those with finite operator norm. Recall in this case that a state $\rho$ on ${\mathcal{H}}$ is a positive semi-definite trace-class operator. (More precisely, if $\rho$ is positive semi-definite, then it is *trace class* if the trace of its matrix defined using any orthonormal basis of ${\mathcal{H}}$ is a convergent series.)
As mentioned in the introduction, each qubit has a Hilbert space ${\mathcal{H}}_Q
\cong {\mathbb{C}}^2$, and a separate germ Hilbert space ${\mathcal{H}}_G$ that could even be infinite-dimensional. If $C$ is a quantum circuit (or a mixed circuit), we expand it to an infected circuit $C'$ as follows: If $C$ creates a qubit with the state ${|0\rangle} \in {\mathcal{H}}_Q$, then $C'$ also creates a germ in some initial state ${|g_0\rangle} \in {\mathcal{H}}_G$. For each gate $$G \in U({\mathcal{H}}_Q^{(1)} {\otimes}{\mathcal{H}}_Q^{(2)} {\otimes}\cdots {\otimes}{\mathcal{H}}_Q^{(k)})$$ that arises in $C$, there is a corresponding germ-mixing operator $$M \in U({\mathcal{H}}_G^{(1)} {\otimes}{\mathcal{H}}_G^{(2)} {\otimes}\cdots {\otimes}{\mathcal{H}}_G^{(k)})$$ that mixes the germ states. Finally, each edge of the circuit $C$ is replaced with an error operator $$E \in U({\mathcal{H}}_G {\otimes}{\mathcal{H}}_Q).$$ Also, we do not assume that the operators $M$ and $E$ and the states ${|g_0\rangle}$ are the same at different positions in $C$. They must satisfy the error bound below, but otherwise they can be different each time and they can be chosen adversarially rather than randomly.
The operators $E$ are subject to an error bound which we explain carefully. Recall the relation $$M({\mathcal{H}}_G {\otimes}{\mathcal{H}}_Q) \cong M({\mathcal{H}}_G) {\otimes}M({\mathcal{H}}_Q).$$ Recall that $M({\mathcal{H}}_Q) \cong M(2)$ can be given a Pauli basis $$\begin{aligned}
P_0 &= I = \begin{pmatrix} 1 & 0 \\ 0 & 1 \end{pmatrix}, &
P_1 &= X = \begin{pmatrix} 0 & 1 \\ 1 & 0 \end{pmatrix} \\
P_2 &= Y = \begin{pmatrix} 0 & -i \\ i & 0 \end{pmatrix}, &
P_3 &= Z = \begin{pmatrix} 1 & 0 \\ 0 & -1 \end{pmatrix}\end{aligned}$$ using any isomorphism ${\mathcal{H}}_Q \cong {\mathbb{C}}^2$. Then we can think of an operator $E \in M({\mathcal{H}}_G {\otimes}{\mathcal{H}}_Q)$ as a superposition of operators acting only on ${\mathcal{H}}_G$: [$$\label{e:superpos}E = \sum_{j=0}^3 E_j {\otimes}P_j.$$]{} The fact that $E$ is unitary implies that $||E_j|| \le 1$, which we can read as saying that each $E_j$ is subunitary; this will be useful in the proof of [Theorem \[th:main\]]{}.
As a warm-up to the main argument, suppose that the germ at a given qubit edge has a pure state ${|g\rangle} \in {\mathcal{H}}_G$. Then the partial evaluation of $E$ on ${|g\rangle}$ gives us a vector $${\vec{F}}\in {\mathcal{H}}_G {\otimes}M({\mathcal{H}}_Q),$$ which then decomposes as [$$\label{e:klz}{\vec{F}}= \sum_{j=0}^3 {|f_j\rangle} {\otimes}P_j.$$]{} Here each ${|f_j\rangle}$ is a non-normalized state representing a vector-valued amplitude of the error mode. Following Knill, Laflamme, and Zurek [@KLZ:resilient §I.B], we can define the size of this error as the sum of the norms of the output germ states ${|f_j\rangle}$ other than the term for the identity. In other words, we can define an error seminorm $$||{\vec{F}}|| = \sum_{j=1}^3 \sqrt{{\langlef_j|f_j\rangle}}.$$ One subtle but standard point, which will be relevant in all of our error bounds, is that the vectors such as ${|f_j\rangle}$ need not be orthogonal. If they are orthogonal, then the different errors to which they are attached are stochastic; if they are parallel, then the errors are coherent or “stoquastic".
If $||{\vec{F}}||$ is large, it means that the error operator $E$ has a large effect on its qubit $Q$. We would like to bound $||{\vec{F}}||$; however for two reasons, we will not do this in all cases. The first reason is that the state $\rho_G$ in general comes from a pure state that is entangled between many germs and computational qubits as well. The second reason is that we assume that if a germ creates an error in a qubit, then it is activated and can cause later errors with high probability.
We address the first issue, and clarify the second one, by passing to a multilinear expansion of all of the error operators using . Instead of directly considering the full vector state of all of the germs and the errors they cause, we can instead consider the amplitude contribution of any particular pattern of Pauli errors. The total error is a superposition of all of these patterns. To prove Theorem \[th:main\], we will bound each term of the superposition separately, and then sum to get the total bound.
If there are $N$ edges, then we expand all possible errors across the circuit $C$: [$$\label{e:mpauli} {\vec{F}}= \sum_{J=0}^{4^N-1} {|f_J\rangle} {\otimes}P_J.$$]{} Here $P_J$ is a *multi-Pauli* operator, a tensor product of Pauli operators including the identity. We consider the partial ordering on qubit edges in which $q_1 \prec q_2$ if there is a directed path from $q_1$ to $q_2$. Then with respect to this partial ordering, some of the Pauli factors of $P_J$ are the earliest among those that are not the identity. We call these qubit edges *locally first diseased* (in superposition).
If $q$ is locally first diseased, then all of the germs that ever interacted with the one at $q$ have an entangled state $${|g\rangle} \in {\mathcal{H}}_G^{(1)} {\otimes}{\mathcal{H}}_G^{(2)}
{\otimes}\cdots {\otimes}{\mathcal{H}}^{(k)}.$$ The state ${|g\rangle}$ is formed from various initial states ${|g_0\rangle}$ with germ-mixing operators $M$ and error components $E_j$ acting on them. Since each $M$ is unitary and each $E_j$ is subunitary, we learn that $|{\langleg|g\rangle}| \le 1$. We can make an error vector ${\vec{F}}$ in the same way as in the warmup case, except using a state ${|g\rangle}$ of many germs rather than one germ. We assume an upper bound [$$\label{e:bound}||{\vec{F}}|| < {\epsilon}(n),$$]{} if the quantum depth of the part of the circuit $C$ that leads to $q$ is at most $n$. We assume that ${\epsilon}(n)$ is a small number when $n$ is small. Otherwise, if the best upper bound ${\epsilon}(n)$ is large for small $n$, then even small quantum circuits are unreliable; with enough such noise, there is no clear reason to expect quantum computation to be possible.
Proof of [Theorem \[th:knill\]]{} {#s:knill}
=================================
The theorem reduces to the existence of quantum teleportation. Quantum teleportation is a mixed quantum-classical circuit $T$ that has one qubit input and one qubit output, and no quantum path from the input to the output. Moreover, the circuit computes the identity: The output agrees with the input and it even inherits any entanglement that the input had with other qubits. The teleportation circuit is given in [Figure \[f:teleport\]]{}; a simplified version is given in [Figure \[f:telesimp\]]{}.
As mentioned, our convention for all diagrams is that qubit edges are red and bit edges are blue. The gates used in the expanded circuit in [Figure \[f:teleport\]]{} are as follows:
- The gate $0$ creates a qubit in the state ${|0\rangle}$.
- The gate $m$ measures a qubit in the computational basis and outputs a bit. The qubit input is destroyed.
- The gate $U$ is unitary with the following action: $$\begin{aligned}
U{|00\rangle} &= \frac{{|00\rangle}+{|01\rangle}}{\sqrt{2}} &
U{|01\rangle} &= \frac{{|10\rangle}-{|11\rangle}}{\sqrt{2}} \\
U{|10\rangle} &= \frac{{|10\rangle}+{|11\rangle}}{\sqrt{2}} &
U{|11\rangle} &= \frac{{|00\rangle}-{|01\rangle}}{\sqrt{2}}.\end{aligned}$$ It can be created as a CNOT gate $${|x,y\rangle} \mapsto {|x+y,y\rangle}$$ followed by a Hadamard gate $$H = \frac1{\sqrt{2}} \begin{pmatrix} 1 & 1 \\ 1 & -1 \end{pmatrix}$$ applied to the second qubit.
- The gate $cX$ applies the one-qubit operator $X$ if its bit input is 1, and the identity $I$ if it is 0.
- The gate $cZ$ applies the one-qubit operator $Z$ if its bit input is 1, and the identity $I$ if it is 0.
If the circuit $T$ is inserted at every edge of a circuit $C$ to make $C'$, then the longest directed qubit path in $C'$ is 6 edges. An initialized qubit lasts for 4 edges inside $T$. The output of $T$ can be the input to some gate $G$ in $C$, and then the output of $G$ can become the input to another copy of $T$ and last for 2 more edges to make 6 total.
It is a celebrated result that the teleportation circuit $T$ computes the identity, even when its input is entangled. In order to prove this, it suffices to check that it is the identity for any spanning set of density operators. For example if the circuit preserves the vector states ${|0\rangle}$, ${|1\rangle}$, ${|\pm\rangle}$, and ${|\pm i\rangle}$, then it also preserves all six corresponding density operators, which implies that it preserves all density operators. It is easy to check that these states are indeed preserved by $T$.
Proof of [Theorem \[th:main\]]{} {#s:main}
================================
[f:telesimp]{}[A simplified diagram of the teleportation circuit in [Figure \[f:teleport\]]{}, in which the gates are combined as much as possible. The gate $b$ creates a Bell pair; the gate $m$ measures two qubits in an entangled basis, and the gate $c$ is a unary quantum gate controlled by two bits]{}
(-1,3) – (1,3); (0,1) – (1,3); (0,1) – (2,0); (.9,3.05) – (1.9,0.05); (1.15,2.95) – (2.15,-0.05); (2,0) – (4,0); (0,1) circle (.25); (0,1) node [$b$]{}; (1,3) circle (.25); (1,3) node [$m$]{}; (2,0) circle (.25); (2,0) node [$c$]{}; (-1,3) node [${|\psi\rangle}$]{}; (4,0) node [${|\psi\rangle}$]{};
We will actually prove two different results that both fit the words of [Theorem \[th:main\]]{}.
Let $C$ be a circuit with bounded quantum depth $n$ and gates that act on at most $k$ qubits. If $C$ is infected with quantum germs with error bound $\delta(n)$ at depth at most $n$, then $C$ has quasi-independent error with bound ${\epsilon}$ depending on $n$, $k$, and $\delta(n)$. Moreover, ${\epsilon}\to 0$ as $\delta(n) \to 0$. \[th:rig\]
If $C$ is a circuit with unbounded quantum depth, then it can be replaced by an equivalent circuit $C'$, so that if $C'$ is infected with quantum germs, then the result is equivalent to quasi-independent noise in $C$. \[c:replace\]
[f:sandwich]{}[A schematic comparison of two kinds of mixed quantum-classical circuits]{}
[Corollary \[c:replace\]]{} is important because $C$ could be constructed according to a fault tolerance threshold theorem that assumes quasi-independent error.
We review the definition of quasi-independent error [@KLZ:resilient]. We assume that the edges of the circuit $C$ are subject to error. Then we consider the multilinear expansion . The operator $P_J$ has a weight $w(J)$, which is the number of tensor factors that are not the identity. The quasi-independent error condition says that $$\sqrt{{\langlef_J|f_J\rangle}} = O({\epsilon}^{w(J)}).$$ for some error bound $0 < {\epsilon}< 1$. (The fault tolerance theorem says that fault tolerance with polylog overhead is possible if ${\epsilon}$ is small enough. Note that Knill, Laflamme, and Zurek call this type of error *monotonic* quasi-independent error.)
Let $P_J$ be a multi-Pauli operator that arises in the multilinear expansion of . Then we claim that $$\sqrt{{\langlef_J|f_J\rangle}} < \delta(n)^{m(J)},$$ where $m(J)$ is the number of locally first errors in $P_J$. In each term $J$, the state ${|f_J\rangle}$ is actually the entangled state ${|g\rangle}$ of all of the germs at the end of the computation. At each position that is a locally first error, the norm of ${|g\rangle}$ decreases by a factor of $\delta(n)$ by . At every other position, the norm does not increase because each operator $E_j$ is subunitary.
Finally, each edge with a locally first error has at most $$1+k+\cdots + k^{n-1} < k^n$$ edges above it that could have errors. It follows that $$w(J) < k^n m(J).$$ Thus $C$ has quasi-independent error with bound $${\epsilon}< \delta(n)^{k^{-n}},$$ as desired.
The original purpose of the quasi-independent error model is that it renormalizes to itself under any map that changes a circuit $C$ to an equivalent circuit $C'$ made by replacing gates and qubit edges by gadgets. This is the technique to prove the fault tolerance theorem using concatenated quantum codes. Such an analysis applies in our case because we can replace each edge by a teleportation gadget, as we did in the proof of [Theorem \[th:knill\]]{}. Indeed, the analysis is particularly simple because every multi-Pauli error in the teleportation gadget in [Figure \[f:teleport\]]{} in the circuit $C'$ is equivalent to a Pauli error or non-error in the original edge in $C$. (This is a standard fact and is left as an exercise to the reader. Note that even though multi-Pauli errors reduce to Pauli errors, the relative phase of two multi-Pauli errors might change.)
Suppose that the circuit $C'$ has quasi-independent error ${\epsilon}$ by [Theorem \[th:main\]]{}. Suppose that each gate in [Figure \[f:teleport\]]{} is available as a single gate in the gate set. Then 1 edge in $C$ is replaced by 8 edges in $C'$. We suppose that $C'$ has quasi-independent error with bound ${\epsilon}$, say by [Theorem \[th:main\]]{}. We suppose for simplicity that the total error amplitude of a multi-Pauli is at most ${\epsilon}^{w(J)}$ rather than $O({\epsilon}^{w(J)})$, although the calculation works either way. The total amplitude of all $4^8$ multi-Pauli operators on the edges of a teleportation gadget is at most $(1+3{\epsilon})^8$. One of these is the term in which all edges are assigned $Z_0 = I$, the non-error; the other errors are bounded by $(1+3{\epsilon})^8 - 1$. It follows that if we interpret $C'$ as an encoding of $C$, then $C$ has quasi-independent error with bound $$\delta < (1+3{\epsilon})^8 - 1.$$ (In fact, we can divide the right side by 3 by symmetry between the non-trivial Pauli errors, but it is not necessary.)
A complexity class interpretation {#s:complexity}
=================================
It is interesting to consider complexity classes with a bounded amount of available quantum computation. One example of such a class is the *Fourier hierarchy* ${\mathsf{FH}}$ [@Shi:tradeoff]. In the Fourier hierarchy, the entire circuit is quantum in the sense that it consists of qubits, but only a bounded number of layers of Hadamard gates are allowed. In between these layers the circuit is *pseudoclassical*, meaning that it is a unitary dilation of classical circuit.
Here we define two other mixed quantum-classical classes, even though we do not know whether they are actually useful. (They should not necessarily be added to the Complexity Zoo [@zoo].) First, we can consider the class ${\mathsf{SQCL}}$, or *sandwiched quantum and classical layers* ([Figure \[f:sandwich\]]{}(a)). We represent this class by a uniform family of polynomial-sized quantum-classical circuits $\{C_n\}$. We assume that qubits are only allowed in the circuit in disjoint, global layers $[t,t+b]$ that are bounded in depth by a constant $b$. In between the layers, all edges have to bit edges Even though a circuit in ${\mathsf{SQCL}}$ can have a polynomial number of layers, the fact that no quantum coherence connects any two of the layers is a much more severe restriction than in ${\mathsf{FH}}$. A single quantum layer is a functional class known as ${\mathsf{QNC}}_0$, and surprisingly even this class seems different from classical computation [@FGHZ:constant; @BJS:collapse]. The class ${\mathsf{SQCL}}$ can be viewed as ${\mathsf{BPP}}^{{\mathsf{QNC}}_0}$, or ${\mathsf{BPP}}$ with oracle access to ${\mathsf{QNC}}_0$, except that it is a semantic type of oracle access in which the oracle output is a probability distribution.
We do not know whether ${\mathsf{QNC}}_0$ is weaker with noisy gates. More precisely, whether there is a fault tolerance noise threshold below which ${\mathsf{QNC}}_0$ is no weaker than before.
We alter the definition of ${\mathsf{SQCL}}$ subtly but dramatically. We define the class ${\mathsf{AQCL}}$, or *asynchronous quantum and classical layers* with quantum circuits as follows: Each qubit path in the circuit has depth at most $b$, but the qubits do not have to disappear at the same time. (See [Figure \[f:sandwich\]]{}(b).) This definition matches the conclusion of [Theorem \[th:knill\]]{}, so we obtain this corollary:
${\mathsf{AQCL}}= {\mathsf{BQP}}$.
[10]{}
Dorit Aharonov and Michael Ben-Or, *Fault-tolerant quantum computation with constant error rate*, STOC ’97 (El Paso, TX), ACM Press, New York, 1999, , pp. 176–188.
Charles H. Bennett, Gilles Brassard, Claude Crépeau, Richard Jozsa, Asher Peres, and William K. Wootters, *Teleporting an unknown quantum state via dual classical and einstein-podolsky-rosen channels*, Phys. Rev. Lett. **70** (1993), 1895–1899.
Michael J. Bremner, Richard Jozsa, and Dan J. Shepherd, *Classical simulation of commuting quantum computations implies collapse of the polynomial hierarchy*, Proc. R. Soc. Lond. Ser. A Math. Phys. Eng. Sci. **467** (2011), no. 2126, 459–472, .
Stephen Fenner, Frederic Green, Steven Homer, and Yong Zhang, *Bounds on the power of constant-depth quantum circuits*, Fundamentals of computation theory, Lecture Notes in Comput. Sci., vol. 3623, Springer, Berlin, 2005, , pp. 44–55.
Grace Hopper, *The first bug*, Ann. Hist. Comput. **3** (1981), no. 3, 285–286.
Gil Kalai, *Quantum computers: noise propagation and adversarial noise models*, 2009, .
[to3em]{}, *How quantum computers fail: quantum codes, correlations in physical systems, and noise accumulation*, 2011, .
Alexei Kitaev, *Quantum error correction with imperfect gates*, Quantum communication, computing, and measurement (Osamu Hirota, Alexander S. Holevo, and Carlton Caves, eds.), Springer, 1997, pp. 181–188.
Emanuel Knill, *Fault-tolerant postselected quantum computation: Schemes*, 2004, .
[to3em]{}, *Scalable quantum computing in the presence of large detected-error rates*, Phys. Rev. A **71** (2005), 042322, .
Emanuel Knill, Raymond Laflamme, and Wojciech H. Zurek, *Resilient quantum computation: error models and thresholds*, Proc. R. Soc. Lond. A **454** (1998), no. 1969, 365–384, .
Greg Kuperberg, *The capacity of hybrid quantum memory*, IEEE Trans. Inform. Theory **49** (2003), no. 6, 1465–1473, .
Michael A. Nielsen and Isaac L. Chuang, *Quantum computation and quantum information*, Cambridge University Press, Cambridge, 2000.
Robert Raussendorf, Jim Harrington, and Kovid Goyal, *A fault-tolerant one-way quantum computer*, Ann. Physics **321** (2006), no. 9, 2242–2270, .
[to3em]{}, *Topological fault-tolerance in cluster state quantum computation*, New J. Phys. **9** (2007), no. 6, 199, .
Yaoyun Shi, *Quantum and classical tradeoffs*, Theoret. Comput. Sci. **344** (2005), no. 2-3, 335–345, .
*[The Complexity Zoo]{}*, <http://www.complexityzoo.com/>.
[^1]: Supported by ERC advanced grant 320924, ISF grant 768/12, and NSF grant DMS-1300120
[^2]: Partly supported by NSF grant CCF-1319245
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'We consider variations of the standard leptogenesis picture arising from the presence of an additional scale related to the breaking of a $U(1)_X$ abelian flavor symmetry. We show that quite generically the presence of an additional energy scale might introduce new qualitative and quantitative changes on leptogenesis. Especially interesting is the possibility of having succesful TeV leptogenesis with a vanishing total CP violating asymmetry. By solving the corresponding Boltzmann equations it is shown that these kind of scenarios encounters no difficulties in generating the Cosmic baryon asymmetry.'
address:
- '$^a$ INFN, Laboratori Nazionali di Frascati,C.P. 13, I00044 Frascati, Italy'
- '$^b$ Instituto de Física, Universidad de Antioquia, A.A.[*[1226]{}*]{}, Medellín, Colombia'
author:
- 'D Aristizabal Sierra$^{a,}$[^1], L A Muñoz$^b$ and E Nardi$^{a,b}$'
title: Implications of an additional scale on leptogenesis
---
Introduction {#sec:int}
============
From observations of light element abundances and of the Cosmic microwave background radiation [@Hinshaw:2008kr] the Cosmic baryon asymmetry, $$\label{eq:baryon-asymm}
{\cal Y}_B=(8.75\pm 0.23)\times 10^{-10}\,,$$ can be inferred. The conditions for a dynamical generation of this asymmetry (baryogenesis) are well known [@Sakharov:1967dj] and depending on how they are realized different scenarios for baryogenesis can be defined (see ref. [@Dolgov:1991fr] for a througout discussion).
Leptogenesis [@Fukugita:1986hr] is a scenario in which an initial lepton asymmetry, generated in the out-of-equilibrium decays of heavy standard model singlet Majorana neutrinos ($N_\alpha$), is partially converted in a baryon asymmetry by anomalous sphaleron interactions [@Kuzmin:1985mm] that are standard model processes. Singlet Majorana neutrinos are an essential ingredient for the generation of light neutrino masses through the seesaw mechanism [@Minkowski:1977sc]. This means that if the seesaw is the source of neutrino masses then qualitatively leptogenesis is unavoidable. Consequently, whether the baryon asymmetry puzzle can be solved within this framework turn out to be a quantitative question. This has triggered a great deal of interest on quantitative analysis of the standard leptogenesis model and indeed a lot of progress during the last years have been achieved (see ref. [@Davidson:2008bu] for details).
Here we focus on variations of the standard leptogenesis picture which can arise if, apart from the lepton number breaking scale ($M_N$), an additional energy scale, related to the breaking of a new symmetry, exist. We consider a simple realization of this idea in which at an energy scale of the order of the lepton number violating scale the tree level coupling linking light and heavy neutrinos is forbiden by an exact $U(1)_X$ flavor symmetry which below (or above) $M_N$ becomes spontaneously broken by the vacuum expectation value, $\sigma$, of a standard model singlet scalar field $S$ and involves heavy vectorlike fields $F_a$. As will be discussed, according to the relative size of the relevant scales of the model ($M_{N_\alpha}$, $\sigma$, $M_{F_a}$), different scenarios for leptogenesis can be defined [@AristizabalSierra:2007ur]. Of particular interest is the case in which the total CP violating (CPV) asymmetries in the decays and scatterings of the singlet seesaw neutrinos vanish. As we will discuss further on two remarkable features distinguish this scenario [@amn]: ($a$) Flavor effects are entirely responsible for successful leptogenesis; ($b$) leptogenesis can be lowered down to the TeV scale.
The rest of this paper is organized as follows: In section \[sec:model\] we briefly discuss the model while in section \[sec:different-possibilities\] we discuss two particular realizations of our scheme paying special attention to the [*purely flavored leptogenesis*]{} (PFL) case for which we analyse the evolution of the generated lepton asymmetry by solving the corresponding Boltzmann equations (BE).
The model {#sec:model}
=========
The model we consider here [@AristizabalSierra:2007ur] is a simple extension of the standard model containing a set of $SU(2)_L\times U(1)_Y$ fermionic singlets, namely three right-handed neutrinos ($N_\alpha = N_{\alpha
R} + N_{\alpha R}^c$) and three heavy vectorlike fields ($F_a=F_{aL} +
F_{aR}$). In addition, we assume that at some high energy scale, taken to be of the order of the leptogenesis scale $M_{N_1}$, an exact $U(1)_X$ horizontal symmetry forbids direct couplings of the lepton $\ell_i$ and Higgs $\Phi$ doublets to the heavy Majorana neutrinos $N_\alpha$. At lower energies, $U(1)_X$ gets spontaneously broken by the vacuum expectation value (vev) $\sigma$ of a $SU(2)$ singlet scalar field $S$. Accordingly, the Yukawa interactions of the high energy Lagrangian read $$\label{eq:lag}
-{\cal L}_Y =
\frac{1}{2}\bar{N}_{\alpha}M_{N_\alpha}N_{\alpha} +
\bar{F}_{a}M_{F_a}F_{a} +
h_{ia}\bar{\ell}_{i}P_{R}F_{a}\Phi +
\bar{N}_{\alpha}
\left( \lambda_{\alpha a} + \lambda^{(5)}_{\alpha a}\gamma_5\right)
F_{a}S
+ \mbox{h.c.}$$ We use Greek indices $\alpha,\beta\dots =1,2,3$ to label the heavy Majorana neutrinos, Latin indices $a,b\dots =1,2,3$ for the vectorlike messengers, and $i, j, k, \dots$ for the lepton flavors $e,\mu,\tau$. Following reference [@AristizabalSierra:2007ur] we chose the simple $U(1)_X$ charge assignments $X(\ell_{L_i},F_{L_a},F_{R_a})=+1$, $X(S)=-1$ and $X(N_{\alpha},\Phi)=0$. This assignment is sufficient to enforce the absence of $\bar N \ell \Phi$ terms, but clearly it does not constitute an attempt to reproduce the fermion mass pattern, and accordingly we will also avoid assigning specific charges to the right-handed leptons and quark fields that have no relevance for our analysis. The important point is that it is likely that any flavor symmetry (of the Froggatt-Nielsen type) will forbid the the same tree-level couplings, and will reproduce an overall model structure similar to the one we are assuming here. Therefore we believe that our results, that are focused on a new realization of the leptogenesis mechanism, can hint to a general possibility that could well occur also in a complete model of flavor.
As discussed in [@AristizabalSierra:2007ur], depending on the hierarchy between the relevant scales of the model ($M_{N_1},\,M_{F_a},\,\sigma$), quite different [*scenarios*]{} for leptogenesis can arise. Here we will concentrate on two cases: ($i$) The standard leptogenesis case ($M_F,\sigma\gg M_N$); ($ii$) the PFL case ($\sigma < M_{N_1}< M_{F_a}$) that is, when the flavor symmetry $U(1)_X$ is still unbroken during the leptogenesis era and at the same time the messengers $F_a$ are too heavy to be produced in $N_1$ decays and scatterings, and can be integrated away [@amn].
As is explicitly shown by the last term in eq. (\[eq:lag\]), in general the vectorlike fields can couple to the heavy singlet neutrinos via scalar and pseudoscalar couplings, In ref. [@AristizabalSierra:2007ur] it was assumed for simplicity a strong hierarchy $\lambda\gg \lambda^{(5)}$ so $\lambda^{(5)}$ was neglected. However, in all the relevant quantities (scatterings, CP asymmetries, light neutrino masses) at leading order the scalar and pseudoscscalar couplings always appear in the combination $\lambda +
\lambda^{(5)}$, and thus such an assumption is not necessary. The replacement $\lambda \to \lambda + \lambda^{(5)}$ would suffice to include in the analysis the effects of both type of interactions.
Effective seesaw and light neutrino masses {#sec:nmgen}
------------------------------------------
![Effective mass operator responsible for neutrino mass generation[]{data-label="fig:neutrino-massmatrix"}](neutrino-mm.eps){width="8cm" height="2.5cm"}
After $U(1)_X$ and electroweak symmetry breaking the set of Yukawa interactions in (\[eq:lag\]) generate light neutrino masses through the effective mass operator shown in figure \[fig:neutrino-massmatrix\]. The resulting mass matrix can be written as [@AristizabalSierra:2007ur] $$\label{eq:nmm}
-{\cal M}_{ij}=
\left[
h^{*}\frac{\sigma}{M_{F}}\lambda^{T}\frac{v^{2}}
{M_{N}}\lambda\frac{\sigma}{M_{F}}h^{\dagger}
\right]_{ij}
= \left[
\tilde{\lambda}^{T}\frac{v^{2}}{M_{N}}\tilde{\lambda}
\right]_{ij} \,.$$ Here we have introduced the seesaw-like couplings $$\label{eq:seesaw-couplings}
\tilde{\lambda}_{\alpha i} =
\left(
\lambda \frac{\sigma}{M_F}h^\dagger
\right)_{\alpha i}\,.$$ Note that, in contrast to the standard seesaw, the neutrino mass matrix is of fourth order in the [*fundamental*]{} Yukawa couplings ($h$ and $\lambda$) and due to the factor $\sigma^2/M_F^2$ is even more suppressed.
Different scenarios for leptogenesis {#sec:different-possibilities}
====================================
In this section we discuss the features of each one of the cases we previously mentioned and derive expressions for the CP asymmetries. Henceforth we will use the following notation for the different mass ratios: $$\label{eq:mass-ratios}
z_{\alpha}=\frac{M_{N_\alpha}^2}{M_{N_1}^2},\qquad \omega_{a}=
\frac{M_{F_a}^2}{M_{F_1}^2}, \qquad r_a=\frac{M_{N_1}}{M_{F_a}}.$$
The standard leptogenesis case {#sec:standard}
------------------------------
When the masses of the heavy fields $F_a$ and the $U(1)_X$ symmetry breaking scale are both larger than the Majorana neutrino masses ($M_F\,, \sigma >
M_N$) there are no major differences from the standard Fukugita-Yanagida leptogenesis model [@Fukugita:1986hr]. After integrating out the $F$ fields one obtains the standard seesaw Lagrangian containing the effective operators $\tilde\lambda_{\alpha i} \bar N_\alpha l_i \Phi$ with the seesaw couplings $\tilde\lambda_{\alpha i}$ given in eq. (\[eq:seesaw-couplings\]). The right handed neutrino $N_1$ decays predominantly via 2-body channels as shown in fig. \[fig:case0\]. This yields the standard results that for convenience we recall here. The total decay width is $\Gamma_{N_1}= \left({M_{N_1}}/{16
\pi}\right) (\tilde\lambda\tilde\lambda^\dagger)_{11}$ and the sum of the vertex and self-energy contributions to the $CP$-asymmetry for $N_1$ decays into the flavor $l_j$ reads [@Covi:1996wh] $$\label{eq:6}
\epsilon_{N_1\to l_j}=\frac{1}{8\pi (\tilde\lambda \tilde\lambda^\dagger)_{11}}
\sum_{\beta\neq 1}{\mathbb I}\mbox{m}
\left\{\tilde\lambda_{\beta j}\tilde\lambda^*_{1j} \left[
% \frac{3}{2\sqrt{z_\beta}}
(\tilde\lambda\tilde\lambda^\dagger)_{\beta1} \tilde F_1(z_\beta) +
% \frac{1}{z_\beta}
(\tilde\lambda\tilde\lambda^\dagger
)_{1\beta}
\tilde F_2(z_\beta)
\right]\right\}\,,$$ where $$\label{eq:tildeF}
\tilde F_1(z )= \frac{\sqrt{z }}{1-z } +
\sqrt{z }\left(1-(1+z )\ln
\frac{1+z }{z }\right),
%= -\frac{3}{2\sqrt{z }} +{\cal O}(\frac{1}{\sqrt{z^3}})
% -\frac{5}{6}x^{-3/2}+{\cal O}(x^{-5/2})
% \\
\qquad \tilde F_2(z)=\frac{1}{1-z}.$$ At leading order in $1/z_\beta$ and after summing over all leptons $l_j$, eq. (\[eq:6\]) yields for the total asymmetry: $$\label{eq:8}
\epsilon_{N_1}=\frac{3}{16\pi (\tilde\lambda \tilde\lambda^\dagger)_{11}}
\sum_{\beta}
{\mathbb I}\mbox{m}
\left\{\frac{1}{\sqrt{z_\beta}}
(\tilde\lambda\tilde\lambda^\dagger)_{\beta1}^2 \right\}.$$ where the sum over the heavy neutrinos has been extended to include also $N_1$ since for $\beta=1$ the corresponding combination of couplings is real.
In the hierarchical case $M_{N_1}\ll M_{N_{2,3}}$ the size of the total asymmetry in (\[eq:8\]) is bounded by the Davidson-Ibarra limit [@Davidson:2002qv] $$\label{eq:DI}
|\epsilon_{N_1}|\leq \frac{3}{16\pi}\frac{M_{N_1}}{v^2} \,(m_{\nu_3}-m_{\nu_1})
\lesssim \frac{3}{16\pi}\frac{M_{N_1}}{v^2} \frac{\Delta m^2_{\text{atm}}}{2m_{\nu_3}}\,,$$ where $m_{\nu_i}$ (with $m_{\nu_1} < m_{\nu_2} < m_{\nu_3}$) are the light neutrinos mass eigenstates and $\Delta m^2_{\text{atm}}\sim
2.5\times 10^{-3}\,$eV$^2$ is the atmospheric neutrino mass difference [@Schwetz:2008er]. It is now easy to see that (\[eq:DI\]) implies a lower limit on $M_{N_1}$. The amount of $B$ asymmetry that can be generated from $N_1$ dynamics can be written as: $$\label{eq:etaB}
\frac{n_B}{s}=-\kappa_s\,\epsilon_{N_1}\,\eta ,$$ where $\kappa_s\approx 1.3\times 10^{-3}$ accounts for the dilution of the asymmetry due to the increase of the Universe entropy from the time the asymmetry is generated with respect to the present time, $\eta $ (that can range between 0 and 1, with typical values $10^{-1}-10^{-2}$) is the [*efficiency factor*]{} that accounts for the amount of $L$ asymmetry that can survive the washout process. Assuming that $\epsilon_{N_1}$ is the main source of the $B-L$ asymmetry [@Engelhard:2006yg], eqs. (\[eq:DI\]) and (\[eq:etaB\]) in addition from the oserved baryon asymmetry eq. (\[eq:baryon-asymm\]) yield: $$\label{eq:M1limit0}
M_{N_1} \gtrsim 10^{9}\,
\frac{m_{\nu_3}}{\eta\, \sqrt{\Delta m^2_{\text{atm}}}} \, {\rm GeV}.$$ This limit can be somewhat relaxed depending on the specific initial conditions [@Giudice:2003jh] or when flavor effects are included [@Nardi:2006fx; @Abada:2006fw; @Abada:2006ea; @JosseMichaux:2007zj] but the main point remains, and that is that the value of $M_{N_1}$ should be well above the electroweak scale.
![Diagrams responsible for the CP violating asymmetry in the standard case[]{data-label="fig:case0"}](standard-lep.eps){width="10cm" height="2.5cm"}
Purely flavored leptogenesis case {#sec:PFL}
---------------------------------
Differently from standard leptogenesis in the present case, since $M_F >
M_{N_1}$, two-body $N_1$ decays are kinematically forbidden. However, via off-shell exchange of the heavy $F_a$ fields, $N_1$ can decay to the three body final states $S\Phi l$ and $\bar S\bar\Phi \bar l$. The corresponding Feynman diagram is depicted in figure \[fig:fig0\]$(a)$. At leading order in $r_a=M_{N_1}/M_{F_a}$, the total decay width reads [@AristizabalSierra:2007ur] $$\label{eq:total-decay-width}
\Gamma_{N_1}\equiv
\sum_j \Gamma(N_1\to S\Phi l_j + \bar S\bar\Phi \bar l_j) =\frac{M_{N_1}}{192\pi^3}
\left(
\frac{M_{N_1}}{\sigma}
\right)^2
(\tilde{\lambda}\tilde{\lambda}^\dagger)_{11}\,.$$ As usual, CPV asymmetries in $N_1$ decays arise from the interference between tree-level and one-loop amplitudes. As was noted in [@AristizabalSierra:2007ur], in this model at one-loop there are no contributions from vertex corrections, and the only contribution to the CPV asymmetries comes from the self-energy diagram \[fig:fig0\]$(b)$. Summing over the leptons and vectorlike fields running in the loop, at leading order in $r_a$ the CPV asymmetry for $N_1$ decays into leptons of flavor $j$ can be written as $$\label{eq:cp-violating-asymm}
\epsilon_{1j} \equiv \epsilon_{N_1\to\ell_j} =
\frac{3}{128\pi}
\frac{\sum_{m} \mathbb{I}\mbox{m}
\left[
\left(
h r^{2} h^{\dagger}
\right)_{mj}\tilde{\lambda}_{1m}\tilde{\lambda}^{*}_{1j}
\right]}{\left(\tilde{\lambda}\tilde{\lambda}^{\dagger}\right)_{11}}\,.$$ Note that since the loop correction does not violate lepton number, the total CPV asymmetry that is obtained by summing over the flavor of the final state leptons vanishes [@Kolb:1979qa], that is $\epsilon_{1}\equiv \sum_j
\epsilon_{1j}=0$. This is the condition that defines PFL; namely there is no CPV [*and*]{} lepton number violating asymmetry, and the CPV lepton flavor asymmetries are the only seed of the Cosmological lepton and baryon asymmetries.
It is important to note that the effective couplings $\tilde\lambda$ defined in eq. (\[eq:seesaw-couplings\]) are invariant under the reparameterization $$\label{eq:couplingtrans}
\lambda\to \lambda\cdot (rU)^{-1},\quad
h^\dagger \to (U r)\cdot h^\dagger\,,$$
![Feynman diagrams responsible for the CPV asymmetry.[]{data-label="fig:fig0"}](cp-asymm-plot.eps){width="10cm" height="3cm"}
where $U$ is an arbitrary $3\times 3$ non-singular matrix. Clearly the light neutrino mass matrix is invariant under this transformation. Moreover, also the flavor dependent washout processes, that correspond to tree level amplitudes that are determined, to a good approximation, by the effective $\tilde \lambda$ couplings, are left essentially unchanged.[^2] On the contrary, the flavor CPV asymmetries eq. (\[eq:cp-violating-asymm\]), that are determined by loop amplitudes containing an additional factor of $h r^2 h^\dagger$, get rescaled as $h r^2 h^\dagger\to h (rUr)^\dagger (rUr) h^\dagger $. Clearly, this rescaling affects in the same way all the lepton flavors (as it should be to guarantee that the PFL conditions $\epsilon_\alpha\equiv \sum\epsilon_{\alpha j}=0$ are not spoiled), and thus for simplicity we will consider only rescaling by a global scalar factor $r.U=U.r=\kappa\,I$ (with $I$ the $3\times 3$ identity matrix) that, for our purposes, is completely equivalent to the more general transformation (\[eq:couplingtrans\]). Thus, while rescaling the Yukawa couplings through $$\label{eq:coupling-rescaling-gen}
\lambda\to \lambda \,\kappa^{-1},\quad
h^\dagger\to\kappa \,h^\dagger\,,$$ does not affect neither low energy neutrino physics nor the washout processes, the CPV asymmetries get rescaled as: $$\label{eq:rescaled-CPV-asymm}
\epsilon_{1j}\to\kappa^2\epsilon_{1j}\,.$$ By choosing $\kappa>1$, all the CPV asymmetries get enhanced as $\kappa^2$ and, being the Cosmological asymmetries generated through leptogenesis linear in the CPV asymmetries, the final result gets enhanced by the same factor. Therefore, for any given set of couplings, one can always find an appropriate rescaling such that the correct amount of Cosmological lepton asymmetry is generated. In practice, the rescaling factors $\kappa$ cannot be arbitrarily large: first, they should respect the condition that all the fundamental Yukawa couplings remain in the perturbative regime; second the size of the $h$ couplings (and thus also of the rescaling parameter $\kappa$) is also constrained by experimental limits on lepton flavor violating decays.
![Feynman diagrams for $1\leftrightarrow 3$ and $2\leftrightarrow 2$ $s$, $t$ and $u$ channel processes after integrating out the heavy vectorlike fields $F_a$.[]{data-label="fig:fig1"}](decaySc.eps){width="13.0cm" height="2.3cm"}
### Boltzmann Equations {#sec:BE}
In this section we compute the lepton asymmetry by solving the appropriate BE. In general, to consistently derive the evolution equation of the lepton asymmetry all the possible processes at a given order in the couplings have to be included. In the present case $1\leftrightarrow 3$ decays and inverse decays, and $2\leftrightarrow
2$ $s$, $t$ and $u$ channel scatterings all occur at the same order in the couplings and must be included altogether in the BE. The Feynman diagrams for these processes are shown in Figure \[fig:fig1\]. In addition, the CPV asymmetries of some higher order multiparticle reactions involving the exchange of one off-shell $N_1$, also contribute to the source term of the asymmetries at the same order in the couplings than the CPV asymmetries of decays and $2\leftrightarrow
2$ scatterings. More precisely, for a proper derivation of the BE it is essential that the CPV asymmetries of the off-shell $3\leftrightarrow 3$ and $2\leftrightarrow 4$ scattering processes are also taken into account [@amn].
As regards the equation for the evolution of the heavy neutrino density $Y_{N_1}$, only the diagrams in fig. \[fig:fig1\], that are of leading order in the couplings, are important [@amn] $$\begin{aligned}
\label{eq:BEforN-section}
\dot{Y}_{N_1} &=
-\left(y_{N_1} - 1 \right) \gamma_{\text{tot}}\,, \\
\label{eq:BEforLA-section}
\dot{Y}_{\Delta \ell_{i}} &=
\left(y_{N_1} - 1 \right)\epsilon_{i}\gamma_{\text{tot}} - \Delta y_{i}
\left[
\gamma_{i} + \left(y_{N_1} - 1 \right)\gamma^{N_{1}\bar\ell_{i}}_{S\Phi}
\right]\,, \end{aligned}$$ Here we have normalized particle densities to their equilibrium densities $y_a\equiv Y_a/Y_a^{\text{eq}}$ where $Y_a=n_a/s$ with $n_a$ the particle number density and $s$ the entropy density. The time derivative is defined as $\dot Y = sHz\,dY/dz$ with $z=M_{N_1}/T$ and $H$ is the Hubble parameter. In the last term of the second equation we have used the compact notation for the reaction densities $\gamma^{N_{1}\bar\ell_{i}}_{S\Phi} =\gamma(N_{1}\bar\ell_{i}\to
{S\Phi})$ and in addition we have defined $$\begin{aligned}
\label{eq:rates}
\gamma_{i} &=&
\gamma^{N_{1}}_{S\ell_{i}\Phi}+
\gamma^{N_{1}\bar{S}}_{\Phi\ell_{i}} +
\gamma^{N_{1}\bar{\Phi}}_{S\ell_{i}} + \gamma_{S\Phi}^{N_{1}\bar\ell_{i}}\,, \\
\gamma_{\text{tot}} &=& \sum_{i=e,\mu,\tau}
\gamma_{i}+\bar\gamma_i\,, \end{aligned}$$ where in the second equation $\bar\gamma_i$ represents the sum of the CP conjugates of the processes summed in $\gamma_{i}$.
Since in this model $N_1$ decays are of the same order in the couplings than scatterings (that is ${\cal O}(\tilde\lambda^2)$), the appropriate condition that defines the [*strong washout*]{} regime in the case at hand reads: $$\label{eq:strong-washout}
\left .\frac{\gamma_{\text{tot}}}{z\,H\,s}\right|_{z\sim 1}>1 \,\qquad {\rm
(strong\ washout)},$$ and conversely $\gamma_{\text{tot}}/(z\,H\,s)|_{z\sim 1}<1$ defines the [ *weak washout*]{} regime. Note that this is different from standard leptogenesis, where at $z\sim 1$ two body decays generally dominate over scatterings, and e.g. the condition for the strong washout regime can be approximated as $\gamma_{\text{tot}}/(z\,H\,s)|_{z\sim 1}\sim
\Gamma_{N_1}/H|_{z\sim 1}>1$.
### Results {#sec:results}
In this section we discuss a typical example of successful leptogenesis at the scale of a few TeV. The example presented is a general one. No particular choice of the parameters has been performed, except for the fact that the low energy neutrino data are reproduced within errors, and that the choice yields an interesting washout dynamics well suited to illustrate how PFL works. The numerical value of the final lepton asymmetry ($Y_{\Delta L} \sim
-7.2\times 10^{-10}$) is about a factor of 3 [*larger*]{} than what is indicated by measurements of the Cosmic baryon asymmetry. This is however irrelevant since, as was already discussed, it would be sufficient a minor rescaling of the couplings (or a slight change in the CPV phases) to obtain the precise experimental result. In the numerical analysis we have neglected the dynamics of the heavier singlet neutrinos since the $N_\alpha$ masses are sufficiently hierarchical to ensure that $N_{2,3}$ related washouts do not interfere with $N_1$ dynamics. Moreover, in the (strong washout) fully flavored regime (that is effective as long as $T < 10^{9}\,$GeV) the $N_{2,3}$ CPV asymmetries do not contribute to the final lepton number asymmetry [@Engelhard:2006yg].
In figure \[fig:fig4\] we show the behavior of the various reaction densities for decays and scatterings, normalized to $sHz$, as a function of $z$. The results correspond to a mass of the lightest singlet neutrino fixed to $M_{N_1}=2.5\,\text{TeV}$, the heavier neutrino masses are $
M_{N_2}=10\,$TeV and $M_{N_3}=15\,$TeV, and the relevant mass ratios $r_a=M_{N_1}/M_{F_a}$ for the messenger fields are $r_{1,2,3}=0.1,0.01,0.001$ (the effects of the lightest $F$ resonances can be seen in the $s$-channel rates in both panels in fig. \[fig:fig4\]). The fundamental Yukawa couplings $h$ and $\lambda$ are chosen to satisfy the requirement that the seesaw formula eq. (\[eq:nmm\]) reproduces within $2\,\sigma$ the low energy data on the neutrino mass squared differences and mixing angles [@Schwetz:2008er]. Typically, when this requirement is fulfilled, one also ends up with a dynamics for all the lepton flavors in the strong washout regime. This is shown in the left panel in fig. \[fig:fig5\] where we present the total rates for the three flavors.
[![ Reaction densities normalized to $zHs$ for $N_1\to S\ell\Phi$ decays (red solid lines), $s$-channel $\bar S N_1
\leftrightarrow \ell \Phi$ scatterings (green dashed lines), and $t,u$-channel scatterings in the point-like approximation (blue dotted lines). Left panel: $\tau$ flavor. Right panel: electron flavor. []{data-label="fig:fig4"}](tau-rates.eps "fig:"){width="6.5cm" height="5.5cm"}]{}
The left panel in fig. \[fig:fig4\] refers to the decay and scattering rates involving the $\tau$-flavor that, in our example, is the flavor more strongly coupled to $N_1$, and that thus suffers the strongest washout. It is worth noticing that, due to the fact that in this model scatterings are not suppressed by additional coupling constants with respect to the decays, the decay rate starts dominating the washouts only at $z\gtrsim 1$. The right panel in fig. \[fig:fig4\] depicts the reaction rates for the electron flavor, that is the more weakly coupled, and for which the strong washout condition eq. (\[eq:strong-washout\]) is essentially ensured by sizeable $s$-channel scatterings. Scatterings and decay rates for the $\mu$-flavor are not shown, but they are in between the ones of the previous two flavors.
[![Left panel: the total washout rates for each lepton flavor normalized to $zHs$ as a function of $z$. Right panel: the evolution of the absolute value of the flavored density asymmetries and of the lepton number asymmetry (black solid line). The flavor CPV asymmetries are $\epsilon_{1e} = -4.7\times 10^{-4}$, $\epsilon_{1\mu} = -1.9\times
10^{-4}$ and $\epsilon_{1\tau} = 6.6\times 10^{-4}$. The final values of the asymmetry densities (at $z\gg 1$) are $Y_{\Delta_{\ell_e}} =-7.1\times
10^{-10}$, $Y_{\Delta_{\ell_\mu}} =-0.3\times 10^{-10}$, $Y_{\Delta_{\ell_\tau}} =0.2\times 10^{-10}$. []{data-label="fig:fig5"}](total-rates.eps "fig:"){width="6.5cm" height="5.5cm"}]{}
The total reaction densities that determine the washout rates for the different flavors are shown in the first panel in figure \[fig:fig5\]. The evolution of these rates with $z$ should be confronted with the evolution of (the absolute value of) the asymmetry densities for each flavor, depicted in the second panel on the right. Since, as already stressed several times, PFL is defined by the condition that the sum of the flavor CPV asymmetry vanishes ($\sum_j \epsilon_{1j}=0$), it is the hierarchy between these washout rates that in the end is the responsible for generating a net lepton number asymmetry. In the case at hand, the absolute values of the flavor CPV asymmetries satisfy the condition $|\epsilon_\mu| < |\epsilon_e| <
|\epsilon_\tau|$, as can be inferred directly by the fact that at $z< 0.1$, when the effects of the washouts are still negligible, the asymmetry densities satisfy this hierarchy. Moreover, since $\epsilon_{\mu,e} < 0$ while $\epsilon_\tau>0$, initially the total lepton number asymmetry, that is dominated by $Y_{\Delta_{\ell_\tau}}$, is positive. As washout effects become important, the $\tau$-related reactions (blue dotted line in the left panel) start erasing $Y_{\Delta_{\ell_\tau}}$ more efficiently than what happens for the other two flavors, and thus the initial positive asymmetry is driven towards zero, and eventually changes sign around $z=0.2$. This change of sign corresponds to the steep valley in the absolute value $|Y_{\Delta L}|$ that is drawn in the figure with a black solid line. Note that when all flavors are in the strong washout regime, as in the present case, the condition for the occurrence of this ‘sign inversions’ is simply given by $ {\rm max_{j\in e,\mu}}
\left(|\epsilon_j|/|\tilde\lambda_{1j}|^2\right)
\gtrsim \epsilon_\tau/|\tilde\lambda_{1\tau}|^2$. From this point onwards, the asymmetry remains negative, and since the electron flavor is the one that suffers the weakest washout, $Y_{\Delta_{\ell_e}}$ ends up dominating all the other density asymmetries. In fact, as can be seen from the right panel in fig. \[fig:fig5\], it is $Y_{\Delta_{\ell_e}}$ that determines to a large extent the final value of the lepton asymmetry $Y_{\Delta L}=-7.2\times 10^{-10}$.
A few comments are in order regarding the role played by the $F_a$ fields. Even if $M_{N_1}\ll M_{F_a}$, at large temperatures $z\gg 1$ the tail of the thermal distributions of the $N_1,\,S$ and $\Phi$ particles allows the on-shell production of the lightest $F$ states. A possible asymmetry generated in the decays of the $F$ fields can be ignored for two reasons: first because due to the rather large $h$ and $\lambda$ couplings $F$ decays occur to a good approximation in thermal equilibrium, ensuring that no sizeable asymmetry can be generated, and second because the strong washout dynamics that characterizes $N_1$ leptogenesis at lower temperatures is in any case insensitive to changes in the initial conditions.
In conclusion, it is clear from the results of this section that the model encounters no difficulties to allow for the possibility of generating the Cosmic baryon asymmetry at a scale of a few TeVs. Moreover, our analysis provides a concrete example of PFL, and shows that the condition $\epsilon_1\neq 0$ is by no means required for successful leptogenesis.
Conclusions
===========
Variations of the standard leptogenesis picture can arise from the presence of an additional energy scale different from that of lepton number violation. Quite generically the resulting scenarios are expected to yield qualitative and quantitative changes on leptogenesis. Here we have considered what we regard as the simplest possibility namely, the presence of an abelian flavor symmetry $U(1)_X$. We have described two possible scenarios within this framework and have explored their implications for leptogenesis.
We have found that as long as the abelian flavor symmetry energy scales remain above the lepton number violating scale neither qualitative nor quantitative differences with the standard leptogenesis model arise. Conversely if the $U(1)_X$ is unbroken during the leptogenesis era and the messengers fields $F_a$ are too heavy to be produced on-shell in $N_1$ decays [*purely flavored leptogenesis*]{} at the TeV scale results. By solving the corresponding BE we have shown that within this scenario the non-vanishing of the CPV lepton flavor asymmetries in addition to the lepton and flavor violating washout processes occuring in the plasma provide the necessary ingredients to generate the Cosmic baryon asymmetry. Accordingly, if below the leptogenesis scale new energy scales are present -as might be expected- the interplay between these scales could have a quite interesting impact on leptogenesis.
References {#references .unnumbered}
==========
[99]{} G. Hinshaw [*et al.*]{} \[WMAP Collaboration\], Astrophys. J. Suppl. [**180**]{}, 225 (2009) \[arXiv:0803.0732 \[astro-ph\]\]. A. D. Sakharov, Pisma Zh. Eksp. Teor. Fiz. [**5**]{} (1967) 32 \[JETP Lett. [**5**]{} (1967 SOPUA,34,392-393.1991 UFNAA,161,61-64.1991) 24\]. A. D. Dolgov, Phys. Rept. [**222**]{}, 309 (1992). M. Fukugita and T. Yanagida, Phys. Lett. B [**174**]{}, 45 (1986). V. A. Kuzmin, V. A. Rubakov and M. E. Shaposhnikov, Phys. Lett. B [**155**]{}, 36 (1985). P. Minkowski, Phys. Lett. B [**67**]{}, 421 (1977); R. N. Mohapatra and G. Senjanovic, Phys. Rev. Lett. [**44**]{}, 912 (1980). S. Davidson, E. Nardi and Y. Nir, Phys. Rept. [**466**]{}, 105 (2008) \[arXiv:0802.2962 \[hep-ph\]\]. D. Aristizabal Sierra, M. Losada and E. Nardi, Phys. Lett. B [**659**]{}, 328 (2008) \[arXiv:0705.1489 \[hep-ph\]\]; D. Aristizabal Sierra, Luis Alfredo Munoz and Enrico Nardi, arXiv:0904.3043 \[hep-ph\].
L. Covi, E. Roulet and F. Vissani, Phys. Lett. B [**384**]{}, 169 (1996) \[arXiv:hep-ph/9605319\]. S. Davidson and A. Ibarra, Phys. Lett. B [**535**]{}, 25 (2002) \[arXiv:hep-ph/0202239\]. T. Schwetz, M. Tortola and J. W. F. Valle, New J. Phys. [**10**]{}, 113011 (2008) \[arXiv:0808.2016 \[hep-ph\]\]. G. Engelhard, Y. Grossman, E. Nardi and Y. Nir, Phys. Rev. Lett. [**99**]{}, 081802 (2007) \[arXiv:hep-ph/0612187\]. G. F. Giudice, A. Notari, M. Raidal, A. Riotto and A. Strumia, Nucl. Phys. B [**685**]{}, 89 (2004) \[arXiv:hep-ph/0310123\]. E. Nardi, Y. Nir, E. Roulet and J. Racker, JHEP [**0601**]{}, 164 (2006) \[arXiv:hep-ph/0601084\]. A. Abada, S. Davidson, F. X. Josse-Michaux, M. Losada and A. Riotto, JCAP [**0604**]{}, 004 (2006) \[arXiv:hep-ph/0601083\]. A. Abada, S. Davidson, A. Ibarra, F. X. Josse-Michaux, M. Losada and A. Riotto, JHEP [**0609**]{}, 010 (2006) \[arXiv:hep-ph/0605281\]. F. X. Josse-Michaux and A. Abada, JCAP [**0710**]{}, 009 (2007) \[arXiv:hep-ph/0703084\]. E. W. Kolb and S. Wolfram, Nucl. Phys. B [**172**]{}, 224 (1980) \[Erratum-ibid. B [**195**]{}, 542 (1982)\].
[^1]: Talk given by D. Aristizabal Sierra at the Discrete’08 Symposium, 11-16 Dec. 2008, Valencia-Spain.
[^2]: The approximation is exact in the limit of pointlike $F$-propagators $(s-M^2_F+iM_F\Gamma_F)\to M^2_F$.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'Emotion cause analysis such as emotion cause extraction (ECE) and emotion-cause pair extraction (ECPE) have gradually attracted the attention of many researchers. However, there are still two shortcomings in the existing research: 1) In most cases, emotion expression and cause are not the whole clause, but the span in the clause, so extracting the clause-pair rather than the span-pair greatly limits its applications in real-world scenarios; 2) It is not enough to extract the emotion expression clause without identifying the emotion categories, the presence of emotion clause does not necessarily convey emotional information explicitly due to different possible causes. In this paper, we propose a new task: Emotion-Cause Span-Pair extraction and classification (ECSP), which aims to extract the potential span-pair of emotion and corresponding causes in a document, and make emotion classification for each pair. In the new ECSP task, ECE and ECPE can be regarded as two special cases at the clause-level. We propose a span-based extract-then-classify (ETC) model, where emotion and cause are directly extracted and paired from the document under the supervision of target span boundaries, and corresponding categories are then classified using their pair representations and localized context. Experiments show that our proposed ETC model outperforms the SOTA model of ECE and ECPE task respectively and gets a fair-enough results on ECSP task.'
author:
- |
Hongliang Bi, Pengyuan Liu\
Beijing Language and Culture University, China\
[[email protected], [email protected]]{}\
bibliography:
- 'coling2020.bib'
title: |
ECSP: A New Task for Emotion-Cause Span-Pair\
Extraction and Classification
---
Introduction
============
Emotion cause analysis such as emotion cause extraction (ECE) and emotion-cause pair extraction (ECPE) have gradually attracted the attention of many researchers, can be constructive to guide the direction of future work, i.e., improving the quality of products or services according to the emotion causes of comments provided by users.
Emotion cause extraction (ECE) was first proposed by , which aims at discovering the potential cause clauses behind a certain emotion expression in the text. Earlier work viewed ECE as a trigger word detection problem and tries to solve it with corresponding tagging techniques. Therefore, primary efforts have been made on discovering refined linguistic features [@chen2010emotion; @lee2013detecting], yielding improved performance. More recently, instead of concentrating on word-level cause detection, clause-level extraction [@gui2016event] was putted forward in that the impact of individual words in a clause can span over the whole sequence in the clause. While ECE has attracted an increasing attention due to its theoretical and practical significance, it requires that the emotion expression annotations should be given in the test set. In light of recent advances in multi-task learning, proposed joint extraction of emotion categories and causes are investigated to exploit the mutual information between two correlated tasks, and proposed emotion-cause pair extraction (ECPE) task, which aims to extract all potential clause-pairs of emotion expression and corresponding cause in a document, and to solve the shortcomings of previous ECE task must be annotated before extraction causes. argues that, while co-extraction of emotion expression and causes are important, ECPE is a more challenging problem that is worth putting more emphases on.
However, ECPE still suffers from two shortcomings: 1) In most cases, emotion expression and cause are not the whole clause, but the span in the clause, so extracting the clause-pair rather than the span-pair greatly limits its applications in real-world scenarios; 2) It is not enough to extract the emotion expression clause without identifying the emotion categories, the presence of emotion clause does not necessarily convey emotional information due to different possible causes such as negative polarity, sense ambiguity or rhetoric. For example, “It feels like the sky is falling right on top of me” is an emotion expression of “fear”.
In this paper, we propose a new task: Emotion-Cause Span-Pair extraction and classification (ECSP), which aims to extract the potential span-pair of emotion and corresponding causes in a document, and make emotion classification for each pair. Therefore, ECE and ECPE can be regarded as two special cases of ECSP at the clause-level. Figure \[example\] is an intuitive example of the difference between the ECE, ECPE and new ECSP task.
**The 38th Document in Benchmark Corpus** clause $c_{1}$: Wang was diagnosed with chronic renal failure last April
clause $c_{2}$: This test result broke the originally happy family of three
clause $c_{3}$: Xu said, “It feels like the sky is falling right on top of me.”
clause $c_{4}$: Xu described how he felt when he learned that his husband was ill
clause $c_{5}$: Because Wang is the support of her and her two-year-old child
**Emotion Cause Extraction (ECE)**\
clause $c_{3}$ $\Longrightarrow$ clause $c_{5}$ **Emotion-Cause Pair Extraction (ECPE)**\
(clause $c_{3}$, clause $c_{5}$)
**Emotion-Cause Span-Pair Extraction and Classification (ECSP)** (It feels like the sky is falling right on top of me,\
Wang is the support of her and her two-year-old child, fear)
Inspired by recent span-based models in syntactic parsing and co-reference resolution [@lee2017end; @stern2017minimal], we propose a span-based model to solve this new ECSP task. The key insight is to annotate each emotion and cause with its span boundary followed by its emotion categories. Under such annotation, we introduce a span-based extract-then-classify (ETC) model that emotion and cause are directly extracted and paired from the document under the supervision of target span boundaries, and corresponding categories are then classified using their pair representations and localized context. The advantage of this method is that clause-based tasks and span-based tasks can be interpreted uniformly. Moreover, since the polarity is decided by using the targeted span representation, the model is able to take all target words into account before making predictions, thus naturally avoiding sentiment inconsistency.
We take BERT [@devlin2019bert] as the default backbone network, and explore the following two aspects. First, we explore the feasibility of the ECSP task under different length search schemes, and the results prove that the ECSP task can be solved well with the increase of the model search length, and there is still some room for improvement. Second, following previous works [@gui2016event; @xia2019emotion], we compare our proposed ETC model and strong baselines under the clause-based search scheme. our proposed ETC model outperforms the SOTA model of ECE and ECPE task respectively and gets a fair-enough results on ECSP task. This proves the feasibility of the ECSP task and the effectiveness of our proposed ETC model.
Proposed model
==============
![Overall illustration of our proposed ETC model.[]{data-label="model"}](model.png){width="\textwidth"}
Instead of traditional clause-based detection methods to identify emotions and causes, we propose use a span-based search scheme as follows: give an input document $D = (x_{1}, \cdots, x_{n})$ with length $n$, and a emotion-cause span-pair list $P = \{p_{1}, \cdots, p_{m}\}$, where the number of emotion-cause span-pair is $m$ and each emotion expression span $e$ and corresponding cause span $c$ in pair $p_{i}$ is annotated with its START position, its END position, and its emotion category. Span $i$ is defined by all the tokens from START($i$) to END($i$) inclusive, for 1 $\leq$ $i$ $\leq$ N.
Our goal is to find all potential span-pair of emotion and corresponding causes in a document, and make emotion classification for each pair. The overall illustration of the proposed ETC model is shown in Figure \[model\]. The basis of our proposed ETC model is the BERT encoder [@devlin2019bert], we map word embeddings into contextualized token representations using pre-trained Transformer blocks [@vaswani2017attention]. A span classifier is first used to propose multiple candidate targets from the sentence. Then, an emotion classifier is designed to predict the emotion labels towards each extracted candidate span-pair using its summarized span representation and and localized context. We further study the performance of different span search schemes.
Span Representation
-------------------
As mentioned before, we first obtain the features of tokens with BERT, which utilizes the abundant language knowledge, position information, and contextual information it contains. Given a document $D={\{x_{t}\}}$ where $t$ is the number of words, BERT begins by converting then sequence of tokens into a sequence of vectors $\mathbf{X}^{0}={\{x_{t}^{0}\}}^{L}_{i}$, $x^{0}_{t} \in \mathcal{R}^{d}$. Each of these vectors is the sum of a token embedding, a positional embedding that represents the position of the token in the sequence, and a segment embedding that represents whether the token is in the source text or the auxiliary text. We only have source text so the segment embeddings are the same for all tokens. Then several Transformer [@vaswani2017attention] layers are applied to get the final representations: $$\mathbf{X}^{i+1} = Transformer(\mathbf{X}^{i}), i \in [0, D-1]
\tag{1}$$ We use the final hidden output of BERT $\mathbf{X}^{D} \in \mathcal{R}^{L\times d}$ as the representations of corresponding tokens.
Attention mechanism [@bahdanau2015neural] can quickly extract important features of sparse data, so it is widely used in natural language. However, the BERT encoder uses a lot of attention mechanism, in order to save resources, it is no longer used. We use the following two convenient functions to create task-specific span features: (1) **sum** of all vectors for the entire span can usually represent the its semantics. (2) **max pooling** is a sample-based discretization process, which the objective is to down-sample an input representation (image, text, hidden-layer output matrix, etc.). For each span $i$, its span representation $\mathbf{g}_{i}$ was defined as: $$\mathbf{g}_{i} = concat(x^{D}_{cls}, sum(i), max(i), \Phi_{i})
\tag{2}$$ where $x^{D}_{cls}$ represents the final hidden output of BERT global context information, which is usually represented by the vector of the first token in BERT. $\Phi_{i}$ encodes the length of span $i$ in number of tokens. Each component of $\mathbf{g}_{i}$ is a span-specific feature that would be difficult to define and use in token-level models.
Jointly Extract Emotion and Cause
---------------------------------
After obtaining span representation, we predict the type for each span. This prediction is done identically and parallelly for each span. For each span we compute a vector of type scores and apply the softmax function to its type score vector to obtain the distribution. For span $i$, $$y_{i}^{span} = softmax(\mathbf{g}_{i}w_{i} + b_{i})
\tag{3}
\label{eq3}$$ where $w_{i}$ and $b_{i}$ are parameters that can be learned.
The predicted type for each span $i$ is the type corresponding to span $i$’s highest span type score. Only spans whose predicted type is not none are selected.
Emotion-Cause Classification
----------------------------
Finally, we obtain a set of emotion expression spans $E = \{\cdots, e_{i}, \cdots\}$ and a set of cause spans $C = \{\cdots,c_{i}, \cdots\}$. Now our goal is then to pair the two sets and construct a set of emotion-cause span-pairs with emotion relationship. Firstly, we apply a Cartesian product to $E$ and $C$, and obtain the set of all possible span-pairs: $$P_{all} = \{\cdots, (e_{i}, c_{j}), \cdots\}
\tag{4}$$
Despite advances in detecting long distance relations using BERT or the attention mechanism, the noise induced with increasing context remains a challenge. By using a Localized Context (LC), i.e. the context between span candidates, the emotion classifier can focus on the sentence’s section that is often most discriminative for the emotion type: $$\begin{aligned}
&p_{(e_{i}, c_{j})} = concat(\mathbf{g}_{e_{i}}, \mathbf{g}_{c_{j}}, LC_{(e_{i}, c_{j})})\\
LC_{(e_{i}, c_{j})} &= concat(sum(i\to j), max(i\to j), \Psi_{i\to j})
\end{aligned}
\tag{5}
\label{eq5}$$ where $e_{i}$ and $c_{j}$ are the representations of the emotion expression span and corresponding cause span respectively, $i\to j$ is localized context between $i$ and $j$, and $\Psi_{i\to j}$ represents the distances (dist) between span $i$ and span $j$.
For each emotion-cause span-pair $(e_{i}, c_{j})$, we obtain a representation by concatenating the respective span embeddings and Localized Context features. Finally, we train a softmax classifier to identify emotion categories: $$y_{(e_{i}, c_{j})}^{pair} = softmax(p_{(e_{i}, c_{j})}w_{(e_{i}, c_{j})} + b_{(e_{i}, c_{j})})
\tag{6}
\label{eq6}$$ where $w_{(e_{i}, c_{j})}$ and $b_{(e_{i}, c_{j})}$ are parameters that can be learned.
Loss Function
-------------
Two learning signals are provided to train this model: the span type information for each span (emotional expression, reason, and none) and the emotion category information for each selected (ordered) span-pair. Both are provided via cross-entropy [@shore1980axiomatic] loss on Eq. \[eq3\] and Eq. \[eq6\] respectively.
Experiments
===========
Corpus
------
We evaluate on the benchmark ECE corpus[^1] [@gui2016event], which was the mostly used corpus for emotion cause extraction. The corpus includes annotations of emotional expressions and corresponding emotional causes. We use the boundary of the annotations as the start and end of the spans. Note that the presence of emotion expression does not necessarily convey emotional information due to different possible causes such as negative polarity, sense ambiguity or rhetoric. And, the presence of emotion expression does not necessarily guarantee the existence of emotional cause neither. Therefore, for each emotion expression, we also use the emotion labels provided by the corpus. There are different lengths for each emotion expression and cause, and the number is shown in Table \[corpus\].
Item Instance Clauses Cause Cause\_1 Cause\_2 Cause\_3
-------- ------------- ----------------- ----------------- ------------------ ------------------ ------------------
Number 2105 11799 2167 2046 56 3
Item Annotations Length $\leq$ 2 Length $\leq$ 5 Length $\leq$ 10 Length $\leq$ 15 Length $\leq$ 20
Number 3879 1841 2587 3193 3655 3812
: Details of the corpus. Cause\_1, Cause\_2 and Cause\_3 represent the documents with 1, 2 and 3 causes, respectively. Length represent the length of each annotation.[]{data-label="corpus"}
Metrics
-------
The precision (P), recall (R), and F1 score are used as the metrics for evaluation. These metrics in emotion cause extraction are defined by: $$Precision = \frac{\sum correct\_items}{\sum proposed\_items},
Recall = \frac{\sum correct\_items}{\sum annotated\_items},
F1 = \frac{2 * Precision * Recall}{Precision + Recall}$$ where $proposed\_items$ denotes the number of items that are predicted, $annotated\_items$ denotes the number of items that in corpus and the $correct\_items$ means the number of items that are correctly predicted. Unlike previous research on clause, a correct item is considered to be correct only if both the start and end of the item are correctly predicted in the new ECSP task.
Experimental Settings
---------------------
We use the BERT-Chinese[^2] model as the default backbone network, which using 12 layers, 768 -dimensional embeddings, 12 heads per layer, resulting in a total 110M parameters. Each span gets a span length feature $\Phi$ which is a learned 25 -dimensional vector representing the number of tokens in that span and each pair also gets a localized context length feature $\Psi$ which is twice as much as $\Phi$. we randomly divide the data with the proportion of 9:1, with 9 folds as training data and remaining 1 fold as testing data. The following results are reported in terms of an average of 10-fold cross-validation. We use Adam optimizer with a linear warmup and linear decay learning rate schedule and a peak learning rate of 5e-5. Dropout is applied with dropout rate 0.1 to all hidden layers of BERT and Classifiers. Mini-batch Size is 1 and early stopping of 20 evaluations on the dev set is used.
Evaluation on the New ECSP Task
-------------------------------
### Overall Performance
Table \[result:ecsp\] shows our proposed ETC model performances with different span lengths on four sub-tasks: (emotion expression span extraction (EESE), emotion cause span extraction (ECSE), emotion-cause span-pair extraction (ECSPE), and emotion-cause span-pair extraction and classification (ECSP)).
Given a document with a $T$ token, there may be $N=T(T+1)/2$ spans. The huge search space makes the task extremely challenging. In this experiment, we created a length-restricted span (rather than just token) representation that achieves a dual goal: to improve memory efficiency and capture the majority (more than 98% of emotions, see Table \[corpus\]) for the span considered.
Compared with ETC-5 and ETC-15, ETC-20 gets great improvements on the ECSP task as well as the two sub-tasks. Specifically, we find that the improvements are mainly in the recall rate on the ECSE task, which finally lead to the great improvement in the recall rate of ECSP. The performance of the model does not decrease sharply as the length of the annotation increases, and our chosen span search scheme is far more memory efficient than a naive search over all possible spans in the input document. Yet our scheme still considers more than 98% of all annotation. Our scheme is linear in the document length, not quadratic; because we limit our proposed ETC model to spans that are wholly in a document and have a max length of $L$ = 20 tokens.
In addition, the model achieved excellent F1 score 88.71 on ESE, but the F1 score on ECSP is 3.14% lower than ECSPE, which indicates that it is not enough to extract the emotion without identifying the emotion categories. The presence of emotion clause does not necessarily convey emotional information explicitly, and emotions need to be classified.
-------- ------- ------- ----------- ------- ------- ----------- ------- ------- ----------- ------- ------- -----------
P R F1 P R F1 P R F1 P R F1
ETC-5 87.78 89.69 **88.71** 63.73 10.48 17.93 56.19 09.31 15.90 54.14 08.90 15.22
ETC-10 87.81 89.53 88.65 61.05 40.69 48.81 54.18 37.23 44.10 50.78 35.14 41.50
ETC-15 86.78 90.10 88.39 61.91 51.78 56.22 53.66 47.35 50.20 50.60 44.40 47.14
ETC-20 87.56 89.31 88.42 60.23 57.78 **58.90** 53.19 51.11 **52.11** 49.91 48.11 **48.97**
-------- ------- ------- ----------- ------- ------- ----------- ------- ------- ----------- ------- ------- -----------
: Experimental results of all proposed ETC model and variants, where ETC-$n$ represents the maximum span length of the model is $n$.[]{data-label="result:ecsp"}
### Effect of Localized Context
As is shown in Table \[result:LC\], localized context can effective slightly improve the performance of the model. The localized context takes advantage of all information between two span, so it is able to enrich the source information when the model predicts the emotion labels, which leads to the performance of the model effective significantly improved.
--------- ------- ------- ----------- ------- ------- ----------- ------- -------- ----------- ------- ------- -----------
P R F1 P R F1 P R F1 P R F1
Without 87.08 89.11 88.07 58.76 57.10 57.79 52.73 520.03 51.31 49.62 47.24 48.35
With 87.56 89.31 **88.42** 60.23 57.78 **58.90** 53.19 51.11 **52.11** 49.91 48.11 **48.97**
--------- ------- ------- ----------- ------- ------- ----------- ------- -------- ----------- ------- ------- -----------
: Experimental results with localized context effects. “With” represents the localized context, “Without” means no localized context.[]{data-label="result:LC"}
Evaluation on the Traditional Task
----------------------------------
By relaxing the ECSP task to the clause-level, we further examine our model by comparing it with state-of-the-art of the traditional ECE and ECPE task.
### Baselines
We employ a hierarchical Bi-LSTM network **Indep** proposed by as baseline in ECPE task. The lower layer consists of a set of word-level Bi-LSTM modules, each of which corresponds to one clause, and accumulate the context information for each word of the clause. Attention mechanism is then adopt to get a clause representation. The upper layer consists of two components: one for emotion expression extraction and another for cause extraction. Each component is a clause-level Bi-LSTM which receives the independent clause representations and finally feed to the softmax layer for emotion prediction and cause predication. It has two interactive variants: **Inter-CE**, where the predictions of cause extraction are used to improve emotion extraction, and **Inter-EC**, where the predictions of emotion extraction are used to enhance cause extraction.
In addition to baselines mentioned above, we also considered several state-of-the-art methods and models in ECE task that need to provide annotations of emotional expressions in the test set in advance to evaluate the results of our proposed ETC model: **RB** is a rule based method [@lee2010text]; **CB** is common-sense based method [@russo2011emocause]; **ConvMS-Memnet** considers emotion cause analysis as a reading comprehension task and designs a multiple-slot deep memory network to model context information [@gui2017question]. **CANN** uses a co-attention neural network to identify emotion causes [@li2018co] and **CANN-E** eliminates the dependence of CANN on emotion annotation in the test data. **HCS** is proposed by using a multiple-level hierarchical network to detect the emotion causes. **MANN** is the current state-of-the-art method employing a multi-attention-based model for emotion cause extraction [@li2019context].
### Results and Analysis
\
The past clause-level models regarded the ECE task as a set of independent clause classification problems. By observing the Table \[result:clause\] (c), we found that the proportions of emotion cause clauses and non-emotion-cause clauses were 18.36% and 81.64%, respectively. It is a serious class-imbalance classification problem and the model tends to predict the clause as non-emotion-cause more often. This is also the reason why their Recall scores were quite low (the highest was 75.87).
By contrast, it can found in Table \[result:clause\] (c) that our proposed ETC model is absolutely higher on each indicator than the other baselines and no need to manually annotate the test set. This is because they can capture the relations of multiple clauses which help inferring the current clause. For example, if no other clauses in a document have been detected as an emotion cause, the model will increase the probability of the current clause being predicted as an emotion cause. This finally increases the Recall score. It is clear that by removing the emotion annotations (CANN-E), the F1 score of CANN drops dramatically (about 34.69%). In contrast, our method does not need the emotion annotations and achieve 89.57% in F1 score, which significantly outperforms the CANN-E model by 51.6%.
guessed that the expression clause extraction and cause clause extraction are not mutually independent. On the one hand, providing emotions can help better discover the causes; on the other hand, knowing causes may also help more accurately extract emotions. Our proposed ETC model uses a classifier to complete the classification of expression and cause, forcing the classifier to learn the intrinsic relationship between them. Thanks to BERT’s self-attention mechanism, our proposed ETC model can capture the relationship between multiple clauses. It can found in Table \[result:clause\] (b) that our proposed ETC model has been greatly improved on both expression clause extraction and cause clause extraction tasks. Compared with Indep, Inter-CE and Inter-EC, our proposed ETC model gets great improvements on the ECPE task as well as the two sub-tasks. Our span-based model achieves 11.57%, 24.77% and 24.5% absolute gains on three sub-task compared to the best classification model, indicating the efficacy of our proposed ETC model.
Related Work
============
First of all, our work is related to extracting causes based on emotions expression presented in documents, i.e., emotion cause extraction (ECE). ECE was first proposed by , given the fact that an emotion is often triggered by cause events and that cause events are integral parts of emotion, they proposed a linguistic-driven rule-based system for emotion cause detection. To solve the insufficient of no formal definition about event in emotion cause extraction and there was no open corpus available for emotion cause extraction, released a corpus and re-formalized the ECE task as a clause classification problem. This corpus has received much attention in the following study and has become a benchmark corpus for ECE task research. Based on this corpus, several traditional rule-based models [@lee2010text; @russo2011emocause; @gui2014emotion], machine learning models [@gui2016event; @gui2017question; @xu2017ensemble] and deep learning models [@gui2017question; @li2018co; @yu2019multiple; @xia2019rthn; @li2019context] were proposed. Recently, To solve the shortcoming of emotion expression must be annotated before cause extraction in the test set, proposed emotion-cause pair extraction (ECPE) task, which aims to extract all potential clause-pairs of emotion expression and corresponding cause in a document.
Conclusions and Future Work
===========================
The key idea of task and model is to build span-based feature representation for emotion expression and causes to efficiently extract document information. Furthermore, our proposed ETC model is able to utilize the information based on an overall understanding of the document and a better localized context of interactions between spans. Comprehensive empirical studies demonstrate the effectiveness of our proposed ETC model. Since our proposed ETC model has a single input structure, so in the future we will explore how to incorporate discourse graphs into our proposed ETC model to further improve performance, and we intend to annotate a large-scale emotion-cause span-pair corpus to facilitate research.
[^1]: Available at: <http://www.hitsz-hlt.com/?page_id=694>
[^2]: Available at: <https://github.com/huggingface/transformers>
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'The fabrication of three-dimensional (3D) nanostructures is of great interest to many areas of nanotechnology currently challenged by fundamental limitations of conventional lithography. One of the most promising direct-write methods for 3D nanofabrication is focused electron beam-induced deposition (FEBID), owing to its high spatial resolution and versatility. Here we extend FEBID to the growth of complex-shaped 3D nanostructures by combining the layer-by-layer approach of conventional macroscopic 3D printers and the proximity effect correction of electron beam lithography. This framework is based on the continuum FEBID model and is capable of adjusting for a wide range of effects present during deposition, including beam-induced heating, defocussing and gas flux anisotropies. We demonstrate the capabilities of our platform by fabricating free-standing nanowires, surfaces with varying curvatures and topologies, and general 3D objects, directly from standard stereolithography (STL) files and using different precursors. Real 3D nanoprinting as demonstrated here opens up exciting avenues for the study and exploitation of 3D nanoscale phenomena.'
author:
- 'L. Skoric'
- 'D. Sanz-Hernández'
- 'F. Meng'
- 'C. Donnelly'
- 'S. Merino-Aceituno'
- 'A. Fernández-Pacheco'
bibliography:
- '3dprinting.bib'
title: 'Layer-by-layer growth of complex-shaped three-dimensional nanostructures with focused electron beams'
---
Introduction
============
The realization of three-dimensional (3D) nanoscale systems is expected to play a key role for the future progress in many areas of nanotechnology such as biology[@qianMicroNanofabricationTechnologies2010], nanomagnetism[@fernandez-pachecoThreedimensionalNanomagnetism2017], metamaterials[@kadic3DMetamaterials2019], and micro-electro-mechanical systems (MEMS)[@vavassoriRemoteMagnetomechanicalNanoactuation2016]. However, conventional lithography methods that excel in planar patterning are not well suited for complex 3D nanofabrication. The development of advanced direct-write micro- and nano-scale fabrication techniques is a possible solution to this problem currently under intense investigation. These include two-photon lithography, [@maruoThreedimensionalMicrofabricationTwophotonabsorbed1997; @seniutinas100NmResolution2018] direct ink writing, [@lewisDirectInkWriting2006] laser assisted methods, [@lewis3DNanoprintingLaserassisted2017; @takaiThreedimensionalMicrofabricationUsing2014] local dispensing of ions in liquid, [@hirtTemplateFree3DMicroprinting2016; @momotenkoWriteRead3D2016] and focused electron and ion beam-induced deposition (FEBID and FIBID). [@winkler3DNanoprintingFocused2019; @matsuiFocusedIonBeamChemicalVaporDepositionFIBCVD2012]
The rapid prototyping of functional geometries with sub-micrometer features is particularly important for applications in areas such as magnetism, superconductivity and metamaterials which are ruled by characteristic length scales typically in the ten to hundred nanometer range. [@fernandez-pachecoThreedimensionalNanomagnetism2017; @espositoNanoscale3DChiral2015; @bezryadinQuantumSuppressionSuperconductivity2000; @belkinFormationQuantumPhase2015; @zhuScalableMultiphotonCoincidence2018]. When it comes to fabricating these geometries, FEBID has demonstrated a number of advantages[@hirtAdditiveManufacturingMetal2017; @jesseDirectingMatterAtomicScale2016], including resolution of a few tens of nanometers, and vertical growth rates in hundreds of nanometers per second [@winklerHighFidelity3DNanoprintingFocused2018]. A large number of available precursors allows for the deposition of metallic, organic, semiconducting, magnetic and superconducting materials. [@utkeGasassistedFocusedElectron2008; @huthFocusedElectronBeam2018]. Although traditionally FEBID-fabricated materials contain a significant proportion of organic impurities, deposition of highly-pure materials has been demonstrated via different strategies, including synthesis of new precursors[@porratiDirectWritingCoFe2015; @makiseMicrostructuralAnalysisTransport2014; @winholdSuperconductivityMetallicBehavior2014], optimization of growth conditions, [@pablo-navarroTuningShapeComposition2017; @bernauTunableNanosynthesisComposite2010; @keLowdosePatterningPlatinum2013] introduction of reactive gases during growth,[@shawravHighlyConductivePure2016] and post-deposition purification. [@serrano-ramonImprovementDomainWall2013; @botmanPurificationPlatinumGold2006; @botmanCreatingPureNanostructures2009].
These advances have already led to a number of important fundamental studies in plasmonics[@winklerDirectWrite3DNanoprinting2017; @espositoNanoscale3DChiral2015], photonic crystals [@koopsTwodimensionalPhotonicCrystals2001], and magnetic nanowires and lattices[@sanz-hernandezFabricationDetectionOperation2017; @sanz-hernandezFabricationScaffoldBased3D2018; @fernandez-pachecoMagnetotransportPropertiesHighquality2009; @wartelleTransmissionXMCDPEEMImaging2017; @kellerDirectwriteFreeformBuilding2018]. Beyond nano-prototyping, a substantial increase in throughput for this technique could be accomplished by orders-of-magnitude increases in deposition rates through optimizations of gas injection systems (GIS) [@friedliOptimizedMoleculeSupply2009] and deposition at cryogenic temperatures [@bresinDirectwrite3DNanolithography2013], and even parallelizing via multiple beams in next-generation tools[@riedeselFirstDemonstration331beam2019].
The application of FEBID for 3D nanofabrication has progressed significantly in recent years by evolving from a trial-and-error approach to systematic generation of electron beam instructions via CAD software solutions [@fowlkesHighFidelity3DNanoprintingFocused2018; @kellerPatternGenerationDirectwrite2018]. These works make an important step towards controllable nanoscale 3D printing of networks of nanowires, correcting effectively for various phenomena present in FEBID when growing pseudo-1D elements using different approaches[@winkler3DNanoprintingFocused2019]. To exploit the full potential of FEBID as a 3D fabrication tool, however, the nanoprinting protocol needs to be redefined in order to generalize it to more complex geometries.
In this article, we have developed a general layer-by-layer framework for the additive manufacturing of 3D nanoscale objects by focused electron beams based on the FEBID continuum model, offering a new level of control over the fabrication of three-dimensional nanosystems. We present an algorithm capable of creating beam scanning patterns adjusted for proximity effects directly from conventional 3D printing stereolithography (STL) file formats. We further show that the model can take account of a range of phenomena, including beam-induced heating, defocussing and gas flux anisotropy. We demonstrate the effectiveness of the model by fabricating geometrically complex 3D structures with high fidelity and with different precursors. This work represents a significant step forward in the capabilities for 3D fabrication at the nanoscale, as well as in the simplification of the nanostructure-design process.
Theoretical description of growth model
=======================================
FEBID deposition is inherently a stochastic process where each electron beam dwell induces the deposition of a small number of precursor molecules. However, when averaged over the characteristic deposition times (usually in the order of a few seconds to tens of minutes), a good approximation can be achieved by describing the height $h$ of the evolving deposit as a continuous function of dwell time $t$ and radial distance $r$ from the beam center $h(t, r)$[@tothContinuumModelsFocused2015]. Overlapping a large number of such deposits produces out of plane growth.
We develop a framework where, similar to conventional 3D printing, structures are grown via the deposition of thin slices/layers. In each layer, a series of deposits $i$ is made at beam scanning positions ${\mathbf{r_i}}$ with dwell times $t_i$, creating deposits $h_i(t_i, {\left\lVert{\mathbf{r}}-{\mathbf{r_i}}\right\rVert})$ (see Figure \[fig:growth\]a).
Provided that the lateral separation between layers is low and that the individual deposits are small and closely spaced, the total height increment at any ${\mathbf{r_i}}$ is the sum over all neighboring deposits. For consistent growth, this needs to be equal to the height of a layer $l$, $\Delta z_l$, resulting in a set of equations for each layer depending on the inter-point distances and dwell times: $$\sum_{j\in N_i} h_j(t_j, {\left\lVert{\mathbf{r_j}} - {\mathbf{r_i}}\right\rVert}) = \Delta z_l,
\label{eqn:consistency}$$ where $N_i$ is the set of all deposits of a layer $l$ in the proximity of ${\mathbf{r_i}}$ (referred to as neighboring points), and $t_j$ is the dwell time associated to each point of the layer.
In general, these equations form a non-linear optimization problem which is computationally expensive to solve. However, under the assumption that the deposit evolution can be well approximated by a separable function of time and space ${h(t, r) = s(r)f(t)}$, we can reduce Equation \[eqn:consistency\] to a matrix equation that is numerically efficiently solvable: $$\Delta z_l = s_{ij}f(t_i),
\label{eqn:matrixdz}$$ where $s_{ij} = s({\left\lVert{\mathbf{r_j}}-{\mathbf{r_i}}\right\rVert})$ is a symmetric matrix, here referred to as the intra-layer proximity matrix. The dwell time values needed for consistent growth can then be calculated numerically by solving this matrix equation and inverting $f$. This framework allows us to adjust for inter-deposit proximity effects within a layer of a 3D object, similarly to “self-consistent” dose correction algorithms employed in electron beam lithography [@owenMethodsProximityEffect1990; @parikhCorrectionsProximityEffects1979]. Here we treat each slice of a 3D object separately, reducing the hard problem of a general 3D growth to a set of 2D consistency equations, significantly reducing the calculation complexity.
To use the above framework for the realistic modeling of a FEBID deposition, an accurate expression for $h(t, r)$ is needed. In the following, we determine this function for the Langmuir FEBID model under equilibrium conditions. [@tothContinuumModelsFocused2015] Using the compact notation based on the characteristic FEBID frequencies that we previously developed [@sanz-hernandezModellingFocusedElectron2017], the dynamics of fractional surface molecule coverage under no diffusion $\theta$ can be described as a competition between the characteristic frequencies of gas adsorption ($\nu_{GAS}$), thermal desorption ($\nu_{d}$) and electron-induced precursor dissociation ($\nu_{el}$): $$\label{eqn:thetadiff}
\frac{\partial \theta}{\partial t} = \nu_{GAS}(1-\theta) - \nu_d\theta - \nu_{el} \theta.$$ The corresponding growth rate ($GR$) is proportional to the amount of gas present and the electron dissociation frequency: $$\label{eqn:growthrate}
GR = \frac{\partial h(t, r)}{\partial t} \propto \theta\nu_{el}.$$ Equations \[eqn:thetadiff\] and \[eqn:growthrate\] can be solved analytically, describing the depletion of gas coverage under electron beam dissociation, and subsequent growth of the induced deposit (see Supporting Information S1.1)[@tothContinuumModelsFocused2015].
To specify the the spatial dependence of deposits, the profile of the effective flux of electrons inducing decomposition as a function of the radial distance from the electron beam $\nu_{el} = \nu_{el}(r)$ is required. For focused beams, this can be approximated by a Gaussian function[@utkeGasassistedFocusedElectron2008]: $$\nu_{el} = \nu_{el}^0 \exp\bigg(\frac{-r^2}{2\sigma^2}\bigg).$$ where $\sigma$ is the effective standard deviation, and $\nu_{el}^0$ the dissociation frequency in the center of the beam.
In general, the surface coverage will evolve during the beam dwell time, converging to an equilibrium distribution on the gas dynamics time scale (see Supporting Information S1.1): $$\tau_{dyn} = \frac{1}{\nu_{GAS}+\nu_d}.$$
However, when growing either under electron limited conditions where the precursor depletion is small ($\nu_{el} \ll \nu_{GAS}+\nu_d$), or in the case when the characteristic dwell time (usually on the order of 1ms) is much longer than the time-scale for convergence to equilibrium ($t_{dwell} \gg \tau_{dyn}$), we can assume that the system is at equilibrium at all times during the growth. This simplifies the expression for $h(t, r)$ to: $$h \propto \underbrace{\frac{\nu_{GAS}\nu_{el}(r)}{\nu_{GAS} + \nu_d + \nu_{el}(r)}}_{s(r)}t.
\label{eqn:simplecontinuum}$$ In this form, $h$ is written as a separable function with no further approximations. We follow this approach in our experiments as discussed below.
To allow for an analytical solution, in the above analysis we did not consider the effect of diffusion, which may become important in FEBID under mass-transport limited conditions. [@tothContinuumModelsFocused2015] However, in the context of Langmuir FEBID, the addition of diffusion would only alter the relevant timescales by speeding up the gas dynamics, as shown in Supporting Information S2. The separable form of deposit evolution could also be recovered in this case, albeit with a more complex form for the spatial part s(r).
Experimental determination of growth parameters
===============================================
For the effective implementation of the model described above, a good approximation for $s(r)$ needs to be determined. In the following, we focus on the standard Pt-based precursor. A generalization to other precursors is discussed later.
To apply our model experimentally, key deposition parameters need to be determined for the given experimental conditions. Although in recent years a significant effort has been made towards determining these both through experiment [@vandorpElectronInducedDissociation2009; @wnukElectronInducedSurface2009; @yangProbingMorphologyEvolving2017; @fowlkesFundamentalElectronPrecursorSolidInteractions2010; @utkeGasassistedFocusedElectron2008; @winklerHighFidelity3DNanoprintingFocused2018] and simulations [@smithNanoscaleThreedimensionalMonte2007; @fowlkesSimulationGuided3DNanomanufacturing2016; @mutungaImpactElectronBeamHeating2019; @sushkoMolecularDynamicsIrradiation2016; @shenPhysisorptionOrganometallicPlatinum2012], the size of the parameter space still prevents confident ab-initio growth predictions. Here we demonstrate a set of experiments (Figures \[fig:growth\]b and \[fig:growth\]c) designed to find the minimal set of parameters which represent our growth conditions, enabling a simple calibration and preventing overfitting.
We measure the growth parameters by building a set of standardized structures, isolating the dependence of the growth rate to particular factors. Firstly, to determine growth rate at the center of the beam, a set of spot depositions is built in parallel with different total deposition times (Figure \[fig:growth\]b). The resulting lengths of the vertical nanowires are measured and plotted against the deposition times.
Under the experimental conditions used in this paper (see Experimental), we have observed no more than 7% variation in the average growth rate when varying the dwell and refresh times used in 3D FEBID deposition (see Supporting Information S3). Therefore, we can safely assume that Equation \[eqn:simplecontinuum\] provides a good approximation of the deposition.
The sublinear dependence of nanowire length on the deposition times can thus be understood via the enhancement of thermal desorption due to beam-induced heating $\nu_d$ [@mutungaImpactElectronBeamHeating2019]. The absolute changes in local temperature during deposition are not expected to be large compared to room temperature, being usually in the order of 10K [@mutungaImpactElectronBeamHeating2019; @bouscaudEstimationElectronBeaminduced2012]. We can thus expand the desorption temperature dependence to first order: $$\nu_d = \nu_0 \exp\bigg(-\frac{E_d}{k_B(T_0+T_{el})}\bigg) \approx \nu_d^0 e^{\beta T_{el}},
\label{eqn:des_approx}$$ where $\nu_0$ is the desorption attempt frequency, $T_0$ is the base deposition temperature, $T_{el} \ll T_0$ is the change in temperature due to electron beam heating, $\nu_d^0$ is the desorption at $T_0$, and $\beta = E_d/(k_BT_0^2)$.
The significant effect of local temperature, together with low equilibrium precursor coverage in Pt FEBID [@friedliMassSensorSitu2007], implies a strong dependence of the growth rates on desorption. Therefore, even though in general we allow for the full form of the equilibrium growth rate (Equation \[eqn:simplecontinuum\]) in our model, we have found that the desorption dominated regime ($\nu_d \gg \nu_{GAS}+\nu_{el}$)[@sanz-hernandezModellingFocusedElectron2017] results in the best fit to our experimental conditions.
Under the assumption that the incident beam heating rate is the same at all dwell points and that the characteristic thermal transport time scale is expected to be short relative to the dwell times[@mutungaImpactElectronBeamHeating2019], a quasi-steady-state temperature can be assumed for every dwell. Hence, the temperature variation during deposition is fully captured by a geometrical factor $R_T$, giving account for the resistance of the 3D structure to heat transport at the point of scanning ($T_{el}\propto R_T$)[@mutungaImpactElectronBeamHeating2019].
Considering the effect of beam heating in the desorption-dominated regime, Equation \[eqn:simplecontinuum\] gives a Gaussian deposit (see inset in Figure \[fig:growth\]a): $$\label{eqn:Ptdep}
h(t, r) = GR_0 e^{-kR_T}\exp\bigg(\frac{-r^2}{2\sigma^2}\bigg)t,$$ where the growth rate in the center of the beam at a base deposition temperature $GR_0$, thermal resistance scaling factor $k$, and the standard deviation of the deposit $\sigma$ are parameters that need to be determined. The exponential temperature dependence for desorption (Equation \[eqn:des\_approx\]) is expressed here through the geometry-dependent factor $e^{-kR_T}$. For details about the implementation of the resistance model and the derivation of Equation \[eqn:Ptdep\] can be found in Supporting Information S4.2.
This experiment, involving vertical deposits, allows us to remove the potential influence of the geometry of the structure and proximity effects on the resulting growth. We extract the expected dependence of lengths $L(t)$ on deposition times $t$ based on Equation \[eqn:Ptdep\], with $w_0$ as their width, and growth rate $GR_0$ and resistance scaling factor $k$ as fitting parameters (see Supporting Information S4.3 for the full derivation): $$L(t) = \frac{w_0}{k}\log\bigg(\frac{kGR_0t}{w_0}+1\bigg).$$
A second experiment (Figure \[fig:growth\]c) is designed to obtain a value for $\sigma$. For this, a set of nominally 150nm wide straight nanowires tilted at 45[$^{\circ}$]{} is built using the algorithm, while varying the value of $\sigma$ in the model with the goal of matching the angle of a single pixel thin nanowire built with the same conditions. For the Pt precursor, we find $GR_0 \approx 100$-$300$nms$^{-1}$, and $\sigma \approx 4$-$5$nm, depending on the SEM system and the conditions inside the microscope chamber, such as temperature, base pressure, and precursor flux.
This two-step calibration procedure allows us to uniquely determine all growth parameters required for implementing an algorithm to pattern 3D complex structures. We further tested the validity of the calibration protocol by growing two sets of additional test structures: nanowires with constant out-of-plane angle and variable controllable width (Figure \[fig:testing\]a), and nominally-straight nanowires, where the $k$ value obtained by the fit in Figure \[fig:growth\]b leads to the growth of a straight structure (see “Optimal correction” in Figure \[fig:testing\]b).
Apart from temperature, we studied other potentially-relevant effects such as electron beam defocussing and gas flux anisotropy. These have been found to have a relatively small effect for our experimental conditions, but can be important for very fine growth in certain structures and deposition regimes. If required, further corrections of this type can be easily implemented in our model at the expense of introducing additional parameters that require independent calibration (see Supporting Information S5).
3D printing algorithm
=====================
We have implemented the layer-by-layer growth model explained above in a three-step algorithm designed to generate beam scanning patterns for FEBID deposition of 3D arbitrary geometries. The algorithm creates deposition sequences directly from STL files that can be designed in any standard 3D CAD software, the same approach followed by standard 3D printers. This approach, in combination with the model described above, is a significant improvement with respect to recent FEBID works [@fowlkesHighFidelity3DNanoprintingFocused2018; @kellerPatternGenerationDirectwrite2018], simplifying the design and enabling the fabrication of complex 3D nano-objects. In this section we give an outline of the algorithm and demonstrate its effectiveness and flexibility.
The workflow of the algorithm is described using the example of a free-standing concave surface (Figure \[fig:algorithm\]a-c), with an SEM image of the fabricated nanostructure given in Figure 3d. Firstly, a geometry is defined with an STL file which is sliced using constant-z planes, defining the layers of the structure (Figure \[fig:algorithm\]a). In all experiments performed here, a maximum layer height of $\Delta z_l \leq 6$nm is set, to guarantee that the approximation of thin layers remains valid. Additionally, an adaptive slicing procedure has been implemented, adjusting the slice thickness based on the local angle of a structure. In this way, low hanging features are more finely defined, and a low lateral displacement between layers is maintained. From each slice, a dense mesh of dwell points is created, with points separated by a lateral “pitch” distance. We set both the displacement between layers, and the pitch to be 3nm. This value has been determined from dedicated experiments (see Supporting Information S7), where single nanowires were found to be widely independent of point pitch, for pitch values below 4nm, in agreement with literature.[@winklerHighFidelity3DNanoprintingFocused2018]
In a second step, geometry-dependent factors affecting the growth are calculated, and included as growth parameters for each dwell point. Specifically, the previously discussed resistance-based model for beam-induced heating is implemented, as displayed in Figure \[fig:algorithm\]b. Further corrections for second-order effects such as beam defocussing and gas flux anisotropy are also possible, but only applied to objects where the added complexity is essential for successful deposition (see Supporting Information S5). We anticipate that this second step of the algorithm could be expanded further by including additional factors relevant in other conditions and geometries, such as diffusion [@mutungaImpactElectronBeamHeating2019] and the dependence of secondary electron emission with local geometry. [@fowlkesFundamentalElectronPrecursorSolidInteractions2010] These, together with the measured calibration parameters, would redefine the model for deposition, $s(r)$, at every dwell point.
Thirdly, we implement the per-layer dwell time solver. Inter-point distances are computed and, together with the growth model, result in the intra-layer proximity matrix $s_{ij}$. The optimization problem for finding the appropriate dwell times at each point (as per Equation \[eqn:matrixdz\]) is then solved using a non-negative linear-least squares solver. The colormap of Figure \[fig:algorithm\]c represents the resulting dwell times at each position.
Finally, the beam scanning pattern is generated out of the computed dwells. In each layer, the beam is set to make multiple passes, reversing its direction from layer to layer in a “serpentine” pattern[@sanz-hernandezFabricationScaffoldBased3D2018], to avoid exceeding a set maximum dwell time (5ms) and improving the smoothness of the deposition. The generated pattern is deposited, resulting in the SEM image shown in Figure \[fig:algorithm\]d.
In what follows, we demonstrate the capabilities of the algorithm by patterning a range of geometries using two different (Pt- and Co-based) precursors. We first show that the fabrication of networks of straight nanowires can be reproduced, similar to the existing FEBID algorithms,[@fowlkesHighFidelity3DNanoprintingFocused2018; @kellerPatternGenerationDirectwrite2018] by building a sequence of nanocubes of different sizes with half-pitch down to 50nm (see Figure \[fig:structures\]a-b and Supporting Information S8 for further details). Secondly, expanding recent achievements in nanowire device fabrication [@sanz-hernandezFabricationDetectionOperation2017; @sanz-hernandezFabricationScaffoldBased3D2018], in Figure \[fig:structures\]c-d we create a free-standing looped nanowire circuit formed by segments at different angles. The high quality and smoothness of the connections achieved here show the potential of our method to nano-prototype advanced 3D nanoelectronic devices. [@sanz-hernandezFabricationDetectionOperation2017].
Finally, the main advantage of our approach is the capability to fabricate arbitrary three-dimensional architectures out of 3D CAD files. As an example, we have fabricated a nanoscale human hand replica (Figure \[fig:structures\]e-g) and a Möbius strip (Figure \[fig:structures\]h-i). These example structures present a wide range of features at different scales which are accurately replicated in experiments. In particular, Figure \[fig:structures\]h-i shows the potential of this approach where, as far as we are aware, the smallest realization of a magnetic Möbius strip has been fabricated. See the video in Supporting Information for a full characterization of this structure at multiple angles, and comparison with the corresponding model file.
We note though, that the thickening of some features is observed for the built structure with respect to the model, e.g. the hand (Figure \[fig:structures\]f) and the associated STL file (Figure \[fig:structures\]e). This edge thickening effect is expected when creating a volumetric object via FEBID due to the penetration of the primary electron beam through the structure, and subsequent generation of backscattered electrons and type II secondary electrons. In order to correct for this additional precursor dissociation, a “shape-dimension adjustment technique”[@parikhCorrectionsProximityEffects1979] as the one employed in electron beam lithography could be implemented.
Another point to remark is that the model has been primarily developed using the standard Pt precursor due to its fast growth rate (of the order of 100nms$^{-1}$), low dependence on refresh times, and a well-known mechanism for decomposition under focused electron beams[@wnukElectronBeamDeposition2011]. Figure \[fig:structures\]h-i demonstrate the algorithm’s robustness when applied to other precursors. The Co-based precursor used to build this structure has a significantly more complex chemical behavior, including autocatalytic effects [@bishopFundamentalAdvancesFocused2019; @muthukumarSpontaneousDissociationCo22012]. We attribute the transferability of our algorithm to complex gases to the ability of the basic Gaussian deposit model with a constant vertical growth rate, with no temperature correction: $$h(t, r) = GR_0\exp\bigg(\frac{-r^2}{2\sigma^2}\bigg)t$$ to capture the effective average of deposition and give satisfactory first-order results for most regimes. This is especially true for wide structures (as opposed to single-pixel nanowires) because scanning along their width introduces a natural refresh rate. Moreover, beam heating effects are greatly reduced due to the higher heat conductivity provided by a larger volume, leading to more consistent growth rates (see Supporting Information S4.2).
Conclusion
==========
We present a framework that makes possible high-fidelity layer-by-layer growth of complex-shaped 3D nanostructures using focused electron beams. Based on a layer-by-layer growth implemented in combination with the FEBID continuum model and proximity effect corrections, we can effectively account for a variety of effects, including beam-induced heating, defocussing, and gas anisotropy. These effects have been studied for the Pt precursor, where we demonstrate how a large number of fundamental parameters can be reduced to only three, which are capable of effectively modeling the 3D deposition process.
The framework has been implemented computationally to generate beam scanning patterns from STL files created by any standard 3D CAD software. Using this approach, a wide range of nanogeometries which were until now inaccessible has been fabricated. These include surfaces with curvature along arbitrary directions, a Möbius strip and a replica of a human hand. We have successfully tested our platform in three SEM systems and using two precursors with different physical and chemical properties, thus demonstrating the robustness and applicability of the technique.
This work paves the way for the advanced nano-manufacturing of 3D objects in a wide range of nanotechnology areas, making a fundamental step towards the study of advanced effects and their future exploitation.
Experimental
============
Three different dual-beam microscope systems were used for FEBID experiments: Helios 600 at the Wolfson Electron Microscopy Suite of University of Cambridge, FEI Nova 200, and Helios 660 NanoLab at the Kelvin Nanocharacterisation Centre of University of Glasgow. Helios 600 and Nova 200 were used for deposition of , while Helios 660 was used with precursor.
All depositions were performed on p-doped Si substrates under 21pA, 30kV electron beams. The algorithm for computing dwell points and corresponding dwell times was implemented in MATLAB using the built-in Optimization Toolbox. Computation times for all structures were in the order of $\sim10$s on most modern desktop computers, with the exception of the hand (Figure \[fig:structures\]e-g) which took several minutes. The algorithm was in some cases supplemented with stage tilting to build features at various angles. In particular, this was exploited for the structures on Figure \[fig:structures\]a-b(c-d) by building at 45[$^{\circ}$]{}(30[$^{\circ}$]{}) tilt in order to realize nanowires parallel to the substrate. The fabrication times ranged from 2s for the shortest pillar in Figure \[fig:growth\]b, 1min for the largest cube (Figure \[fig:structures\]a-b), and up to 65min and 75min for the Möbius and hand respectively (Figure \[fig:structures\]e-g). For more information about the access to the software used in this paper, please contact the corresponding authors.
This work was supported by the EPSRC Cambridge NanoDTC EP/L015978/1, an EPSRC Early Career Fellowship EP/M008517/1, and the Winton Program for the Physics of Suistanability. L. Skoric acknowledges support from St Johns College of the University of Cambridge. C.Donnelly was supported by the Leverhulme Trust (ECF-2018-016), the Isaac Newton Trust (18-08) and the L’Oréal-UNESCO UK and Ireland Fellowship For Women In Science. Sara Merino-Aceituno was supported by the Vienna Research Groups grant number VRG17-014 by the Vienna Science and Technology Fund. Dédalo Sanz-Hernández acknowledges a Girton College Pfeiffer Scholarship.
The authors thank the staff of the Wolfson Electron Microscopy Suite of University of Cambridge and the Kelvin Nanocharacterisation Centre of University of Glasgow for technical support. We thank Jason D. Fowlkes, A. Hierro-Rodriguez and A. Welbourne for helpful discussions.
- Experimental and technical details (PDF)
- Video showing the Möbius strip from different angles compared to the corresponding 3D model (AVI)
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
author:
- |
V. E. Korepin\
[*C.N. Yang Institute for Theoretical Physics*]{}\
[*State University of New York at Stony Brook*]{}\
[*Stony Brook, NY 11794–3840, USA*]{}\
\
S. Lukyanov\
[*Department of Physics and Astronomy, Rutgers University*]{}\
[*Piscataway, NJ 08855-0849, USA*]{}\
\
Y. Nishiyama\
[*Department of Physics, Faculty of Science, Okayama University*]{}\
[*Okayama 700-8530, Japan*]{}\
\
M. Shiroishi\
[*Institute for Solid State Physics, University of Tokyo*]{}\
[*Kashiwanoha 5-1-5, Kashiwa, Chiba, 277-8571, Japan*]{}
title: ' **Asymptotic Behavior of the Emptiness Formation Probability in the Critical Phase of XXZ Spin Chain** '
---
We consider interacting spin 1/2 on the one-dimensional infinite lattice. The model is described by the Hamiltonian $${\cal H}=\sum_{j=-\infty}^{\infty}
\left\{ S^x_j S^x_{j+1} + S^y_j S^y_{j+1}
+ \Delta \left(S^z_j S^z_{j+1} -\frac{1}{4} \right) \right\}, \label{XXZ}$$ where $S= \sigma /2$ and $\sigma$ are Pauli matrices. It exhibits diverse physics with varying the anisotropy parameter $\Delta$. For $\Delta \le -1$ the ground state has a long-range ferromagnetic order, and a finite excitation gap opens in the low-lying spectrum. For $\Delta > 1$ the ground state develops antiferromagnetic order, the gap also exists. For moderate anisotropy $-1 < \Delta \le 1$, the antiferromagnetic order is dissolved by quantum fluctuations. The gap closes, and the long-range asymptotic of correlation function $\langle S_i S_j \rangle$ decays as a power law at zero temperature.
In this article we study the Emptiness Formation Probability (EFP), $$P(n) = \Big\langle \prod_{j=1}^n \left(S_j^z + \frac{1}{2} \right)
\Big\rangle,
\label{EFP}$$ in the domain of critical phase. The importance of EFP was emphasized in the book [@BIK93]. It reflects the nature of the ground-state and it is a good indicator of the phase-separation.
In spite of its seemingly complicated definition, EFP has emerged quite naturally in the course of studies of the quantum integrability of the $XXZ$ spin chain. The first progress in calculating of EFP was achieved in Ref. [@Takahashi77], where $P(3)$ was obtained for the isotropic antiferromagnetic chain $(\Delta=1)$. Subsequently, Jimbo and Miwa derived multiple integral representation for all correlation functions by means of vertex operators approach [@Jimbo95; @Jimbo96]. Their integral formula for EFP is: $$\begin{aligned}
P(n) &=& \left(- \nu \right)^{-\frac{n(n+1)}{2}} \int_{-\infty}^{\infty}
\frac{{\rm d} \lambda_1}{2 \pi} \cdots \int_{-\infty}^{\infty}
\frac{{\rm d} \lambda_n}{2 \pi} \prod_{a>b} \frac{\sinh (\lambda_a - \lambda_b)}
{\sinh \left( \left(\lambda_a-\lambda_b - {\rm i} \pi \right) \nu \right)} \nonumber \\
& & \hspace{1cm} \times \prod_{k=1}^{n} \frac{\sinh^{n-k} \left( \left( \lambda_k + {\rm i} \pi/2
\right) \nu \right) \sinh^{k-1} \left( \left(\lambda_k - {\rm i} \pi/2 \right)
\nu \right)}{\cosh^n (\lambda_k)},
\label{Jimbo-Miwa}\end{aligned}$$ where the parameter ${\nu}$ is related to the anisotropy $\Delta$ as $$\nu = \frac{1}{\pi} \cos^{-1}(\Delta). \label{nu}$$ The integral representation was reproduced in the framework of algebraic Bethe ansatz in [@Kitanine00]. A good deal of new developments have been reported for EFP. For example, in the case of ${\Delta=1/2}$ the simple formula for $P(n)$ has been conjectured in [@Razumov00] and then proved in [@Kitanine02]. For ${\Delta=0}$ the asymptotic form of $P(n)$ as $n\gg 1$ was found in [@Shiroishi01]. In either case the large-$n$ asymptotic behavior is given by $ P(n) \sim n^{-\gamma}\, C^{- n^2}
$ with $$\begin{aligned}
C &=& \sqrt{2}, \ \ \ \ \ \gamma = \frac{1}{4},
\ \ \ \ \ (\nu = 1/2), \nonumber \\
C &=& \frac{8}{3\sqrt{3}}, \ \ \ \gamma = \frac{5}{36}, \ \ \ \ (\nu=1/3).
\label{exact}\end{aligned}$$ Moreover, ${P(n)}$ was calculated recently for $\Delta=1$ for strings of the length $n\le6$, see papers [@Boos01; @Boos02; @Boos02b]. Their result suggests that ${P(n)}$ at $\Delta=1$ decays in a Gaussian way as well. Meanwhile, the asymptotic behavior of EFP for the whole critical regime was analyzed in field-theoretical framework, by Abanov and Korepin [@Abanov02]. The Gaussian decay naturally appears in this approach.
Based on those developments, it would appear reasonable that the following asymptotic form of EFP $$P(n) \simeq A\ n^{-\gamma}\, C^{- n^2}, \label{Asym}$$ holds for all over the critical regime ${(-1 < \Delta \le 1)}$. We propose the explicit expressions for the rate of Gaussian decay ${C}$ and the power-law exponent $\gamma$: $$\begin{aligned}
C &=& \frac{\Gamma^2(1/4)}{\pi \sqrt{2 \pi}} \exp \left\{ - \int_0^{\infty}
\frac{{\rm d} t}{t}
\frac{\sinh^2(t \nu) {\rm e}^{-t}}{\cosh(2 t \nu) \sinh(t)} \right\},
\label{C} \\
\gamma &=& \frac{1}{12} + \frac{\nu^2}{3(1-\nu) }. \label{gamma}\end{aligned}$$ One can confirm that above formulas reproduce the exact results for $\Delta =0$ and $\Delta =1/2$ (\[exact\]).
In order to confirm the validity of the formulas for general anisotropies ${0 \le \nu < 1 }$, we have performed extensive numerical calculations of two kinds: One is the first-principle simulation method, namely, the density-matrix renormalization group (DMRG) [@White92; @White93; @Peschel99], and the other is the Monte-Carlo numerical integration of the multiple integration formula (\[Jimbo-Miwa\]) of Jimbo and Miwa. As noted afterwards, those methods are compensative, and we were able to perform reliable simulations for over all the critical regime. As a result, we could confirm fairly definitely that the above general formulas are indeed correct.
Let us turn to addressing the numerical-simulation results. In Tables 1-4, we have listed the DMRG results of EFP for $\nu =0.2,...,0.8$. (The methodological details will be explained afterwards.)
In addition, in each table, we have presented the logarithm of the ratio of two adjacent EFPs, which should behave in the form $$\ln \left(\frac{P(n)}{P(n+1)}\right) \simeq \gamma \ln (1+\frac{1}{n})
+ (2n+1) \ln C, \hspace{1.5cm} \label{Asym2}$$ according to the long-distance asymptotic formula (\[Asym\]). Note that, after taking the ratio, we are able to kill the contribution of the constant factor $A$ of (\[Asym\]), for which, at present, we have no analytical prediction. (We, however, could estimate ${A}$ numerically from our DMRG data as is shown in Table 5.) In this way, the resultant processed data can be directly comparable with the analytical conjecture.
We see that our processed DMRG data $\ln(P(n)/P(n+1))$ are extremely close to the analytical prediction (\[Asym2\]). Actually we find that they coincide up to two or three digits for general values of $n$ and $\nu$. Therefore we conclude that our formulas for the asymptotic form are valid indeed for all over the regime ${0 \le \nu <1}$. Especially see the Table 6, where we compare the asymptotic formula with the known exact values for ${\Delta=1}$ [@Boos01; @Boos02; @Boos02b]. Note that, in this case, we have ${C = \Gamma^2(1/4)/(\pi \sqrt{2 \pi}) = 1.66925..., \ \ \gamma = 1/12}$, and ${A = 0.841.}$
We, however, remark that rather large discrepancies are seen for such cases either very large ${n}$ for small ${\nu}$ or small $n$ for large $\nu$. The former deviations are merely due to the numerical round off errors. Since in computers, real numbers are stored in 8-byte space, and the precision is of the order of $10^{-15}$ at best. Hence it is in principle difficult to calculate the correlations less than the magnitude $< 10^{-12}$ reliably. The latter deviations are not so surprising, because our general formulas should be justified for long distances of EFP. This short-range deviation will be further exploited in the succeeding Monte-Carlo numerical integration analyses.
$n$ ${P(n)}$ ${\ln(P(n)/P(n+1))}$ Asymptotics (\[Asym2\])
----- ----------------------------- ---------------------- -------------------------
1 ${0.5}$ 1.505 1.512
2 ${1.111 \times 10^{-1}}$ 2.447 2.447
3 ${9.614 \times 10^{-3}}$ 3.397 3.397
4 ${3.218 \times 10^{-4}}$ 4.353 4.353
5 ${4.140 \times 10^{-6}}$ 5.312 5.312
6 ${2.041 \times 10^{-8}}$ 6.276 6.271
7 ${3.837 \times 10^{-11}}$ 7.187 7.232
8 ${2.901 \times 10^{-14}}$ 4.683 8.192
9 ${2.682 \times 10^{-16}}$ 3.923 9.155
: DMRG data for ${P(n)}$ : ${\nu=0.2}$[]{data-label="table1"}
$n$ ${P(n)}$ ${\ln(P(n)/P(n+1))}$ Asymptotics (\[Asym2\])
----- ----------------------------- ---------------------- -------------------------
1 ${0.5}$ 1.319 1.320
2 ${1.337 \times 10^{-1}}$ 2.070 2.071
3 ${1.687 \times 10^{-2}}$ 2.851 2.851
4 ${9.752 \times 10^{-4}}$ 3.640 3.640
5 ${2.561 \times 10^{-5}}$ 4.434 4.433
6 ${3.039 \times 10^{-7}}$ 5.231 5.230
7 ${1.626 \times 10^{-9}}$ 6.026 6.025
8 ${3.924 \times 10^{-12}}$ 5.862 6.823
9 ${1.117 \times 10^{-14}}$ 2.824 7.621
: DMRG data for ${P(n)}$ : ${\nu=0.4}$
$n$ ${P(n)}$ ${\ln(P(n)/P(n+1))}$ Asymptotics (\[Asym2\])
----- ----------------------------- ---------------------- -------------------------
1 ${0.5}$ 1.106 1.125
2 ${1.655 \times 10^{-1}}$ 1.594 1.588
3 ${3.360 \times 10^{-2}}$ 2.121 2.115
4 ${4.028 \times 10^{-3}}$ 2.667 2.663
5 ${2.798 \times 10^{-4}}$ 3.223 3.221
6 ${1.115 \times 10^{-5}}$ 3.784 3.783
7 ${2.534 \times 10^{-7}}$ 4.348 4.348
8 ${3.278 \times 10^{-9}}$ 4.909 4.915
9 ${2.420 \times 10^{-11}}$ 5.231 5.483
: DMRG data for ${P(n)}$ : ${\nu=0.6}$
$n$ ${P(n)}$ ${\ln(P(n)/P(n+1))}$ Asymptotics (\[Asym2\])
----- ---------------------------- ---------------------- -------------------------
1 ${0.5}$ 0.896 1.250
2 ${2.041 \times 10^{-1}}$ 1.120 1.221
3 ${6.663 \times 10^{-2}}$ 1.361 1.387
4 ${1.709 \times 10^{-2}}$ 1.615 1.615
5 ${3.400 \times 10^{-3}}$ 1.879 1.870
6 ${5.196 \times 10^{-4}}$ 2.150 2.139
7 ${6.050 \times 10^{-5}}$ 2.430 2.417
8 ${5.336 \times 10^{-6}}$ 2.708 2.701
9 ${3.549 \times 10^{-7}}$ 2.996 2.988
: DMRG data for ${P(n)}$ : ${\nu=0.8}$
${\nu}$ ${\Delta}$ ${C}$ ${\gamma}$ ${A}$
--------- ------------ --------- ------------ -------
0.0 1.0 1.66925 0.0833 0.841
0.2 0.8090 1.61803 0.1 0.816
0.4 0.3090 1.49207 0.1722 0.747
0.6 -0.3090 1.33168 0.3833 0.68
0.8 -0.8090 1.16287 1.15 0.9
: Numerical estimation of the factor ${A}$
$n$ ${P(n)}$ ${\ln(P(n)/P(n+1))}$ Asymptotics (\[Asym2\])
----- ------------------------------- ---------------------- -------------------------
1 ${0.5}$ 1.58685 1.59489
2 ${1.02284 \times 10^{-1}}$ 2.59643 2.59566
3 ${7.62415 \times 10^{-3}}$ 3.60989 3.61060
4 ${2.06270 \times 10^{-4}}$ 4.63019 4.62998
5 ${2.01172 \times 10^{-6}}$ 5.65115 5.65133
6 ${7.06812 \times 10^{-9}}$
: Comparison with exact values for ${P(n)}$ at ${\Delta=1}$
So far, we had analyzed processed EFP data. In Fig. 1, we present the raw (unprocessed) EFP data obtained with the DMRG method. The dotted lines are the analytical conjecture with the factors given in Table 5. We see that our first-principle data are well-fitted by our general formulas (\[C\]) and (\[gamma\]) for various values of $\nu$. However, as noted above, for some cases, there appear deviations.
![EFP is plotted for various valued of distances $n$ and anisotropies $\nu$. We have employed the DMRG method. The dotted lines are our conjecture (\[Asym\]) with the formulas (\[C\]) and (\[gamma\]).[]{data-label="DMRG"}](fig1.eps)
Let us turn to addressing details of the DMRG method [@White92; @White93; @Peschel99]. DMRG is a sort of computer-aided real-space decimation, where the number of states for a renormalized block is retained within a bound tractable in computers. In this way, through the successive applications of DMRG, we are able to access very large system sizes. Hence, this method is suitable for surveying the long-distance asymptotic behavior of EFP. As a matter of fact, efficiency of the method was demonstrated in the preceding works [@Shiroishi01; @Boos02]: This success was rather unexpected, because in general, DMRG is not very efficient for such systems exhibiting criticality. In particular, long-distance asymptotic behavior of two-point correlation function is deteriorated rather severely. Those pathologies would be due to the truncation of bases through numerical renormalizations. However, as far as EFP is concerned, DMRG is proven to be free from such difficulties. Below, we will overview technical points of our numerical simulation. Our algorithm is standard [@White92; @White93], and we refer readers to a recent proceeding [@Peschel99] for full account of methodological details. We have employed the so-called infinite algorithm, which is adequate to investigate bulk properties at the ground state. (For the purpose of studying finite-size scaling behavior, the finite algorithm would be more suitable.) We have remained, at most, $m=300$ bases for a renormalized block. The density matrix eigenvalue $\{w_\alpha\}$ of remaining bases indicates the statistical weight of each remained state. We found $w_\alpha > 5 \cdot 10^{-12}$: That is, we have remained almost all relevant states with appreciable statistical weight $w_\alpha > 5 \cdot 10^{-12}$, which may indicate error of the present simulation. We have repeated 300 renormalizations, and hence, total system size extends to $L=600$. The DMRG data alternates in turn through renormalizations: Note that the number of spins consisting a renormalized block increases by one after another through renormalizations. The problem is that the Hilbert-space structures are incompatible with respect to those cases whether the block contains even or odd spins. (Haldane system ($S=1$) is not affected by this difficulty.) Therefore, we have taken arithmetic mean over those two cases.
The DMRG method is not efficient in the close vicinity of the ferromagnetic isotropic point $\nu=1$. Because in the vicinity of that point, the spin-wave velocity tends to vanish, and there appear numerous nearly-degenerated low-lying levels. Those nearly-degenerated levels are very hard to resolve in the process of numerical diagonalization. The diagonalization is a significant part in the DMRG procedure, and hence, DMRG becomes hardly applicable. In order to compensate this drawback, we have employed another numerical method, that is, the Monte-Carlo integration. We will explain it in the following.
![EFP is plotted for various valued of distances $n$ and anisotropies $\nu$. We have employed the Monte-Carlo integration method for Eq. (\[Jimbo-Miwa\]). The dotted line is Eq. (\[Asym\]) for ${\nu=0.8}$. []{data-label="EFP_fig"}](fig2.eps)
In Fig. 2, we have presented numerical data with Monte-Carlo integration for the multiple integral formula (\[Jimbo-Miwa\]) of Jimbo and Miwa. The technical details will be explained afterwards. From the plot, we see that the decay of EFP is gradually modified as for $\nu \to 1$, and surprisingly enough, the decay approaches the simple exponential formula of $(1/2)^n$ eventually. This feature is precisely in accord with the aforementioned finding that DMRG data deviate from the Gaussian formula as for $\nu \to 1$. Therefore, we see that at least for short-distance asymptotic form of EFP is governed by the simple exponential decay $0.5^n$ as $\nu \to 1$. As a matter of fact, right at $\nu=1$, one may easily verify the pure exponential decay $P(n)$ from the Jimbo-Miwa formula.
The crossover from Gaussian to exponential decay suggests that the quantum fluctuations are suppressed as $\nu \to 1$. As as matter of fact, field theoretical consideration [@Abanov02] reveled that EFP measures the probability of formations of ferromagnetic islands with size $n$ surrounded by antiferromagnetic background from the viewpoint of Euclidean space-time. Therefore, our result suggests that this Euclidean space-time picture is deteriorated for $\nu \to 1$, because the spin-wave velocity tends to vanish, and thus, the size of the islands grows along the imaginary-time direction abruptly.
In the following, we explain technical details of the Monte-Carlo integration. We have used the subroutine package described in the textbook [@NRF]. As a random number generator, we have employed “Mersenne Twistor", MT19937 in Ref. [@Matsumoto98]. For each plot, we have performed eight-billion Monte-Carlo steps. This main Monte-Carlo procedure is preceded by preliminary five million Monte-Carlo steps which are aimed to improve the efficiency of Monte-Carlo sampling by surveying the integrand in the multi-dimensional space. Irrespective of $n$ and $\nu$, we found that the statistical errors are of the order of $10^{-8}$.
To summarize, we have advocated compact explicit formulas (\[C\]) and (\[gamma\]) for the long-distance asymptotic form behavior of EFP. The formulas reproduce the presently-available exact results obtained at special solvable points of $\Delta=0$ and $1/2$, and therefore, we expect that our formulas are valid for over all the critical regime. We have performed extensive simulations of DMRG and Monte-Carlo integration. As a result, we found that our general formulas are indeed correct over the critical regime; see Tables and Fig. 1. In other words, EFP decays obeying the Gaussian formula (\[Asym\]) in the domain of criticality, and the decay rate becomes slower in the ferromagnetic side. This tendency is understood by a physical picture that EFP detects a formation of condensed particles (ferromagnetic string). Apparently, for ferromagnetic side, the condensation becomes promoted, and EFP gets enhanced, albeit the asymptotic form is still maintained to be Gaussian. However, as for the extreme limit to the ferromagnetic isotropic point, the Gaussian decay is smeared out by an exponential form at least for short distances. This fact reflects the suppression of quantum fluctuations in the ferromagnetic side. In other words, this crossover may be regarded as the precursor of the onset of true (long-range order) phase-separation for $\Delta<-1$, where EFP should not decay, but may saturate to a certain finite resident value. Finally we like to remark that the rate of Gaussian decay ${C}$, Eq.(\[C\]), can be evaluated analytically for some other rational values of ${\nu}$. For example, we have found $$\begin{aligned}
C|_{\nu=1/4} &=& 2^{\frac{1}{4}} {\rm e}^{\frac{\rm G}{\pi}}, \ \ \ \ \ \
({\rm G : Catalan \ number}), \nonumber \\
C|_{\nu=1/6} &=& 2 \sqrt{\frac{2}{3}}, \ \ \ \
C|_{\nu=2/3} = \frac{16(\sqrt{2}-1)}{3 \sqrt{3}}, \end{aligned}$$ by use of the integral representations for the logarithm of (multiple) Gamma functions. These facts may indicate a possibility to prove our analytic formulas (\[Asym\]) rigorously for these values of ${\nu}$ just in the similar way as ${\nu=1/2}$ and ${1/3}$.
[**Acknowledgment**]{} We thank A. Abanov, H.Boos, F.Essler, F.Smirnov, M.Takahashi and A. Zamolodchikov for useful discussions. V. Korepin was supported by by NSF Grant PHY-9988566, S. Lukyanov was supported in part by DOE Grant DE-FG02-96ER40959, Y. Nishiyama was supported by Grant-in-Aid for Young Scientists No. 13740240, M. Shiroishi was supported by Grant-in-Aid for Young Scientists No. 14740228.
[99]{}
M. Takahashi: J. Phys. C: Solid State Phys. [**10**]{} (1977) 1289.
M. Jimbo and T. Miwa : “[*Algebraic analysis of solvable lattice models*]{}", AMS, 1995. M. Jimbo and T. Miwa : J. Phys. [**A 29**]{} (1996) 2923.
N. Kitanine, J. M. Maillet, V. Terras : Nucl. Phys. [**B 567**]{}, (2000) 554.
M. Shiroishi, M. Takahashi, Y. Nishiyama : J. Phys. Soc. Jpn [**70**]{} (2001) 3535.
N. M. Bogoliubov, A. G. Izergin, V. E.Korepin : “[*Quantum Inverse Scattering Method and Correlation Functions*]{}”, Cambridge University Press, 1993.
H. E. Boos and V. E. Korepin, : J. Phys. A: Math. Gen. [**34**]{} (2001) 5311. H. E. Boos, V. E. Korepin, Y. Nishiyama and M. Shiroishi: J. Phys. A: Math. Gen. [**35**]{} (2002) 4443. H. E. Boos, V. E. Korepin and F. A. Smirnov, hep-th/0209246.
A. V. Razumov and Yu. G. Stroganov : J. Phys. A: Math. Gen. [**34**]{} (2001) 3185, [*ibid.*]{} 5335.
N. Kitanine, J. M. Maillet, N. A. Slavnov and V. Terras : J. Phys. A: Math. Gen. [**35**]{} (2002) L385.
A. G. Abanov and V. E. Korepin : cond-mat/0206353.
S. R. White : Phys. Rev. Lett. [**69**]{}, 2863 (1992).
S. R. White : Phys. Rev. B [**48**]{}, 10345 (1993). I. Peschel, X. Wang, M. Kaulke and K. Hallberg ed. : “[*Density-Matrix Renormalization — A New Numerical Method in Physics*]{}“”, (Springer-Verlag, Berlin/Heidelberg 1999).
W. H. Press, S. A. Teukolsky, W. T. Vetterling, and B. P. Flannery, [*Numerical recipes in FORTRAN*]{} (Cambridge university press, 1992).
M. Matsumoto and T. Nishimura : ACM Trans. on Modeling and Computer Simulations, [**8**]{}, 3 (1998).
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'Raman spectroscopy, a fast and nondestructive imaging method, can be used to monitor the doping level in graphene devices. We fabricated chemical vapor deposition (CVD) grown graphene on atomically flat hexagonal boron nitride (hBN) flakes and SiO$_2$ substrates. We compared their Raman response as a function of charge carrier density using an ion gel as a top gate. The G peak position, 2D peak position, 2D peak width and the ratio of the 2D peak area to the G peak area show a dependence on carrier density that differs for hBN compared to SiO$_2$. Histograms of two-dimensional mapping are used to compare the fluctuations in the Raman peak properties between the two substrates. The hBN substrate has been found to produce fewer fluctuations at the same charge density owing to its atomically flat surface and reduced charged impurities.'
address:
- '$^1$ Department of Physics, University of Arizona, Tucson, AZ, 85721 USA.'
- '$^2$ National Institute for Materials Science, 1-1 Namiki, Tsukuba, 305-0044 Japan.'
- '$^3$ College of Optical Sciences, University of Arizona, Tucson, AZ, 85721 USA.'
author:
- 'Kanokporn Chattrakun$^1$, Shengqiang Huang$^1$, K. Watanabe$^2$, T. Taniguchi$^2$, A. Sandhu$^{1,3}$ and B. J. LeRoy$^1$'
title: Gate dependent Raman spectroscopy of graphene on hexagonal boron nitride
---
One of the major limits for graphene device performance is the presence of electron and hole puddles caused by charged impurities in the SiO$_2$ substrate and its roughness. Using hexagonal boron nitride (hBN) as a substrate has lead to significantly increased mobility [@Dean2010; @Kim2011] and drastically reduced electron and hole puddles as compared to SiO$_2$ [@Xue2011; @Decker2011] for both exfoliated and chemical vapor deposition (CVD) grown graphene. These improvements are due to the atomic flatness of the hBN substrate and the reduction in charged impurities. Having an improved substrate for graphene presents an opportunity to fabricate high mobility nanoelectronic devices. There are several approaches to obtain a single layer graphene sheet. One of which is the CVD method which can produce monolayer areas that are single domain over areas as large as 20 $\mu$m [@Reina2009; @Li2011]. Using the CVD growth method gives an advantage over other methods such as mechanical exfoliation which produces considerably smaller graphene flakes and is therefore not scalable for large area applications. Moreover, since we would like to compare results between two substrates, hBN and SiO$_2$, we want to use large domains of graphene so that an individual domain will be on both substrates.
Raman spectroscopy is a fast and nondestructive imaging method that can be used to monitor many properties of graphene such as the number of layers [@Ferrari2006], defect density [@Cancado2011], impurity density [@Casiraghi2007], and doping level [@Das2008]. There are two prominent peaks in the Raman spectrum of high quality monolayer graphene; the G peak and the 2D peak. The G peak originates from a first order scattering process involving a doubly degenerate phonon, E$_{2g}$, at the [**$\Gamma$**]{} point, the center of the graphene Brillouin zone. The 2D peak originates from a second order, double resonance process consisting of two phonons both located near the [**K**]{} point, the corner of the Brillouin zone. The peak position of the 2D peak is excitation energy dependent due to the combination of the phonon dispersion relation and the requirement for energy and momentum conservation [@Malard2009; @Ferrari2013]. Doping of graphene shifts its Fermi level leading to controllable changes in its electronic properties as well as the Raman peaks. Doping can be accomplished by electrostatic backgating [@Kim2011; @Ponomarenko2009; @Schwierz2010], chemical methods [@Wang2012a], and electrochemical methods [@Das2008]. Electrochemical doping is able to induce a high charge density in the graphene and is therefore used in this study. Properties such as the peak positions (frequencies), widths, and the ratio of the 2D and G intensities can be used to probe the doping level in graphene. It has been shown that the G peak width [@Yan2007] and position [@Das2008; @Yan2007; @Pisana2007] on SiO$_2$ are sensitive to the charge density. The G peak blue shifts with increasing doping of graphene and saturates at high doping due to the non-adiabatic removal of the Kohn anomaly at the [**$\Gamma$**]{} point [@Pisana2007]. The G phonon lifetime quickly increases with carrier density because the excited electron cannot decay due to the Pauli exclusion principle and momentum conservation. Thus the G peak width, which depends inversely on the phonon lifetime strongly depends on the charge carrier density. The 2D peak intensity [@Casiraghi2009; @Basko2009], and position [@Stampfer2007] on SiO$_2$ also depend on doping. The 2D peak width does not demonstrate such sensitivity since the decay process is unaffected by the Pauli exclusion principle. The 2D peak intensity decreases with increasing charge density due to the increasing of the electron-electron scattering rate at high density [@Basko2009]. When graphene is placed on hBN, the G peak has been reported to red-shift [@Wang2012b; @Forster2012; @Ahn2013] while the shift of the 2D peak is reported to depend on the sample preparation procedure [@Wang2012b; @Ahn2013]. The carrier dependent Raman properties of graphene have been widely and intensively studied on SiO$_2$ but the systematic doping dependence on hBN has not yet been investigated.
Here we report Raman spectroscopy measurements as a function of doping level for graphene on hBN and SiO$_2$. Graphene was synthesized by low-pressure CVD on Cu foil [@Li2011]. The foil was heated to 1040 $^{\circ}$C and annealed for 20 minutes under a H$_2$ gas flow rate of 4 sccm at a pressure of 20 mTorr. Then the Cu was exposed to CH$_4$ and H$_2$ gas at 1.3 and 4 sccm respectively for 1 hour. The system was cooled down to 350 $^{\circ}$C under the same CH$_4$ and H$_2$ flow rate. Using large area CVD graphene allows the same graphene flake to be probed both on the hBN and on the underlying SiO$_2$. Single crystal hBN flakes [@Taniguchi2007] were mechanically exfoliated onto a Si substrate with 285 nm thick SiO$_2$. After exfoliating the hBN, the samples were annealed under O$_2$ and Ar gas at 500 $^{\circ}$C for 3 hours to remove any excess tape residue. PMMA was spun on top of the Cu foil then an aqueous HCl solution was used to etch the Cu. The graphene on PMMA was thoroughly rinsed in deionized water and deposited on the hBN flakes. The sample was soaked in acetone to dissolve the PMMA, then rinsed in IPA and dried with ultra-high purity nitrogen gas. Thermal annealing under H$_2$ and Ar gas was performed at 350 $^{\circ}$C for 3 hours to remove any PMMA residue [@Ishigami2007]. Cr/Au contacts were written using electron beam lithography for electrical transport measurements. The device was annealed in H$_2$ and Ar gas at 350 $^{\circ}$C for another 3 hours after deposition of the metal contacts before any measurements. The doping was accomplished using electrochemical gating using a polymer electrolyte, LiClO$_4$:PEO [@Das2008] (1:8 by weight). Raman spectroscopy was performed using a Nd:YAG laser (532 nm) on graphene with and without hBN underneath, the spot size was 1 $\mu$m with a 100x (NA = 0.8) objective lens. The voltage on the ion gel was varied from -1.5 V to 3 V. Two-dimensional Raman mapping was performed at -1.5, 0, and 2 V on the ion gel in the area shown in figure 1a. We studied the G peak position, 2D peak position, 2D peak width, and the ratio of the 2D to G peak areas as a function of doping level on both the hBN flakes and on the SiO$_2$ substrate.
Figure 1(a) shows an optical image of one of the devices measured. The entire area is covered with graphene; the darker area on the left has hBN with a uniform thickness of 10.5 nm determined from its optical contrast [@Golla2013]; the bright green area near the center of the image is thicker hBN; the area on the right has only graphene on SiO$_2$. The total size of the image is 12.6 by 3.6 $\mu$m and the scale bar is 2 $\mu$m. The dashed line indicates a region of bilayer graphene (the thicker hBN and bilayer graphene regions, shown in dark grey in figures 2(a), 3(a), and 4(a), are excluded from all of our analyses). Figure 1(b) shows a schematic diagram of the experimental setup used for the two-dimensional Raman mapping measurements. A Nd:YAG laser, wavelength of 532 nm, is sent through a 100x objective lens and focused onto the sample. The reflected light is dispersed using a 600 lines/mm grating and the resulting spectrum is imaged on a thermoelectrically cooled CCD. The spectral resolution is 1 cm$^{-1}$. For two-dimensional mapping, the sample position is moved using a closed-loop piezoelectric stage. The Raman spectrum has two peaks from the graphene; the G peak at 1590 cm$^{-1}$, and the 2D peak at 2690 cm$^{-1}$ [@Ferrari2006; @Graf2007], one peak due to hBN at 1365 cm$^{-1}$ [@Geick1966] and two additional peaks from the ion gel near 1280 cm$^{-1}$ and 1480 cm$^{-1}$ [@Yoshihara1964]. A typical Raman spectrum is shown in figure 1(c) showing a 2D peak that is several times stronger than the G peak. The spectra are fitted with a single Lorentzian for all G, 2D and hBN peaks. The 2D peak can be fitted with a single Lorenztian because the graphene is supported on hBN or SiO$_2$ and the system has sufficiently high level of doping to overcome the intrinsic bimodal lineshape of the 2D peak seen in freestanding graphene [@Berciaud2013]. Figure 1(d) plots the ratio of the integrated area of the 2D and G peak for graphene on hBN (blue) and SiO$_2$ (red) as a function of the voltage applied to the ion gel. For stability of the sample, small steps and extended time between each voltage step are required. The voltage was applied in steps of 0.05 V, after each step there was a waiting time of one minute for the ion gel charge to stabilize before collecting data. The area ratio on hBN is greater than that on SiO$_2$ and both vary as a function of gate voltage or charge density. The ratio has a maximum value near the Dirac point of the sample ( -1V) and decreases as the charge density is increased on the graphene for both holes and electrons. As the voltage on the ion gel is changed, the Fermi energy of the graphene shifts. As the Fermi energy shifts, the intensity of the G peak remains constant but the intensity of the 2D peak is reduced leading to a decrease in the ratio of the 2D peak area to the G peak area. The 2D peak intensity depends on doping through the scattering rate of the photoexcited electrons and holes. The rate of electron-electron collisions is sensitive to the density of carriers, increasing the scattering rate as the charge density increases resulting in a decrease of the 2D peak intensity [@Basko2009].
Figure 2(a) shows a two dimensional Raman map of the G peak position ($\omega_G$) for the area shown in figure 1(a). From the map, the G peak position is lower on hBN than on SiO$_2$. The data in figure 2(a) was acquired with 0 V on the ion gel. The range of peak positions is 1584-1598 cm$^{-1}$ on hBN and 1588-1602 cm$^{-1}$ on SiO$_2$. We have performed similar analysis on ten other flakes and found that the G peak is on average 3 cm$^{-1}$ lower on hBN compared to SiO$_2$. We have found that the thickness of the hBN has a strong effect on the strength of the G and 2D peaks in graphene and therefore we have only performed the analysis for thin hBN flakes of uniform thickness. G peak positions on both hBN (blue) and SiO$_2$ (red) shift to higher frequency for charge densities that are away from the neutrality point for both electrons and holes as shown in figure 2(b). The increase of the G peak position along with the narrowing of the peak, which is also observed in our measurements, is a result of increased charge density [@Pisana2007; @Yamamoto2012]. From the shift of the G peak position with the applied voltage, we determined the maximum induced carrier density to be $1\times10^{13}$ cm$^{-2}$ when 3 V was applied to the ion gel [@Lazzeri2006; @Chen2011]. From our two-dimensional mapping of ten different graphene on hBN flakes, we found that the average standard deviation of the histogram of the G peak position was 1 cm$^{-1}$ on hBN and 1.6 cm${-1}$ on SiO$_2$. These values can be converted to fluctuations in energy by using the movement of the G peak position with energy (42 cm$^{-1}$/eV) [@Chen2011]. These fluctuations in energy are then converted to variations in charge density using $n=\frac{1}{\pi}(\frac{E_F}{\hbar v_F})^2$, where $v_F$ is the Fermi velocity which is $1.1\times10^6$ m/s in graphene, giving average charge fluctuations of $6.1\times10^{10}$ cm$^{-2}$ on hBN and $1.56\times10^{11}$ cm$^{-2}$ on SiO$_2$. As observed in other local spectroscopy measurements, the charge fluctuations in graphene are reduced on the hBN as compared to the SiO$_2$ [@Xue2011; @Decker2011]. Figures 2(c), 2(d), and 2(e) are histograms extracted from the two dimensional mappings at voltages on the ion gel of -1.5, 0, and 2 V respectively. The G peak positions are lower on hBN (blue) than SiO$_2$ (red) for the 0 and 2 V gate voltages, but approximately the same at -1.5 V. The difference of the peak positions at 0 V is more pronounced than 2 V. This indicates the G peak on hBN moves more slowly than on SiO$_2$ at negative voltages and faster at positive voltages.
A two-dimensional map of the 2D peak position ($\omega_{2D}$) with 2 V on the ion gel ($n = 9\times10^{12}$ cm$^{-2}$) is shown in figure 3(a). The figure indicates a significant difference in the 2D peak position on the hBN flake compared to the SiO$_2$ substrate. The shift of the 2D peak position can be used to indicate the doping level. Figure 3(b) plots the 2D peak position as a function of doping on local spots of graphene on SiO$_2$ and hBN. It decreases for $n$ doping and increases for $p$ doping on both substrates. At high electron concentrations, the 2D peak strongly blue-shifts on SiO$_2$ but there is only a much weaker shift on hBN. The shift in the 2D position on hBN results from a reduction of the Kohn anomaly at the K point influenced by an enhanced screening from the dielectric substrate, which is not a direct interaction between graphene and the substrate. The electron-electron interaction depends on the electronic screening by the environment. Therefore such interaction is expected to be reduced by a dielectric substrate. Although hBN has a similar dielectric constant to SiO$_2$, its atomically flat surface provides a different dielectric environment for graphene. The reduced charged impurities and increased flatness affects the Fermi velocity of the carriers as well as the electron-electron and electron phonon coupling in graphene, thus enhancing the dielectric screening for graphene on hBN [@Forster2012; @Graf2007]. Figures 3(c), 3(d), and 3(e) are histograms extracted from the two-dimensional mapping at -1.5, 0, and 2 V on the ion gel respectively. The histograms of the 2D peak position at 0 V (figure 2(c)) are nearly identical between the two substrates; hBN (blue) and SiO$_2$ (red). For the particular location chosen for figure 3(b) the 2D peak is slightly blue shifted on the SiO$_2$ compared to the hBN. At both positive and negative gate voltage, the 2D peak positions show more spatial variation on both substrates compared to the 0 V. The reduced movement of the 2D peak on hBN is clearly seen in figure 3(d), where the two distributions no longer overlap.
Figure 4(a) maps the 2D peak width ($\Gamma_{2D}$) at 0 V on the ion gel. The map indicates the 2D peak is narrower on hBN compared to SiO$_2$. Using hBN as a substrate for graphene, has been shown to lead to a reduction in surface roughness and charged impurities [@Dean2010; @Xue2011; @Decker2011]. Both of these effects reduce the phonon scattering rate and therefore lead to a longer phonon lifetime and hence a narrower 2D peak [@Pisana2007; @Forster2012; @Graf2007]. Figure 4(b) plots the 2D peak width as a function of the applied voltage on the ion gel. The hBN substrate is shown in blue, and the SiO$_2$ substrate is red. Since the peak width originates from the scattering of phonons, the SiO$_2$ substrate that has greater scattering responds less to the applied gate voltage or carrier concentration because of the high scattering rate even near the Dirac point. On the other hand, the hBN substrate that has less scattering at low charge density responds faster to the increasing carrier density as seen by the increasing 2D width at higher applied voltage. Figures 4(c), 4(d), and 4(e) are histograms of the 2D peak widths at -1.5, 0, and 2 V gate voltages respectively. They show that the graphene 2D peak is narrower on hBN than SiO$_2$ for all charge densities however the width of the 2D peak increases at higher charge density (2 V on the ion gel). The distribution of the histogram gives a measure of the fluctuations on the two substrates. Near the Dirac point (figure 4(c)), the widths of the distributions are comparable as the FWHM is 3.3 cm$^{-1}$ on hBN and 3.7 cm$^{-1}$ on SiO$_2$. When further away from the Dirac point (figures 4(d) and 4(e)), the hBN substrate shows a narrower distribution, indicating less charge fluctuations [@Wang2012b]. The histograms in figures 4(d) and 4(e) have FWHM of 2.7, 4.7 cm$^{-1}$ on hBN and 4.2, 6.0 cm$^{-1}$ on SiO$_2$ respectively.
We have investigated the Raman properties of graphene such as the G peak position, the 2D peak position, the 2D peak width, and the area ratio of the 2D peak to the G peak as a function of carrier density on two different substrates; hBN and SiO$_2$. We have found that the area ratio has a higher value on hBN and varies with the carrier density. The G peak position depends on the carrier density, and the position of the G peak on hBN is lower than on SiO$_2$ for the same charge density. The analysis of two-dimensional mapping results show less fluctuation of the G peak position on hBN indicating less charge fluctuation on the hBN substrate as compared to SiO$_2$. We have found a reduction by a factor of three in the charge fluctuation on hBN as compared to SiO$_2$. The 2D peak positions are comparable near the Dirac point on both substrates but show a strong blue-shift at high density on SiO$_2$ while remaining approximately constant on hBN. The 2D peak width has a lower value on hBN but shows more variation at high carrier concentration indicating less scattering. The distribution of the histograms has been found to be narrower on hBN, and becomes more pronounced at higher density. This reflects a lower fluctuation of carrier density on graphene when hBN is used as a substrate. The variation of the Raman peaks with charge density can be used to quickly identify the doping in graphene on hBN devices.
K. C., S. H. and B. J. L. acknowledge support from the U. S. Army Research Office under contract W911NF-09-1-0333 and NSF CAREER Award No. DMR/0953784.
References {#references .unnumbered}
==========
[99]{}
Dean C R, Young A F, Meric I, Lee C, Wang L, Sorgenfrei S, Watanabe K, Taniguchi T, Kim P, Shepard K L and Hone J 2010 [*Nat. Nanotechnol.*]{} [**5**]{} 722-726. Kim E, Yu T, Song E S, and B. Yu, Appl. Phys. Lett. [**98**]{} 262103 (2011). Xue J, Sanchez-Yamagishi J, Bulmash D, Jacquod P, Deshpande A, Watanabe K, Taniguchi T, Jarillo-Herrero P and LeRoy B J 2011 [*Nat. Mater.*]{} [**10**]{} 282-285. Decker R, Wang Y, Brar V W, Regan W, Tsai H Z, Wu Q, Gannett W, Zettl A and Crommie M F 2011 [*Nano Lett.*]{} [**11**]{} 2291-2295. Reina A, Jia X, Ho J, Nezich D, Son H, Bulovic V, Dresselhaus MS and Kong J 2009 [*Nano Lett.*]{} [**9**]{} 30-35. Li X, Magnuson CW, Venugopal A, Tromp RM, Hannon JB, Vogel EM, Colombo L and Ruoff RS 2011 [*J. Am. Chem. Soc.*]{} [**133**]{} 2816-2819. Ferrari AC, Meyer JC, Scardaci V, Casiraghi C, Lazzeri M, Mauri F, Piscanec S, Jiang D, Novoselov KS, Roth S and Geim AK 2006 [*Phys. Rev. Lett.*]{} [**97**]{} 187401. Cancado L G, Jorio A, Martins Ferreira E H, Stavale F, Achete C A, Capaz R B, Moutinho M V O, Lombardo A, Kulmala T S and Ferrari A C 2011 [*Nano Lett.*]{} [**11**]{} 3190-3196. Casiraghi C, Pisana S, Novoselov K S and Ferrari A C 2007 [*Appl. Phys. Lett.*]{} [**91**]{} 233108. Das A, Pisana S, Chakraborty B, Piscanec S, Saha S K, Waghmare U V, Novoselov K S, Krishnamurthy H R, Giem A K, Ferrari A C and Sood A K 2008 [*Nat. Nanotechnol.*]{} [**3**]{} 210-215. Malard L M, Pimenta M A, Dresselhaus G and Dresselhaus M S 2009 [*Phys. Rep.*]{} [**437**]{} 51-87. Ferrari A C and Basko D M 2013 [*Nat. Nanotechnol.*]{} [**8**]{} 235-246. Ponomarenko L A, Yang R, Mohiuddin T M, Katsnelson M I, Novoselov K S, Morozov S V, Zhukov A A, Schedin F, Hill E W and Geim A K 2009 [*Phys. Rev. Lett.*]{} [**102**]{} 206603. Schwierz F 2010 [*Nat. Nanotechnol.*]{} [**5**]{} 487-496. Wang Q H, Jin Z, Kim K K, Hilmer A J, Paulus G L, Shih C J, Ham M H, Sanchez-Yamagishi J D, Watanabe K, Taniguchi T, [*et al.*]{} 2012 [*Nat. Chem.*]{} [**4**]{} 724-732. Yan J, Zhang Y, Kim P and Pinczuk A 2007 [*Phys. Rev. Lett.*]{} [**98**]{} 166802. Pisana S, Lazzeri M, Casiraghi C, Novoselov K S, Geim A K, Ferrrari A C and Mauri F 2007 [*Nat. Mater.*]{} [**6**]{} 198-201. Casiraghi C 2009 [*Phys. Rev. B*]{} [**80**]{} 233407. Basko D M, Piscanec S and Ferrari A C 2009 [*Phys. Rev. B*]{} [**80**]{} 165413. Stampfer C, Molitor M, Graf D, Ensslin K, Jungen A, Hierold C and Wirtz L 2007 [*Appl. Phys. Lett.*]{} [**91**]{} 241907. Wang L, Chen Z, Dean C R, Taniguchi T, Watanabe K, Brus L E and Hone J 2012 [*ACS Nano*]{} [**6**]{} 9314-9319. Forster F, Molina-Sanchez A, Engels S, Epping A, Watanabe K, Taniguchi T, Wirtz L and Stampfer C 2013 [*Phys. Rev. B*]{} [**88**]{} 085419. Ahn G, Kim H R, Ko T Y, Choi K, Watanabe K, Taniguchi T, Hong B H and Ryu S 2013 [*ACS Nano*]{} [**7**]{} 1533-1541. Taniguchi T and Watanabe K 2007 [*J. Cryst. Growth*]{} [**303**]{} 525-529. Golla D, Chattrakun K, Watanabe K, Taniguchi T, LeRoy B J and Sandhu A 2013 [*Appl. Phys. Lett.*]{} [**102**]{} 161906. Graf D, Molitor F, Ensslin K, Stampfer C, Jungen A, Hierold C and Wirtz L 2007 [*Nano Lett.*]{} [**7**]{} 238-242. Geick R and Perry C H 1966 [*Phys. Rev.*]{} [**146**]{} 543-547. Yoshihara T, Tadokoro H and Murahashi S 1964 [*J. Chem. Phys.*]{} [**41**]{} 2902-2911. Berciaud S, Li X, Htoon H, Brus L E, Doorn S K and Heniz T F 2013 [*Nano Lett.*]{} [**13**]{} 3517-3523. Yamamoto M, Einstein T L, Fuhrer M S and Cullen W G 2012 [*ACS Nano*]{} [**6**]{} 8335-8341. Lazzeri M and Mauri F 2006 [*Phys. Rev. Lett.*]{} [**97**]{} 266407. Chen C F, Park C H, Boudouris B W, Horng J, Geng B, Girit C, Zettl A, Crommie M F, Segalman R A, Louie S G and Wang F 2011 [*Nature*]{} [**471**]{} 617-620. Ishigami M, Chen J H, Cullen W G, Fuhrer M S and Williams E D 2007 [*Nano Lett.*]{} [**7**]{} 1643-1648.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'The Duffin-Kemmer-Petiau (DKP) equation with a square step potential is used in a simple way with polymorphic purposes. It proves adequate to refuse a proposed new current that is currently interpreted as a probability current,to show that the Klein paradox does exist in the DKP theory and to revise other minor misconceptions diffused in the literature.'
author:
- |
T.R. Cardoso, L.B. Castro[^1], A.S. de Castro[^2]\
\
UNESP - Campus de Guaratinguetá\
Departamento de Física e Química\
12516-410 Guaratinguetá SP - Brazil
title: |
Inconsistencies of a purported\
probability current in the Duffin-Kemmer-Petiau theory
---
The first-order Duffin-Kemmer-Petiau (DKP) equation has experimented a renewal of life due to the discovery of a new conserved four-vector current density [@gho1]-[@gho4], whose positive-definite time component would be a candidate to a probability density, and as a bonus a hope for avoiding Klein's paradox for bosons [@gho4]. The DKP equation for a boson minimally coupled to the electromagnetic field is given by$$\left( i\beta ^{\mu }D_{\mu }-m\right) \psi =0 \label{dkp}$$where the matrices $\beta ^{\mu }$ satisfy the algebra$$\beta ^{\mu }\beta ^{\nu }\beta ^{\lambda }+\beta ^{\lambda }\beta ^{\nu
}\beta ^{\mu }=g^{\mu \nu }\beta ^{\lambda }+g^{\lambda \nu }\beta ^{\mu }
\label{beta}$$the covariant derivative is given by $D_{\mu }=\partial _{\mu
}+ieA_{\mu }$ and the metric tensor is $g^{\mu \nu }=\,$diag$\,(1,-1,-1,-1)$. The second-order Klein-Gordon and Proca equations are obtained when one selects the spin-0 and spin-1 sectors of the DKP theory. A well-known conserved four-current is given by $$J^{\mu }=\bar{\psi}\beta ^{\mu }\psi \label{corrente}$$where the adjoint spinor $\bar{\psi}=\psi ^{\dagger }\eta ^{00}$ with$$\eta ^{\mu \nu }=\beta ^{\mu }\beta ^{\nu }+\beta ^{\nu }\beta ^{\mu
}-g^{\mu \nu } \label{etamunu}$$ and $\left( \beta ^{\mu }\right) ^{\dagger }=\eta ^{00}\beta ^{\mu }\eta
^{00}$. The time component of this current is not positive definite but it may be interpreted as a charge density. An alleged new conserved current, however, is written as [@gho1], [@gho3] $$S^{\mu }=\bar{\psi}\eta ^{\mu \nu }\psi \,u_{\nu } \label{corrente s}$$Here $u_{\nu }$ is the unity timelike four-velocity of the observer ($u^{\nu }u_{\nu }=1$). Since $S^{0}=\psi ^{\dagger }\psi \geqslant
0$ in the lab frame, it might be tempting to interpret this alternative current as a probability current.
In the present work, the simple problem of scattering in a square step potential, considered as a time-component of the electromagnetic four-potential, is used to show not only that this new current leads to inconsistencies sufficient enough to reject it as a true probability current, but also to show that Klein's paradox is absent in Ref. [@gho4] just because it was not searched for, and that it is not necessary to refer to limiting cases of smooth potentials for finding the appropriate boundary conditions for discontinuous potentials as done in [@che].
Criticism on $S^{\mu }$ has already a precedent. Struyve et al. [@stru] have shown that $S^{\mu }$ is not conserved when the electromagnetic interaction is introduced in the DKP equation and so $S^{0}$ can not be interpreted as a particle probability. They found that$$\partial _{\mu }\Theta ^{\mu \nu }=\frac{e}{m}F^{\nu \mu }\,J_{\mu }
\label{tensor}$$where the energy-momentum tensor $\Theta ^{\mu \nu }=\bar{\psi}\eta ^{\mu
\nu }\psi $ and the electromagnetic field tensor $F_{\mu \nu }=\partial
_{\mu }A_{\nu }-\partial _{\nu }A_{\mu }$. Therefore, $S^{\mu }=\Theta ^{\mu
\nu }u_{\nu }$ is not conserved for an arbitrary observer. Struyve et al. [@stru] argued that the energy-momentum tensor is not conserved because a charged particle exchanges energy and momentum with the electromagnetic field. Recently, Datta [@dat] has tried to save the status of the new current by considering$$S^{\prime \mu }=S^{\mu }+\frac{e}{m}\int dx^{\mu }F^{\lambda \nu
}\,J_{\lambda }u_{\nu } \label{newS}$$with $\partial _{\mu }S^{\prime \mu }=0$. As a matter of fact, Datta tried to take care of the remark in Ref. [@stru] about the exchanges of energy and momentum of the particle with the external field by including the electromagnetic field as a part of the system.
Let us consider the one-dimensional time component of the static electromagnetic potential, so that the time-independent DKP equation can be written as$$\{\beta ^{0}[E-eA_{0}(z)]+i\beta ^{3}\frac{d}{dz}-m\}\varphi (z)=0
\label{dkp ind temp}$$where the decomposition $\psi (z,t)=\varphi (z)\exp (-iEt)$ has been used.
For the case of spin 0, we use the representation for the $\beta ^{\mu }$matrices given by [@ned1]$$\beta ^{0}=%
\begin{pmatrix}
\theta & \overline{0} \\
\overline{0}^{T} & \mathbf{0}%
\end{pmatrix}%
\quad \mathrm{and}\quad \beta ^{i}=%
\begin{pmatrix}
\widetilde{0} & \rho _{i} \\
-\rho _{i}^{T} & \mathbf{0}%
\end{pmatrix}%
,\quad i=1,2,3 \label{m1}$$where$$\begin{aligned}
\ \theta &=&%
\begin{pmatrix}
0 & 1 \\
1 & 0%
\end{pmatrix}%
\quad \quad \rho _{1}=%
\begin{pmatrix}
-1 & 0 & 0 \\
0 & 0 & 0%
\end{pmatrix}
\notag \\
&& \label{m2} \\
\rho _{2} &=&%
\begin{pmatrix}
0 & -1 & 0 \\
0 & 0 & 0%
\end{pmatrix}%
\quad \quad \rho _{3}=%
\begin{pmatrix}
0 & 0 & -1 \\
0 & 0 & 0%
\end{pmatrix}
\notag\end{aligned}$$$\overline{0}$, $\widetilde{0}$ and $\mathbf{0}$ are 2$\times $3, 2$\times $2 and 3$\times $3 zero matrices, respectively, while the superscript T designates matrix transposition. The five-component spinor can be written as $\varphi ^{T}=\left( \varphi _{1},...,\varphi _{5}\right) $ in such a way that the DKP equation decomposes into$$\text{O}_{{\verb|KG|}}\,\varphi _{1}=0,\quad \varphi _{2}=\frac{E-eA_{0}}{m}%
\varphi _{1},\quad \varphi _{3}=\varphi _{4}=0,\quad \varphi _{5}=\frac{i}{m}%
\frac{d\varphi _{1}}{dz} \label{esp1}$$where O$_{{\verb|KG|}}=-d^{2}/dz^{2}+m^{2}-\left( E-eA_{0}\right)
^{2}$. Using now the components of the spinor, the time and space components of the currents $J^{\mu }$ and $S^{\mu }$ can be written as $$\begin{aligned}
J^{0} &=&2\Re\left( \varphi _{1}\varphi _{2}^{\ast }\right) ,\quad
J^{3}=-2\Re(\varphi _{1}\varphi _{5}^{\ast }) \notag \\
&& \label{corrent} \\
S^{0} &=&\left\vert \varphi _{1}\right\vert ^{2}+\left\vert \varphi
_{2}\right\vert ^{2}+\left\vert \varphi _{5}\right\vert ^{2},\quad
S^{3}=-2\Re\left( \varphi _{2}\varphi _{5}^{\ast }\right) \notag\end{aligned}$$Note that there is no reason to require that the spinor and its derivative are continuous across finite discontinuities of the potential, as naively advocated in Ref. [@che]. A little careful analysis reveals, though, that proper matching conditions follow from the differential equations obeyed by the spinor components, as they should be, avoiding in this manner the hard tasking of recurring to the limit process of smooth potentials. Only the first component of the spinor satisfies the second-order Klein-Gordon equation, so that $\varphi _{1}$ and its first derivative are continuous even the potential suffers finite discontinuities. In this case of a discontinuous potential, $\varphi _{2}$ is discontinuous and so are $J^{0}$, $S^{0}$ and $S^{3}$. The discontinuity of $J^{0}$ does not matter if it is to be interpreted as a charge density. As for $S^{\mu }$, it is an obvious nonsense to interpret it as a probability current seeing that a probability density should always be continuous and that the probability flux should be uniform in a stationary regime. In this point we are faced with serious defects of $S^{\mu }$. Nevertheless, despite those unpleasant properties of $S^{\mu }$ we shall explore the scattering in a square step potential in order to clarify additional misapprehensions of the DKP theory.
The one-dimensional square step potential is expressed as $$A_{0}\left( z\right) =V_{0}\,\theta \left( z\right) \label{pot}$$where $\theta \left( z\right) $ denotes the Heaviside step function. For $z<0$ the DKP equation has the solution $$\varphi \left( z\right) =\varphi _{+}e^{+ikz}+\varphi _{-}e^{-ikz}
\label{sol z<0}$$where$$\varphi _{\pm }^{T}=\frac{a_{\pm }}{\sqrt{2}}\left( 1,\frac{E}{m},0,0,\mp
\frac{k}{m}\right) \label{sol2}$$and $k=\sqrt{E^{2}-m^{2}}$. For $\left\vert E\right\vert >m$, the solution expressed by (\[sol z<0\]) and (\[sol2\]) describes plane waves propagating on both directions of the $Z$-axis with group velocity $%
v_{g}=dE/dk$ equal to the classical velocity. If we choose particles inciding on the potential barrier $\left( E>m\right) $, $\varphi _{+}\exp
(+ikz)$ will describe incident particles ($v_{g}=+k/E>0$), whereas $\varphi
_{-}\exp (-ikz)$ will describe reflected particles ($v_{g}=-k/E<0$). The flux related to the standard current $J^{\mu }$, corresponding to $\varphi $ given by (\[sol z<0\]), is expressed as $$J^{3}=\frac{k}{m}\left( \left\vert a_{+}\right\vert ^{2}-\left\vert
a_{-}\right\vert ^{2}\right) \label{j1}$$Note that the relation $J^{3}=J^{0}v_{g}$ maintains for the incident and reflected waves, since$$J_{\pm }^{0}=\frac{E}{m}\left\vert a_{\pm }\right\vert ^{2} \label{j2}$$On the other hand, for $z>0$ one should have $v_{g}\geq 0$ in such a way that the solution in this region of space describes an evanescent wave or a progressive wave running away from the potential interface. The general solution has the form $$\varphi _{\text{t}}\left( z\right) =\left( \varphi _{\text{t}}\right)
_{+}e^{+iqz}+\left( \varphi _{\text{t}}\right) _{-}e^{-iqz} \label{sol z0}$$where$$\left( \varphi _{\text{t}}\right) _{\pm }^{T}=\frac{b_{\pm }}{\sqrt{2}}%
\left( 1,\frac{E-eV_{0}}{m},0,0,\mp \frac{q}{m}\right) \label{sol3}$$and $q=\sqrt{\left( E-eV_{0}\right) ^{2}-m^{2}}$. Due to the twofold possibility of signs for the energy of a stationary state, the solution involving $b_{-}$ can not be ruled out a priori. As a matter of fact, this term may describe a progressive wave with negative energy and phase velocity $v_{ph}=|E|/q>0$. One can readily envisage that three different classes of solutions can be segregated:
- **Class A.** For $eV_{0}<E-m$ one has $q\in
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
$, and the solution describing a plane wave propagating in the positive direction of the $Z$-axis with group velocity $v_{g}=q/\left(
E-eV_{0}\right) $ is possible only if $b_{-}=0$. In this case the components of the standard current are given by
$$J^{0}=\frac{E-eV_{0}}{m}\left\vert b_{+}\right\vert ^{2},\quad J^{3}=\frac{q%
}{m}\left\vert b_{+}\right\vert ^{2} \label{c11}$$
- **Class B.** For $E-m<eV_{0}<E+m$ one has that $q=\pm i\left\vert
q\right\vert $, and (\[sol z0\]) with $b_{\mp }=0$ describes an evanescent wave. The condition $b_{\mp }=0$ is necessary for furnishing a finite current as $z\rightarrow \infty $. In this case $$J^{0}=\frac{E-eV_{0}}{m}e^{-2\left\vert q\right\vert z}\left\vert b_{\pm
}\right\vert ^{2},\quad J^{3}=0 \label{c2}$$
- **Class C.** With $eV_{0}>E+m$ it appears again the possibility of propagation in the positive direction of the $Z$-axis, now with $b_{+}=0$ and a group velocity given by $v_{g}=q/\left( eV_{0}-E\right) $. The standard current takes the form$$J^{0}=\frac{E-eV_{0}}{m}\left\vert b_{-}\right\vert ^{2},\quad J^{3}=-\frac{q%
}{m}\left\vert b_{-}\right\vert ^{2} \label{c33}$$In this last class we meet a bizarre circumstance as long as both $J^{0}$ and $J^{3}$ are negative quantities. The maintenance of the relation $%
J^{3}=J^{0}v_{g}$, though, is a license to interpret the solution $\left(
\varphi _{\text{t}}\right) _{-}\exp (-iqz)$ as describing the propagation, in the positive direction of the $Z$-axis, of particles with electric charges of opposite sign to the incident particles. This interpretation is consistent if the particles moving in this region have energy $-E$ and are under the influence of a potential $-eV_{0}$. It means that, in fact, the progressive wave describes the propagation of antiparticles in the positive direction of the $Z$-axis.
The demand for continuity of $\varphi _{1}$ and $d\varphi _{1}/dz$ at $z=0$ fixes the wave amplitudes in terms of the amplitude of the incident wave, viz. $$\frac{a_{-}}{a_{+}}=\left\{
\begin{array}{c}
\frac{k-q}{k+q} \\
\\
\frac{\left( k-i|q|\right) ^{2}}{k^{2}+|q|^{2}} \\
\\
\frac{k+q}{k-q}%
\end{array}%
\begin{array}{c}
\text{\textrm{for the class}}{\text{ }\mathbf{A}} \\
\\
\text{\textrm{for the class}}{\text{ }\mathbf{B}} \\
\\
\text{\textrm{for the class}}{\text{ }\mathbf{C}}%
\end{array}%
\right. \label{12}$$$$\frac{b_{+}}{a_{+}}=\left\{
\begin{array}{c}
\frac{2k}{k+q} \\
\\
\frac{2k\left( k-i|q|\right) }{k^{2}+|q|^{2}} \\
\\
0%
\end{array}%
\begin{array}{c}
\text{\textrm{for the class}}{\text{ }\mathbf{A}} \\
\\
\text{\textrm{for the class}}{\text{ }\mathbf{B}} \\
\\
\text{\textrm{for the class}}{\text{ }\mathbf{C}}%
\end{array}%
\right. \label{13}$$$$\frac{b_{-}}{a_{+}}=\left\{
\begin{array}{c}
0 \\
\\
0 \\
\\
\frac{2k}{k-q}%
\end{array}%
\begin{array}{c}
\text{\textrm{for the class}}{\text{ }\mathbf{A}} \\
\\
\text{\textrm{for the class}}{\text{ }\mathbf{B}} \\
\\
\text{\textrm{for the class}}{\text{ }\mathbf{C}}%
\end{array}%
\right. \label{14}$$
Now we focus attention on the calculation of the reflection ($R$) and transmission ($T$) coefficients. The reflection (transmission) coefficient is defined as the ratio of the reflected (transmitted) flux to the incident flux. Since $\partial J^{0}/\partial t=0$ for stationary states, one has that $J^{3}$ is independent of $z$. This fact implies that $$R=\left\{
\begin{array}{c}
\left( \frac{k-q}{k+q}\right) ^{2} \\
\\
1 \\
\\
\left( \frac{k+q}{k-q}\right) ^{2}%
\end{array}%
\begin{array}{c}
\text{\textrm{for the class}}{\text{ }\mathbf{A}} \\
\\
\text{\textrm{for the class}}{\text{ }\mathbf{B}} \\
\\
\text{\textrm{for the class}}{\text{ }\mathbf{C}}%
\end{array}%
\right. \label{15}$$ $$T=\left\{
\begin{array}{c}
\frac{4kq}{(k+q)^{2}} \\
\\
0 \\
\\
-\frac{4kq}{(k-q)^{2}}%
\end{array}%
\begin{array}{c}
\text{\textrm{for the class}}{\text{ }\mathbf{A}} \\
\\
\text{\textrm{for the class}}{\text{ }\mathbf{B}} \\
\\
\text{\textrm{for the class}}{\text{ }\mathbf{C}}%
\end{array}%
\right. \label{16}$$
For all the classes one has $R+T=1$ as should be expected for a conserved quantity. The class C presents $R>1$, the alluded Klein's paradox, implying that more particles are reflected from the potential barrier than those incoming. Contrary to the assertion of Ghose et al. [gho4]{}, Klein's paradox there exists for bosons in the DKP theory. It must be so because, as seen before, the potential stimulates the production of antiparticles at $%
z=0$. Due to the charge conservation there is, in fact, the creation of particle-antiparticle pairs. Since the potential in $z>0$ is repulsive for particles they are necessarily reflected. From the previous discussion related to the classes B and C, one can realize that the threshold energy for the pair production is given by $eV_{0}=2m$. The propagation of antiparticles inside the potential barrier can be interpreted as due to the fact that each antiparticle is under the influence of an effective potential given by $-eV_{0}$. In this way, each antiparticle has an available energy (rest energy plus kinetic energy) given by $eV_{0}-E$, accordingly one concludes about the threshold energy. One can also say that there is an ascending step for particles and a descending step for antiparticles.
Note that the currents $J^{\mu }$ and $S^{\mu }$ are simply related by$$S^{\mu }=\frac{E-eA_{0}}{m}\,J^{\mu } \label{rel}$$in all the classes of solutions. In this manner, the conservation law $\partial _{\mu }J^{\mu }=0$ is not compatible with $\partial _{\mu
}S^{\mu }=0$, at least for the case under investigation where $\partial
S^{0}/\partial t=0$ but $S^{3}$ is not uniform. In order to understand the behaviour of $S^{\mu }$ let us recall that the DKP equation can be recast into the Hamiltonian form [@now], [@pim1] $$i\frac{\partial \psi }{\partial t}=H\psi \label{eqham}$$where $$H=i\left[ \beta ^{i},\beta ^{0}\right] D_{i}+eA_{0}+m\beta ^{0}+i\frac{e}{2m}%
F_{\mu \nu }\left( \beta ^{\mu }\beta ^{0}\beta ^{\nu }+\beta ^{\mu }g^{0\nu
}\right) \label{ham}$$At this point is worthwhile to mention that $H$ is not Hermitian, in opposition what was adverted in [@now], since $$\left( iF_{0i}\beta ^{i}\left( \beta ^{0}\right) ^{2}\right) ^{\dagger
}=-\left( iF_{0i}\beta ^{i}\left( \beta ^{0}\right) ^{2}\right)
+iF_{0i}\beta ^{i} \label{her}$$There results that $$\partial _{\mu }S^{\mu }=i\psi ^{\dagger }\left( H-H^{\dagger }\right) \psi
=-\frac{e}{m}F_{0i}\,\psi ^{\dagger }\left[ \beta ^{i},\left( \beta
^{0}\right) ^{2}\right] \psi =\frac{e}{m}F_{0i}\,J^{i} \label{source1}$$This result clearly shows that the electromagnetic coupling induces a source term in the current $S^{\mu }$, as has already been shown in Ref. [@stru]. It is curious that the source term is due to the non-Hermitian piece of the anomalous term in (\[ham\]). Now, coming back to the square step potential (\[pot\]), one can write $$\partial _{\mu }S^{\mu }=-\frac{eV_{0}}{m}\,\delta \left( z\right)
J^{3}\left( z\right) \label{source2}$$in such a way that the jumping of $S^{3}$ at $z=0$ reads $$S^{3}(0_{+})-S^{3}(0_{-})=-\frac{eV_{0}}{m}J^{3}(0) \label{source3}$$a result in perfect agreement with (\[rel\]).
As for the current proposed by Datta [@dat], one can see that there is a spurious factor of 4 in the second term of $S^{\prime \mu }$ in (\[newS\]) due to the process of summation of four identical terms (four uses of the fundamental theorem of calculus involving $F^{\mu \nu }\,J_{\mu }u_{\nu }$) for considering the source term in (\[tensor\]) as an additional current term, not this merely but also $\partial _{\mu }S^{\prime \mu }\neq 0$. The new conserved current should be written as $$\tilde{S}^{\mu }=S^{\mu }+\frac{1}{4}\frac{e}{m}\int dx^{\mu }F^{\lambda \nu
}\,J_{\lambda }u_{\nu }$$so that $\tilde{S}^{\mu }$ without any question satisfies $\partial _{\mu }%
\tilde{S}^{\mu }=0$. It is worthwhile to note, though, that all components of the conserved current are nonvanishing for an arbitrary direction of motion. Furthermore, $\tilde{S}^{0}$ carries a temporal dependence even for the case of a time-independent DKP equation and it is infinite at the points of space where the potential suffers finite discontinuities. These are a few undesirable features which defy the candidature of $\tilde{S}^{\mu }$ as a current probability.
For short, the DKP equation with a square step potential is a test ground to refuse $S^{\mu }$ (and $\tilde{S}^{\mu }$) as a probability current as well as to show that Klein's paradox is alive and well in the DKP theory (we have talking about the spin-0 sector of the DKP theory but the state of affairs is not different for the spin-1 sector as one can see in Appendix A).
**Appendix A**
For the case of spin 1, the $\beta ^{\mu }$ matrices are [@ned2]$$\beta ^{0}=%
\begin{pmatrix}
0 & \overline{0} & \overline{0} & \overline{0} \\
\overline{0}^{T} & \mathbf{0} & \mathbf{I} & \mathbf{0} \\
\overline{0}^{T} & \mathbf{I} & \mathbf{0} & \mathbf{0} \\
\overline{0}^{T} & \mathbf{0} & \mathbf{0} & \mathbf{0}%
\end{pmatrix}%
\quad \quad \beta ^{i}=%
\begin{pmatrix}
0 & \overline{0} & e_{i} & \overline{0} \\
\overline{0}^{T} & \mathbf{0} & \mathbf{0} & -is_{i} \\
-e_{i}^{T} & \mathbf{0} & \mathbf{0} & \mathbf{0} \\
\overline{0}^{T} & -is_{i} & \mathbf{0} & \mathbf{0}%
\end{pmatrix}%$$where $s_{i}$ are the 3$\times $3 spin-1 matrices $\left(
s_{i}\right) _{jk}=-i\varepsilon _{ijk}$, $e_{i}$ are the 1$\times $3 matrices $\left( e_{i}\right) _{1j}=\delta _{ij}$ and $\overline{0}=%
\begin{pmatrix}
0 & 0 & 0%
\end{pmatrix}%
$, while** **$\mathbf{I}$ and $\mathbf{0}$** **designate the 3$%
\times $3 unit and zero matrices, respectively. With the spinor written as $%
\varphi ^{T}=\left( \varphi _{1},...,\varphi _{10}\right) $, and defining $%
\Psi ^{T}=\left( \varphi _{2},\varphi _{3},\varphi _{7}\right) $, $\Phi
^{T}=\left( \varphi _{5},\varphi _{6},\varphi _{4}\right) $and $\Theta
^{T}=\left( \varphi _{9},-\varphi _{8},\varphi _{1}\right) $ as done in Ref. [@che], the DKP equation (\[dkp ind temp\]) can be expressed in terms of the following equations$$\text{O}_{{\verb|KG|}}\Psi =0,\quad \Phi =\frac{E-eV_{0}}{m}\,\Psi ,\quad
\Theta =\frac{i}{m}\frac{d\Psi }{dz},\quad \varphi _{10}=0 \label{spin1a}$$Now the components of the four-currents are given by$$\begin{aligned}
J^{0} &=&2\Re\left( \varphi _{2}\varphi _{5}^{\ast }+\varphi _{3}\varphi
_{6}^{\ast }+\varphi _{4}\varphi _{7}^{\ast }\right) ,\quad
J^{3}=-2\Re\left( \varphi _{1}\varphi _{7}^{\ast }+\varphi _{2}\varphi
_{9}^{\ast }-\varphi _{3}\varphi _{8}^{\ast }\right) \notag \\
&& \label{spin1b} \\
S^{0} &=&\sum_{i=1}^{9}|\varphi _{i}|^{2},\quad S^{3}=-2\Re\left( \varphi
_{1}\varphi _{4}^{\ast }+\varphi _{5}\varphi _{9}^{\ast }-\varphi
_{6}\varphi _{8}^{\ast }\right) \notag\end{aligned}$$A discontinuous potential makes $\varphi _{4}$, $\varphi _{5}$ and $\varphi _{6}$ discontinuous, and as an immediate consequence $J^{0}$, $%
S^{0} $ and $S^{3}$ are also discontinuous. The plane wave solutions for the potential given by (\[pot\]) in the region $z<0$ can be written as$$\varphi \left( z\right) =\varphi _{+}e^{+ikz}+\varphi _{-}e^{-ikz}
\label{spin1c}$$where$$\varphi _{\pm }^{T}=\left( \mp \frac{k}{m}\gamma _{\pm },\alpha _{\pm
},\beta _{\pm },\frac{E}{m}\gamma _{\pm },\frac{E}{m}\alpha _{\pm },\frac{E}{%
m}\beta _{\pm },\gamma _{\pm },\pm \frac{k}{m}\beta _{\pm },\mp \frac{k}{m}%
\alpha _{\pm },0\right) \label{spin1d}$$and $\alpha _{\pm }$, $\beta _{\pm },$and $\gamma _{\pm }$ are arbitrary amplitudes. Defining$$a_{\pm }=\sqrt{2\left( |\alpha _{\pm }|^{2}+|\beta _{\pm }|^{2}+|\gamma
_{\pm }|^{2}\right) }$$it follows that the components of the current can be written in the same form as (\[j1\]) and (\[j2\]). A similar procedure for the region $z>0$ allows one to obtain the results (\[12\])-(\[16\]). Even though $\varphi _{2}$, $\varphi _{3}$ and $\varphi _{7}$ obey the Klein-Gordon equation there is no reason why they have the same amplitudes, as assumed in Ref. [@che]. As a matter of fact, a nontrivial spinor with only three nonvanishing components would be possible.
**Acknowledgments**
This work was supported in part by means of funds provided by CAPES, CNPq and FAPESP. The authors would like to thank Professor P. Holland for drawing attention to Ref. [@stru] after submission of our Letter to Physics Letters A and to an anonymous referee for Ref. [@dat].
[99]{} P. Ghose, D. Home, M.N.S. Roy, Phys. Lett. A 183 (1993) 267.
P. Ghose, Phys. Lett. A 191 (1994) 362;
P. Ghose, Found. Phys. 26 (1996) 1441.
M. Nowakowski, Phys. Lett. A 244 (1998) 329.
P. Ghose, M.K. Samal, Phys. Rev. E 64 (2001) 036620.
P. Ghose et al., Phys. Lett. A 290 (2001) 205;
C.A. Bonin et al., quant-ph/0608002;
H. Nikolić, hep-th/0702060.
P. Ghose, M.K. Samal, A. Datta, Phys. Lett. A 315 (2003) 23.
L. Chetouani et al., Int. J. Theor. Phys. 43 (2004) 1147;
B. Boutabia-Chéraitia, T. Boudjedaa, Phys. Lett. A 338 (2005) 97;
M. Merad, Int. J. Theor. Phys. 46 (2007) 2105.
W. Struyve et al., Phys. Lett. A 322 (2004) 84.
A. Datta, High Spin Field Theories and Relativistic Quantum Mechanics of Bosons, in Bosons, Ferromagnetism and Crystal Growth Research, Horizons in World Physics, edited by E. Seifer, Vol. 257, Chap. 4, pp. 119-149, Nova Publishers, New. York, 2007.
Y. Nedjadi, R.C. Barret, J. Phys. G 19 (1993) 87.
Y. Nedjadi, R.C. Barret, J. Math. Phys. 35 (1994) 4517.
J.T. Lunardi et al., Phys. Lett. A 268 (2000) 165.
[^1]: [email protected]
[^2]: [email protected].
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'Treating differentials as independent algebraic units have a long history of use and abuse. It is generally considered problematic to treat the derivative as a fraction of differentials rather than as a holistic unit acting as a limit, though for practical reasons it is often done for the first derivative. However, using a revised notation for the second and higher derivatives will allow for the ability to treat differentials as independent units for a much larger number of cases.'
author:
- Jonathan Bartlett
- 'Asatur Zh. Khurshudyan'
bibliography:
- 'References.bib'
title: Extending the Algebraic Manipulability of Differentials
---
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
author:
- Jin Xu
title: |
Theory on Structure and Coloring\
of Maximal Planar Graphs (I):\
Relationship between Structure and\
Coloring
---
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: |
Given two rooted, ordered, and labeled trees $P$ and $T$ the tree inclusion problem is to determine if $P$ can be obtained from $T$ by deleting nodes in $T$. This problem has recently been recognized as an important query primitive in XML databases. Kilpeläinen and Mannila \[*SIAM J. Comput. 1995*\] presented the first polynomial time algorithm using quadratic time and space. Since then several improved results have been obtained for special cases when $P$ and $T$ have a small number of leaves or small depth. However, in the worst case these algorithms still use quadratic time and space. Let $n_S$, $l_S$, and $d_S$ denote the number of nodes, the number of leaves, and the depth of a tree $S \in \{P, T\}$. In this paper we show that the tree inclusion problem can be solved in space $O(n_T)$ and time: $$O\left( \min\left\{
\begin{array}{l}
l_Pn_T \\
l_Pl_T\log \log n_T + n_T \\
\frac{n_Pn_T}{\log n_T} + n_{T}\log n_{T}
\end{array}\right\}
\right)$$
This improves or matches the best known time complexities while using only linear space instead of quadratic. This is particularly important in practical applications, such as XML databases, where the space is likely to be a bottleneck.
author:
- 'Philip Bille[^1]'
- 'Inge Li G[ø]{}rtz[^2]'
bibliography:
- 'pub.bib'
title: 'The Tree Inclusion Problem: In Linear Space and Faster[^3]'
---
Introduction
============
Let $T$ be a rooted tree. We say that $T$ is *labeled* if each node is assigned a character from an alphabet $\Sigma$ and we say that $T$ is *ordered* if a left-to-right order among siblings in $T$ is given. All trees in this paper are rooted, ordered, and labeled. A tree $P$ is *included* in $T$, denoted $P\sqsubseteq T$, if $P$ can be obtained from $T$ by deleting nodes of $T$. Deleting a node $v$ in $T$ means making the children of $v$ children of the parent of $v$ and then removing $v$. The children are inserted in the place of $v$ in the left-to-right order among the siblings of $v$. The *tree inclusion problem* is to determine if $P$ can be included in $T$ and if so report all subtrees of $T$ that include $P$.
&&$Q$&&&&&$D$&\[name=cat\] catalog\
&&\[name=book\] book &&&&&&\[name=book2\] book & \[name=book3\] & \[name=book4\]\
& \[name=author\] author &&\[name=chapter\] chapter&& \[name=author2\] author && \[name=chapter2\] chapter && \[name=chapter3\] chapter\
& \[name=john\] john && \[name=xml\] XML && \[name=name\] name && \[name=title\] title & \[name=section\] section & \[name=title2\] title\
&&&&& \[name=john2\] john && \[name=DB\] databases & \[name=xml2\] XML & \[name=queries\] queries\
&& (a) &&&&&& (b)\
&&&&&&&&\[name=cat\] catalog\
&&\[name=book\] book &&&&&&\[name=book2\] book & \[name=book3\] & \[name=book4\]\
& \[name=author\] author &&\[name=chapter\] chapter&& \[name=author2\] author && \[name=chapter2\] chapter && \[name=chapter3\] chapter\
& \[name=john\] john && \[name=xml\] XML && \[name=name\] name && \[name=title\] title & \[name=section\] section & \[name=title2\] title\
&&&&& \[name=john2\] john && \[name=DB\] databases & \[name=xml2\] XML & \[name=queries\] queries\
&&&&&&(c)&&&&\
Recently, the problem has been recognized as an important query primitive for XML data and has received considerable attention, see e.g., [@SM2002; @YLH2003; @YLH2004; @ZADR03; @SN2000; @TRS2002]. The key idea is that an XML document can be viewed as a tree and queries on the document correspond to a tree inclusion problem. As an example consider Figure \[inclusionexample\]. Suppose that we want to maintain a catalog of books for a bookstore. A fragment of the tree, denoted $D$, corresponding to the catalog is shown in (b). In addition to supporting full-text queries, such as find all documents containing the word “John”, we can also utilize the tree structure of the catalog to ask more specific queries, such as “find all books written by John with a chapter that has something to do with XML”. We can model this query by constructing the tree, denoted $Q$, shown in (a) and solve the tree inclusion problem: is $Q \sqsubseteq D$? The answer is yes and a possible way to include $Q$ in $D$ is indicated by the dashed lines in (c). If we delete all the nodes in $D$ not touched by dashed lines the trees $Q$ and $D$ become isomorphic. Such a mapping of the nodes from $Q$ to $D$ given by the dashed lines is called an *embedding* (formally defined in Section \[sec:recursion\]). We note that the ordering of the XML document, and hence the left-to-right order of siblings, is important in many cases. For instance, in the above example, the relative order of contents of the chapters is most likely important. Also, in biological databases, order is of critical importance. Consequently, standard XML query languages, such as XPath [@CD99] and XQuery [@BCFF01], require the output of queries to be ordered.
The tree inclusion problem was initially introduced by Knuth [@Knuth1969 exercise 2.3.2-22] who gave a sufficient condition for testing inclusion. Motivated by applications in structured databases [@KM93; @MR90] Kilpeläinen and Mannila [@KM1995] presented the first polynomial time algorithm using $O(n_Pn_T)$ time and space, where $n_P$ and $n_T$ is the number of nodes in $P$ and $T$, respectively. During the last decade several improvements of the original algorithm of [@KM1995] have been suggested [@Kilpelainen1992; @AS2001; @Richter1997a; @Chen1998]. The previously best known bound is due to Chen [@Chen1998] who presented an algorithm using $O(l_Pn_T)$ time and $O(l_P \cdot \min\{d_T, l_T\})$ space. Here, $l_S$ and $d_S$ denote the number of leaves and the depth of a tree $S$, respectively. This algorithm is based on an algorithm of Kilpeläinen [@Kilpelainen1992]. Note that the time and space is still $\Theta(n_Pn_T)$ for worst-case input trees.
In this paper we present three algorithms which combined improve all of the previously known time and space bounds. To avoid trivial cases we always assume that $1 \leq n_P \leq n_T$. We show the following theorem:
\[thm:main\] For trees $P$ and $T$ the tree inclusion problem can be solved in $O(n_T)$ space with the following running time: $$O\left( \min\left\{
\begin{array}{l}
l_Pn_T \\
l_Pl_T\log \log n_T + n_T \\
\frac{n_Pn_T}{\log n_T} + n_{T}\log n_{T}
\end{array}\right\}
\right)$$
Hence, when $P$ has few leaves we obtain a fast algorithm and even faster if both $P$ and $T$ have few leaves. When both trees have many leaves and $n_{P} = \Omega (\log^{2} n_{T})$, we instead improve the previous quadratic time bound by a logarithmic factor. Most importantly, the space used is linear. In the context of XML databases this will likely make it possible to query larger trees and speed up the query time since more of the computation can be kept in main memory.
The extended abstract of this paper [@BG05] contained an error. The algorithms in the paper [@BG05] did not use linear space. The problem was due to a recursive traversal of $P$ which stored too many sets of nodes leading to a worst-case space complexity of $\Omega(d_P
l_T)$. In this paper we fix this problem by recursively visiting the nodes such that the child with the largest number of descendant leaves is visited first, and by showing that the size of the resulting stored node sets exponentially decrease. With these ideas we show that all of our algorithms use $O(n_T)$ space. Additionally, our improved analysis of the sizes of the stored node sets also leads to an improvement in the running time of the algorithm in the second case above. In the previous paper the running time was $O(n_pl_T\log\log n_T+ n_T)$.
Techniques
----------
Most of the previous algorithms, including the best one [@Chen1998], are essentially based on a simple dynamic programming approach from the original algorithm of [@KM1995]. The main idea behind this algorithm is the following: Let $v$ be a node in $P$ with children $v_1, \ldots, v_i$ and let $w$ be a node in $T$. Consider the subtrees rooted at $v$ and $w$, denoted by $P(v)$ and $T(w)$. To decide if $P(v)$ can be included in $T(w)$ we try to find a sequence $w_{1}, \ldots, w_{i}$ of left-to-right ordered descendants of $w$ such that $P(v_k) \sqsubseteq T(w_{k})$ for all $k$, $1 \leq k \leq i$. The sequence is computed greedily from left-to-right in $T(w)$ effectively finding the *leftmost inclusion* of $P(v)$ in $T(w)$. Applying this approach in a bottom-up fashion we can determine, if $P(v) \sqsubseteq T(w)$, for all pairs of nodes $v$ in $P$ and $w$ in $T$.
In this paper we take a different approach. The main idea is to construct a data structure on $T$ supporting a small number of procedures, called the *set procedures*, on subsets of nodes of $T$. We show that any such data structure implies an algorithm for the tree inclusion problem. We consider various implementations of this data structure which all use linear space. The first simple implementation gives an algorithm with $O(l_Pn_T)$ running time. As it turns out, the running time depends on a well-studied problem known as the *tree color problem*. We show a direct connection between a data structure for the tree color problem and the tree inclusion problem. Plugging in a data structure of Dietz [@Die89] we obtain an algorithm with $O(l_Pl_T\log \log n_T + n_T)$ running time.
Based on the simple algorithms above we show how to improve the worst-case running time of the set procedures by a logarithmic factor. The general idea used to achieve this is to divide $T$ into small trees called *clusters* of logarithmic size which overlap with other clusters in at most $2$ nodes. Each cluster is represented by a constant number of nodes in a *macro tree*. The nodes in the macro tree are then connected according to the overlap of the cluster they represent. We show how to efficiently preprocess the clusters and the macro tree such that the set procedures use constant time for each cluster. Hence, the worst-case quadratic running time is improved by a logarithmic factor.
Our algorithms recursively traverse $P$ top-down. For each node $v \in V(P)$ we compute a set of nodes representing all of the subtrees in $T$ that include $P(v)$. To avoid storing too many of these node sets the traversal of $P$ visits the child with the largest number of descendant leaves first. For the first two algorithms this immediately implies a space complexity of $O(l_T \log l_P)$, however, by carefully analyzing the sizes of stored node sets we are able to show that they decrease exponentially leading to the linear space bound. In the last algorithm the node sets are compactly encoded in $O(n_T/\log n_T)$ space and therefore our recursive traversal alone implies a space bound of $O(n_T/\log n_T
\cdot \log l_P) = O(n_T)$.
Throughout the paper we assume a unit-cost RAM model of computation with word size $\Theta(\log n_T)$ and a standard instruction set including bitwise boolean operations, shifts, addition, and multiplication. All space complexities refer to the number of words used by the algorithm.
Related Work
------------
For some applications considering *unordered* trees is more natural. However, in [@MT1992; @KM1995] this problem was proved to be NP-complete. The tree inclusion problem is closely related to the *tree pattern matching problem* [@CO1982; @Kosaraju1989; @DGM1990; @CHI1999]. The goal is here to find an injective mapping $f$ from the nodes of $P$ to the nodes of $T$ such that for every node $v$ in $P$ the $i$th child of $v$ is mapped to the $i$th child of $f(v)$. The tree pattern matching problem can be solved in $(n_P+n_T) \log^{O(1)} (n_P+n_T)$ time. Another similar problem is the *subtree isomorphism* problem [@Chung1987; @ST1999], which is to determine if $T$ has a subgraph isomorphic to $P$. The subtree isomorphism problem can be solved efficiently for ordered and unordered trees. The best algorithms for this problem use $O(\frac{n_P^{1.5}n_T}{\log
n_P} + n_T)$ time for unordered trees and $O(\frac{n_Pn_T}{\log n_P} + n_T)$ time for ordered trees [@Chung1987; @ST1999]. Both use $O(n_Pn_T)$ space. The tree inclusion problem can be considered a special case of the *tree edit distance problem* [@Tai1979; @ZS1989; @Klein1998; @DMRW07]. Here one wants to find the minimum sequence of insert, delete, and relabel operations needed to transform $P$ into $T$. Currently the best algorithm for this problem uses $O(n_T n_P^2 (1 + \log \frac{n_T}{n_P}))$ time [@DMRW07]. For more details and references see the survey [@Bille2005].
Outline
-------
In Section \[def\] we give notation and definitions used throughout the paper. In Section \[sec:recursion\] a common framework for our tree inclusion algorithms is given. Section \[simple\] presents two simple algorithms and then, based on these results, we show how to get a faster algorithm in Section \[micromacro\].
Notation and Definitions {#def}
========================
In this section we define the notation and definitions we will use throughout the paper. For a graph $G$ we denote the set of nodes and edges by $V(G)$ and $E(G)$, respectively. Let $T$ be a rooted tree. The root of $T$ is denoted by ${\ensuremath{\mathrm{root}}}(T)$. The *size* of $T$, denoted by $n_T$, is $|V(T)|$. The *depth* of a node $v\in V(T)$, ${\ensuremath{\mathrm{depth}}}(v)$, is the number of edges on the path from $v$ to ${\ensuremath{\mathrm{root}}}(T)$ and the depth of $T$, denoted $d_T$, is the maximum depth of any node in $T$. The parent of $v$ is denoted ${\ensuremath{\mathrm{parent}}}(v)$ and the set of children of $v$ is denoted ${\ensuremath{\mathrm{child}}}(v)$. We define ${\ensuremath{\mathrm{parent}}}({\ensuremath{\mathrm{root}}}(T))=\bot$, where $\bot \not\in V(T)$ is a special *null node*. A node with no children is a leaf and otherwise an internal node. The set of leaves of $T$ is denoted $L(T)$ and we define $l_T = |L(T)|$. We say that $T$ is *labeled* if each node $v$ is a assigned a character, denoted ${\ensuremath{\mathrm{label}}}(v)$, from an alphabet $\Sigma$ and we say that $T$ is *ordered* if a left-to-right order among siblings in $T$ is given. Note that we *do not* require that the size of the alphabet is bounded by a constant. All trees in this paper are rooted, ordered, and labeled.
#### Ancestors and Descendants
Let $T(v)$ denote the subtree of $T$ rooted at a node $v \in
V(T)$. If $w\in V(T(v))$ then $v$ is an ancestor of $w$, denoted $v \preceq w$, and if $w\in V(T(v))\backslash \{v\}$ then $v$ is a proper ancestor of $w$, denoted $v \prec w$. If $v$ is a (proper) ancestor of $w$ then $w$ is a (proper) descendant of $v$. A node $z$ is a common ancestor of $v$ and $w$ if it is an ancestor of both $v$ and $w$. The nearest common ancestor of $v$ and $w$, ${\ensuremath{\mathrm{nca}}}(v,w)$, is the common ancestor of $v$ and $w$ of greatest depth. The *first ancestor of $w$ labeled $\alpha$*, denoted ${\ensuremath{\mathrm{fl}}}(w,\alpha)$, is the node $v$ such that $v \preceq w$, ${\ensuremath{\mathrm{label}}}(v)
= \alpha$, and no node on the path between $v$ and $w$ is labeled $\alpha$. If no such node exists then ${\ensuremath{\mathrm{fl}}}(w,\alpha) = \bot$.
#### Traversals and Orderings
Let $T$ be a tree with root $v$ and let $v_1, \ldots ,v_k$ be the children of $v$ from left-to-right. The *preorder traversal* of $T$ is obtained by visiting $v$ and then recursively visiting $T(v_i)$, $1 \leq i \leq k$, in order. Similarly, the *postorder traversal* is obtained by first visiting $T(v_i)$, $1 \leq i \leq k$, in order and then $v$. The *preorder number* and *postorder number* of a node $w \in T(v)$, denoted by ${\ensuremath{\mathrm{pre}}}(w)$ and ${\ensuremath{\mathrm{post}}}(w)$, is the number of nodes preceding $w$ in the preorder and postorder traversal of $T$, respectively. The nodes to the *left* of $w$ in $T$ is the set of nodes $u \in V(T)$ such that ${\ensuremath{\mathrm{pre}}}(u) < {\ensuremath{\mathrm{pre}}}(w)$ and ${\ensuremath{\mathrm{post}}}(u) < {\ensuremath{\mathrm{post}}}(w)$. If $u$ is to the left of $w$, denoted by $u
\lhd w$, then $w$ is to the *right* of $u$. If $u \lhd w$ or $u
\preceq w$ or $w \prec u$ we write $u \unlhd w$. The null node $\bot$ is not in the ordering, i.e., $\bot \ntriangleleft v$ for all nodes $v$.
#### Minimum Ordered Pairs
A set of nodes $X \subseteq V(T)$ is *deep* if no node in $X$ is a proper ancestor of another node in $X$. For $k$ deep sets of nodes $X_1, \ldots, X_k$ let $\Phi(X_1,\ldots,X_k) \subseteq (X_1
\times \cdots \times X_k)$, be the set of tuples such that $(x_1,
\ldots, x_k) \in \Phi(X_1,\ldots,X_k)$ iff $x_1 \lhd \cdots \lhd
x_k$. If $(x_1, \ldots, x_k) \in \Phi(X_1,\ldots,X_k) $ and there is no $(x_1', \ldots, x_k') \in \Phi(X_1,\ldots,X_k) $, where either $x_1 \lhd x_1' \lhd x_k' \unlhd x_k$ or $x_1 \unlhd x_1'
\lhd x_k' \lhd x_k$ then the pair $(x_1, x_k)$ is a *minimum ordered pair*. Intuitively, $(x_1, x_k)$ is a closest pair of nodes from $X_1$ and $X_k$ in the left-to-right order for which we can find $x_2, \ldots, x_{k-1}$ such that $x_1 \lhd \cdots
\lhd x_k$. The set of minimum ordered pairs for $X_1, \ldots,
X_k$ is denoted by ${\ensuremath{\mathrm{mop}}}(X_1, \ldots, X_k)$. Figure \[fig:mopexample\] illustrates these concepts on a small example.
&&&& (0,.3)\
& (0,.3) &&& (-.3,.3) && (0,.3)\
(-.3,0)[$v_1$]{}(0,-.3) & (-.3,0)[$v_2$]{}(0,-.3) & (-.2,-.3) & (0,-.3) & & (0,-.3)(-.3,0)[$v_5$]{} & (-.2,-.3)(-.3,0)[$v_6$]{} & (0,-.3)(-.3,0)[$v_7$]{}&&&\
&& (0,-.3) & (0,-.3)(-.3,0)[$v_3$]{} & (0,-.3)(-.3,0)[$v_4$]{}&& (0,-.3)\
&&&&(0,0)[(a)]{}\
(.5,0)[=$S_1$]{} && (.5,0)[=$S_2$]{} && (.5,0)[=$S_3$]{} && (.5,0)[=$S_4$]{} &&
&&&& (0,.3)\
& (0,.3) &&& (-.3,.3) && (0,.3)\
(-.3,0)[$v_1$]{}(0,-.3) & (-.3,0)[$v_2$]{}(0,-.3) & (-.2,-.3) & (0,-.3)(-.3,0)[$v_5$]{} && (0,-.3) & (-.2,-.3)(-.3,0)[$v_8$]{} & (0,-.3)(-.3,0)[$v_9$]{}\
&& (0,-.3)(-.3,0)[$v_3$]{} & (0,-.3)(-.3,0)[$v_4$]{} & (0,-.3)(-.3,0)[$v_6$]{}&& (0,-.3)(-.3,0)[$v_7$]{}\
&&&&(0,0)[(b)]{}\
For any set of pairs $Y$, let ${\ensuremath{\mathop{Y|}_{1}}}$ and ${\ensuremath{\mathop{Y|}_{2}}}$ denote the *projection* of $Y$ to the first and second coordinate, that is, if $(y_1,y_2) \in Y$ then $y_1 \in
{\ensuremath{\mathop{Y|}_{1}}}$ and $y_2 \in {\ensuremath{\mathop{Y|}_{2}}}$. We say that $Y$ is *deep* if ${\ensuremath{\mathop{Y|}_{1}}}$ and ${\ensuremath{\mathop{Y|}_{2}}}$ are deep. The following lemma shows that given deep sets $X_1, \ldots, X_k$ we can compute ${\ensuremath{\mathrm{mop}}}(X_1,\ldots,X_k)$ iteratively by first computing ${\ensuremath{\mathrm{mop}}}(X_1,X_2)$ and then ${\ensuremath{\mathrm{mop}}}({\ensuremath{\mathop{{\ensuremath{\mathrm{mop}}}(X_1,X_2)|}_{2}}},X_3)$ and so on.
\[lem:nnsmaller2\] For any deep sets of nodes $X_1, \ldots, X_k$, $k>2$, we have, $(x_1,x_k) \in {\ensuremath{\mathrm{mop}}}(X_1, \ldots, X_k)$ iff there exists a node $x_{k-1}$ such that $(x_1,x_{k-1}) \in {\ensuremath{\mathrm{mop}}}(X_1, \ldots,
X_{k-1})$ and $(x_{k-1},x_k) \in {\ensuremath{\mathrm{mop}}}({\ensuremath{\mathop{{\ensuremath{\mathrm{mop}}}(X_1,
\ldots, X_{k-1})|}_{2}}}, X_k)$.
We start by showing that if $(x_1,x_k) \in {\ensuremath{\mathrm{mop}}}(X_1, \ldots,
X_k)$ then there exists a node $x_{k-1}$ such that $(x_1,x_{k-1}) \in
{\ensuremath{\mathrm{mop}}}(X_1, \ldots, X_{k-1})$ and $(x_{k-1},x_k) \in
{\ensuremath{\mathrm{mop}}}({\ensuremath{\mathop{{\ensuremath{\mathrm{mop}}}(X_1, \ldots, X_{k-1})|}_{2}}}, X_k)$.
First note that $(z_1,\ldots,z_k) \in
\Phi(X_1,\ldots,X_k)$ implies $(z_1,\ldots,z_{k-1}) \in
\Phi(X_1,\ldots,X_{k-1})$. Since $(x_1,x_k) \in
{\ensuremath{\mathrm{mop}}}(X_1,\ldots,X_k)$ there must be a minimum node $x_{k-1}$ such that the tuple $(x_1,\ldots,x_{k-1})$ is in $\Phi(X_1,\ldots,X_{k-1})$. We have $(x_1, x_{k-1}) \in {\ensuremath{\mathrm{mop}}}(X_1,\ldots,X_{k-1})$. We need to show that $(x_{k-1},x_k) \in {\ensuremath{\mathrm{mop}}}({\ensuremath{\mathop{{\ensuremath{\mathrm{mop}}}(X_1, \ldots, X_{k-1})|}_{2}}}, X_k)$. Since $(x_1,x_k) \in {\ensuremath{\mathrm{mop}}}(X_1, \ldots,
X_k)$ there exists no node $z \in X_k$ such that $x_{k-1} \lhd z
\lhd x_k$. If such a $z$Êexisted we would have $(x_1,\ldots, x_{k-1},z)\in \Phi(X_1,\ldots,X_k)$, contradicting that $(x_1,x_k)\in {\ensuremath{\mathrm{mop}}}(X_1, \ldots, X_k)$. Assume there exists a node $z \in {\ensuremath{\mathop{{\ensuremath{\mathrm{mop}}}(X_1, \ldots,
X_{k-1})|}_{2}}}$ such that $x_{k-1} \lhd z
\lhd x_k$. Since $(x_1,x_{k-1})\in {\ensuremath{\mathrm{mop}}}(X_1,\ldots,X_{k-1})$ this implies that there is a node $z' \rhd x_1$ such that $(z',z) \in
{\ensuremath{\mathrm{mop}}}(X_1,\ldots,X_{k-1})$. But this implies that the tuple $(z',\ldots,z,x_k)$ is in $\Phi(X_1,\ldots,X_k)$ contradicting that $(x_1,x_k) \in {\ensuremath{\mathrm{mop}}}(X_1,\ldots,X_k)$.
We will now show that if there exists a node $x_{k-1}$ such that $(x_1,x_{k-1}) \in
{\ensuremath{\mathrm{mop}}}(X_1, \ldots, X_{k-1})$ and $(x_{k-1},x_k) \in
{\ensuremath{\mathrm{mop}}}({\ensuremath{\mathop{{\ensuremath{\mathrm{mop}}}(X_1, \ldots, X_{k-1})|}_{2}}}, X_k)$ then the pair $(x_1,x_k) \in {\ensuremath{\mathrm{mop}}}(X_1, \ldots, X_k)$. Clearly, there exists a tuple $(x_1,\ldots,x_{k-1},x_k)\in \Phi(X_1,\ldots,X_k)$. Assume that there exists a tuple $(z_1,\ldots,z_k)\in
\Phi(X_1,\ldots,X_k)$ such that $x_1 \lhd z_1 \lhd z_k \unlhd
x_k$. Among the tuples satisfying these constraints let $(y_1, \ldots, y_{k-1}, y_k)$Êbe the one with maximum $y_1$, minimum $y_{k-1}$, and maximum $y_k$. It follows that $(y_1,y_{k-1})\in {\ensuremath{\mathrm{mop}}}(X_1,\ldots,X_{k-1})$. Since $(x_1,x_{k-1}) \in {\ensuremath{\mathrm{mop}}}(X_1,\ldots,X_{k-1})$ we must have $x_{k-1}\lhd y_{k-1}$. But this contradicts $(x_{k-1},x_k) \in {\ensuremath{\mathrm{mop}}}({\ensuremath{\mathop{{\ensuremath{\mathrm{mop}}}(X_1, \ldots, X_{k-1})|}_{2}}}, X_k)$, since node $y_{k-1}\in {\ensuremath{\mathop{{\ensuremath{\mathrm{mop}}}(X_1,\ldots,X_{k-1})|}_{2}}}$.
Assume that there exists a tuple $(z_1,\ldots,z_k)\in
\Phi(X_1,\ldots,X_k)$ such that $x_1 \unlhd z_1 \lhd z_k \lhd
x_k$. Since $(x_1,x_{k-1})\in {\ensuremath{\mathrm{mop}}}(X_1,\ldots,X_{k-1})$ we have $x_{k-1} \unlhd z_{k-1}$ and thus $x_{k-1} \lhd z_k \lhd x_{k}$ contradicting $(x_{k-1},x_k) \in {\ensuremath{\mathrm{mop}}}({\ensuremath{\mathop{{\ensuremath{\mathrm{mop}}}(X_1, \ldots, X_{k-1})|}_{2}}}, X_k)$.
The following lemma is the reverse of the previous lemma and shows that given deep sets $X_1, \ldots, X_k$ we also can compute ${\ensuremath{\mathrm{mop}}}(X_1,\ldots,X_k)$ iteratively from right to left. The proof is similar to the proof of Lemma \[lem:nnsmaller2\].
\[lem:nnsmallerleft\] For any deep sets of nodes $X_1, \ldots, X_k$, $k > 2$, we have, $(x_1,x_k) \in {\ensuremath{\mathrm{mop}}}(X_1, \ldots, X_k)$ iff there exists a node $x_{2}$ such that $(x_2,x_{k}) \in {\ensuremath{\mathrm{mop}}}(X_2, \ldots,
X_{k})$ and $(x_1,x_2) \in {\ensuremath{\mathrm{mop}}}(X_1, {\ensuremath{\mathop{{\ensuremath{\mathrm{mop}}}(X_2,
\ldots, X_{k})|}_{1}}})$.
#### Heavy Leaf Path Decomposition
We construct a *heavy leaf path decomposition* of $P$ as follows. We classify each node as *heavy* or *light*. The root is light. For each internal node $v$Êwe pick a child $v_j$Êof $v$Êwith maximum $l_{P(v_j)}$ and classify it as heavy. The remaining children of $v$ are light. An edge to a light node is a *light edge*, and an edge to a heavy node is a *heavy edge*. The heavy child of a node $v$Êis denoted heavy($v$). Let ${\ensuremath{\mathrm{ldepth}}}(v)$ denote the number of light edges on the path from $v$Êto root($P$).
Note that a heavy leaf path decomposition is the same as the classical *heavy path decomposition* [@HT1984] except that the heavy child is defined as the child with largest number of the descendant leaves and not the child with the largest number of descendants. This distinction is essential for achieving the linear space bound of our algorithms. Note that heavy path decompositions have previously been used in algorithms for the related tree edit distance problem [@Klein1998].
For any tree $P$Êand node $v \in V(P)$, $$l_{P(v)}\leq \frac{l_P}{2^{{\ensuremath{\mathrm{ldepth}}}(v)}}\;.$$
By induction on ${\ensuremath{\mathrm{ldepth}}}(v)$. For ${\ensuremath{\mathrm{ldepth}}}(v)=0$ it is trivially true. Let ${\ensuremath{\mathrm{ldepth}}}(v)=\ell$. Assume wlog. that $v$Êis light. Let $w$Êbe the unique light ancestor of $v$ with ${\ensuremath{\mathrm{ldepth}}}(w)=\ell-1$. By the induction hypothesis $l_{P(w)}\leq l_P/2^{\ell-1}$.ÊNow $v$ has a sibling ${\ensuremath{\mathrm{heavy}}}({\ensuremath{\mathrm{parent}}}(v))$ and thus at most half of the leaves in $P({\ensuremath{\mathrm{parent}}}(v))$ can be in the subtree rooted at $v$. Therefore, $l_{P(v)}\leq l_{P(w)}/2 \leq l_P/2^{\ell}$.
\[cor:ldepth\] For any tree $P$Êand node $v \in V(P)$, ${\ensuremath{\mathrm{ldepth}}}(v) \leq \log l_P$.
#### Notation
When we want to specify which tree we mean in the above relations we add a subscript. For instance, $v \prec_{T} w$ indicates that $v$ is an ancestor of $w$ *in $T$*.
Computing Deep Embeddings {#sec:recursion}
=========================
In this section we present a general framework for answering tree inclusion queries. As in [@KM1995] we solve the equivalent *tree embedding problem*. Let $P$ and $T$ be rooted labeled trees. An *embedding* of $P$ in $T$ is an injective function $f : V(P) \rightarrow V(T)$ such that for all nodes $v,u \in
V(P)$,
- ${\ensuremath{\mathrm{label}}}(v) = {\ensuremath{\mathrm{label}}}(f(v))$. (label preservation condition)
- $v \prec u$ iff $f(v) \prec f(u)$. (ancestor condition)
- $v \lhd u$ iff $f(v) \lhd f(u)$. (order condition)
An example of an embedding is given in Figure \[inclusionexample\](c).
For any trees $P$ and $T$, $P \sqsubseteq T$ iff there exists an embedding of $P$ in $T$.
We say that the embedding $f$ is *deep* if there is no embedding $g$ such that $f({\ensuremath{\mathrm{root}}}(P)) \prec g({\ensuremath{\mathrm{root}}}(P))$. The *deep occurrences* of $P$ in $T$, denoted ${\ensuremath{\mathrm{emb}}}(P, T)$ is the set of nodes, $${\ensuremath{\mathrm{emb}}}(P,T) = \{f({\ensuremath{\mathrm{root}}}(P)) \mid \text{$f$ is a deep embedding of $P$ in $T$}\}.$$ By definition the set of ancestors of nodes in ${\ensuremath{\mathrm{emb}}}(P,T)$ is exactly the set of nodes $\{u \mid P \sqsubseteq T(u)\}$. Hence, to solve the tree inclusion problem it is sufficient to compute ${\ensuremath{\mathrm{emb}}}(P,T)$ and then, using additional $O(n_T)$ time, report all ancestors of this set. We note that Kilpeläinen and Mannila [@KM1995] used the similar concept of *left embeddings* in their algorithms. A left embedding of $P$ in $T$ is an embedding such that the root of $P$ is mapped to the node in $T$ with the smallest postorder number, i.e., the deepest node among the nodes furthest to the left. Our definition of ${\ensuremath{\mathrm{emb}}}(P,T)$ only requires that the root is mapped to a deepest node. In the following we show how to compute deep embeddings. The key idea is to build a data structure for $T$ allowing a fast implementation of the following procedures. For all $X \subseteq
V(T)$, $Y \subseteq V(T)\times V(T)$, and $\alpha \in \Sigma$ define:
Return the set $\{{\ensuremath{\mathrm{parent}}}(x) \mid x \in X\}$.
Return the set $\{{\ensuremath{\mathrm{nca}}}(y_1,y_2) \mid (y_1,y_2) \in Y\}$.
Return the set $\{x \in X\mid \text{there is no }
z \in X \text{ such that } x \prec z\}$.
Return the set of pairs $R$ such that for any pair $(y_1,y_2)
\in Y$, $(y_1,x) \in R$ iff $(y_2, x) \in
{\ensuremath{\mathrm{mop}}}({\ensuremath{\mathop{Y|}_{2}}}, X)$.
Return the set of pairs $R$ such that for any pair $(y_1,y_2)
\in Y$, $(x,y_2) \in R$ iff $(x,y_1) \in
{\ensuremath{\mathrm{mop}}}(X,{\ensuremath{\mathop{Y|}_{1}}})$.
Return the set ${\ensuremath{\textsc{Deep}}}(\{{\ensuremath{\mathrm{fl}}}(x, \alpha) \mid x \in X\})$.
Collectively we call these procedures the *set procedures*. The procedures ${\ensuremath{\textsc{Parent}}}$ and ${\ensuremath{\textsc{Nca}}}$ are self-explanatory. ${\ensuremath{\textsc{Deep}}}(X)$ returns the set of all nodes in $X$ that have no descendants in $X$. Hence, the returned set is always deep. ${\ensuremath{\textsc{MopRight}}}$ and ${\ensuremath{\textsc{MopLeft}}}$ are used to iteratively compute minimum ordered pairs. ${\ensuremath{\textsc{Fl}}}(X, \alpha)$ returns the deep set of first ancestors with label $\alpha$ of all nodes in $X$. If we want to specify that a procedure applies to a certain tree $T$ we add the subscript $T$. With the set procedures we can compute deep embeddings. The following procedure ${\ensuremath{\textsc{Emb}}}(v)$, $v \in V(P)$, recursively computes the set of deep occurrences of $P(v)$ in $T$. Figure \[fig:embexample\] illustrates how ${\ensuremath{\textsc{Emb}}}$ works on a small example.
Let $v_1, \ldots, v_k$ be the sequence of children of $v$ ordered from left to right. There are three cases:
1. $k=0$ ($v$ is a leaf). Compute $R := {\ensuremath{\textsc{Fl}}}(L(T),{\ensuremath{\mathrm{label}}}(v))$.
2. $k=1$. Recursively compute $R_1 := {\ensuremath{\textsc{Emb}}}(v_1)$.
Compute $R := {\ensuremath{\textsc{Fl}}}({\ensuremath{\textsc{Deep}}}({\ensuremath{\textsc{Parent}}}(R_1)),{\ensuremath{\mathrm{label}}}(v))$.
3. $k > 1$. Let $v_j$Êbe the heavy child of $v$.
Recursively compute $R_j := {\ensuremath{\textsc{Emb}}}(v_j)$ and set $U_j := \{(r,r) \mid r \in R_j\}$.
For $i := j+1$ to $k$, recursively compute $R_i := {\ensuremath{\textsc{Emb}}}(v_i)$ and $U_i := {\ensuremath{\textsc{MopRight}}}(U_{i-1}, R_i)$.
Set $U_j:=U_k$.
For $i := j-1$ to $1$, recursively compute $R_i := {\ensuremath{\textsc{Emb}}}(v_i)$ and $U_i := {\ensuremath{\textsc{MopLeft}}}(R_i,U_{i+1})$.
Finally, compute $R := {\ensuremath{\textsc{Fl}}}({\ensuremath{\textsc{Deep}}}({\ensuremath{\textsc{Nca}}}(U_1)), {\ensuremath{\mathrm{label}}}(v))$.
If $R = \emptyset$ stop and report that there is no deep embedding of $P(v)$ in $T$. Otherwise return $R$.
&&&(0,0)[$P$]{} &&&&&&&&&& (0,0)[$T$]{}\
&&& (.3,0)[$a$]{}(-.3,0)[$1$]{} &&&&&&&&&& (0,.3)[$a$]{}\
&& (.3,0)[$b$]{}(-.3,0)[$2$]{} && (.3,0)[$a$]{}(-.3,0)[$4$]{} &&&&&& (0,.3)[$b$]{} &&& (-.3,.3)[$b$]{} &&& (0,.3)[$a$]{}\
&& (.3,0)[$a$]{}(-.3,0)[$3$]{} && &&&&&& (0,-.3)[$a$]{} && (0,-.3)[$a$]{} && (-.3,-.3)[$b$]{} & (0,-.3)[$a$]{} && (0,-.3)[$b$]{}\
&&&&&&&&&&&&&&(0,-.3)[$a$]{}\
&&&(0,0)[(a)]{} &&&&&&&&&& (0,0)[(b)]{}\
\
&&& (0,.3)[$a$]{} &&&&&&&&&& (0,.3)[$a$]{}\
(0,.3)[$b$]{} &&& (-.3,.3)[$b$]{} &&& (0,.3)[$a$]{} &&&& (0,.3)[$b$]{} &&& (-.3,.3)[$b$]{} &&& (0,.3)[$a$]{}\
(0,-.3)[$a$]{} && (0,-.3)[$a$]{} && (-.3,-.3)[$b$]{} & (0,-.3)[$a$]{} && (0,-.3)[$b$]{} &&& (0,-.3)[$a$]{} && (0,-.3)[$a$]{} && (-.3,-.3)[$b$]{} & (0,-.3)[$a$]{} && (0,-.3)[$b$]{}\
&&&& (0,-.3)[$a$]{} &&&&&&&&&& (0,-.3)[$a$]{}\
&&&(0,0)[(c)]{} &&&&&&&&&& (0,0)[(d)]{}\
\
&&& (0,.3)[$a$]{} &&&&&&&&&& (0,.3)[$a$]{}\
(0,0)[(.9,0)(1.2,.3)]{}(0,.35)[$b$]{} &&& (-.3,.3)[$b$]{} &&& (0,.3)[$a$]{} &&&& (0,.3)[$b$]{} &&& (-.3,.3)[$b$]{} &&& (0,.3)[$a$]{}\
(0,-.3)[$a$]{} &&(0,-.3)[$a$]{} && (.38,0)(.7,.25) (-.3,-.3)[$b$]{} & (0,-.3)[$a$]{} && (0,-.3)[$b$]{} &&& (0,-.3)[$a$]{} && (0,-.3)[$a$]{} && (-.3,-.3)[$b$]{} & (0,-.3)[$a$]{} && (0,-.3)[$b$]{}\
&&&& (0,-.3)[$a$]{} &&&&&&&&&& (0,-.3)[$a$]{}\
&&&(0,0)[(e)]{} &&&&&&&&&& (0,0)[(f)]{}\
To prove the correctness of the ${\ensuremath{\textsc{Emb}}}$ procedure we need the following two propositions. The first proposition characterizes for node $v \in V(P)$ the set ${\ensuremath{\mathrm{emb}}}(v)$ using ${\ensuremath{\mathrm{mop}}}$, ${\ensuremath{\mathrm{nca}}}$, and ${\ensuremath{\mathrm{fl}}}$. The second proposition shows that the set $U_1$ computed in step 3 of the ${\ensuremath{\textsc{Emb}}}$ procedure is the set ${\ensuremath{\mathrm{mop}}}({\ensuremath{\textsc{Emb}}}(v_1), \ldots, {\ensuremath{\textsc{Emb}}}(v_k))$.
\[prop:embchildren\] Let $v$ be a node in $P$ and let $v_1, \ldots ,v_k$ be the sequence of children of $v$Êordered from left to right, where $k\geq 2$. For any node $w\in {\ensuremath{\mathrm{emb}}}(P(v),T)$, there exists a pair of nodes $(w_1,w_k)\in {\ensuremath{\mathrm{mop}}}({\ensuremath{\mathrm{emb}}}(P(v_1),T),\ldots,{\ensuremath{\mathrm{emb}}}(P(v_k),T))$ such that $w={\ensuremath{\mathrm{fl}}}({\ensuremath{\mathrm{nca}}}(w_1,w_k),{\ensuremath{\mathrm{label}}}(v))$.
Since $w$Êis the root of an occurrence of $P(v)$Êin $T$Êthere must exist a set of disjoint occurrences of $P(v_1),\ldots,P(v_k)$ in $T(w)$ with roots $w_1 \lhd \ldots \lhd w_k$, such that $w$ is an ancestor of $w_1,\ldots, w_k$. Since the $w_i$’s are ordered $w$ must be an ancestor of ${\ensuremath{\mathrm{nca}}}(w_1,w_k)$. Since $w$ is the root of a *deep* occurrence of $P(v)$Êin $T$ it follows that $w={\ensuremath{\mathrm{fl}}}({\ensuremath{\mathrm{nca}}}(w_1,w_k),{\ensuremath{\mathrm{label}}}(v))$.
It remains to show that we can assume $(w_1,w_k)\in {\ensuremath{\mathrm{mop}}}({\ensuremath{\mathrm{emb}}}(P(v_1),T),\ldots,{\ensuremath{\mathrm{emb}}}(P(v_k),T))$. It follows from the previous discussion that $(w_1,\ldots,w_k)\in \Phi({\ensuremath{\mathrm{emb}}}(P(v_1),T),\ldots,{\ensuremath{\mathrm{emb}}}(P(v_k),T))$. Assume for the sake of contradiction that $(w_1,w_k)$Êis *not* a minimum ordered pair. Then there exists a set of disjoint occurrences of $P(v_1),\ldots,P(v_k)$ in $T(w)$ with roots $u_1 \lhd \ldots \lhd u_k$, such that either $w_1 \lhd u_1 \lhd u_k \unlhd w_k$ or $w_1 \unlhd u_1 \lhd u_k \lhd w_k$, and $(u_1,u_k)\in{\ensuremath{\mathrm{mop}}}({\ensuremath{\mathrm{emb}}}(P(v_1),T),\ldots,{\ensuremath{\mathrm{emb}}}(P(v_k),T))$. Therefore $u={\ensuremath{\mathrm{fl}}}({\ensuremath{\mathrm{nca}}}(u_1,u_k),{\ensuremath{\mathrm{label}}}(v))$ is an embedding of $P(v)$Êin $T$. Now either $w \prec u$Êcontradicting the assumption that $w$ is a deep embedding or $w=u$ in which case $(u_1,u_k)$ satisfies the properties of the lemma (see also Figure \[fig:MopSubtree\](a)).
![(a) For all $i$,Ê$w_i$ and $u_i$ are roots of occurrences of $P(v_i)$ in $T$, and $w$ and $u$ is the nearest common ancestor of $(w_1,w_3)$ and $(u_1,u_3)$, respectively. Since $w_1 \lhd u_1$ and $u_3 \lhd w_3$ we cannot have $u \prec w$. (b) For all $i$,Ê$w_i$ is an embedding of $P(v_i)$ in $T$, $(w_1,w_4)$ is a minimum ordered pair, and $w$ is the nearest common ancestor of all the $w_i$’s. The number of leaves in $T(w)$ is at least $\sum_{i=1}^4 l_{T(w_i)} \geq \sum_{i=1}^4 l_{P(v_i)}$.[]{data-label="fig:MopSubtree"}](MopSubtree)
\[prop:Mop\] For $j+1\leq l \leq k$, $$\label{eq:uk}
U_l =
{\ensuremath{\mathrm{mop}}}({\ensuremath{\textsc{Emb}}}(v_j), \ldots, {\ensuremath{\textsc{Emb}}}(v_l)),$$ For $1\leq l \leq j-1$, $$\label{eq:uj}
U_l =
{\ensuremath{\mathrm{mop}}}({\ensuremath{\textsc{Emb}}}(v_l), \ldots, {\ensuremath{\textsc{Emb}}}(v_k)).$$
We first show Equation by induction on $l$. For $l=j+1$ it follows from the definition of ${\ensuremath{\textsc{MopRight}}}$ that $U_l$ is the set of minimum ordered pairs of ${\ensuremath{\textsc{Emb}}}(v_j)$ and ${\ensuremath{\textsc{Emb}}}(v_{j+1})$, i.e., $U_l = {\ensuremath{\mathrm{mop}}}({\ensuremath{\textsc{Emb}}}(v_j), {\ensuremath{\textsc{Emb}}}(v_{l}))$. Hence, assume that $l > j+1$. By the induction hypothesis we have $$U_l = {\ensuremath{\textsc{MopRight}}}(U_{l-1}, {\ensuremath{\textsc{Emb}}}(v_l)) =
{\ensuremath{\textsc{MopRight}}}({\ensuremath{\mathrm{mop}}}({\ensuremath{\textsc{Emb}}}(v_j), \ldots, {\ensuremath{\textsc{Emb}}}(v_{l-1})), R_l)\;.$$ By definition of ${\ensuremath{\textsc{MopRight}}}$, $U_l$ is the set of pairs such that for any pair $(r_j, r_{l-1}) \in {\ensuremath{\mathrm{mop}}}({\ensuremath{\textsc{Emb}}}(v_j), \ldots,
{\ensuremath{\textsc{Emb}}}(v_{l-1}))$, $(r_j, r_l) \in U_l$ iff $(r_{l-1}, r_l) \in
{\ensuremath{\mathrm{mop}}}({\ensuremath{\mathop{{\ensuremath{\mathrm{mop}}}({\ensuremath{\textsc{Emb}}}(v_j), \ldots, {\ensuremath{\textsc{Emb}}}(v_{l-1}))|}_{2}}},
R_l)$. By Lemma \[lem:nnsmaller2\] it follows that $(r_j, r_l)
\in U_l$ iff $(r_j, r_l) \in {\ensuremath{\mathrm{mop}}}({\ensuremath{\textsc{Emb}}}(v_j), \ldots,
{\ensuremath{\textsc{Emb}}}(v_l))$.
We can now similarly show Equation by induction on $j'=j-l$. By Equation we have $U_j= {\ensuremath{\mathrm{mop}}}({\ensuremath{\textsc{Emb}}}(v_j), \ldots, {\ensuremath{\textsc{Emb}}}(v_k))$ when we begin computing $U_{j-1}$. For $j'=1$ ($l=j-1$) it follows from the definition of ${\ensuremath{\textsc{MopLeft}}}$ that $U_{j-1} = {\ensuremath{\mathrm{mop}}}({\ensuremath{\textsc{Emb}}}(v_{j-1}), {\ensuremath{\textsc{Emb}}}(v_j))$. Hence, assume that $j' > 1$. Using Lemma \[lem:nnsmallerleft\] the Equation follows similarly to the proof of Equation .
By Proposition \[prop:Mop\], $U_1={\ensuremath{\mathrm{mop}}}({\ensuremath{\textsc{Emb}}}(v_1),\ldots,{\ensuremath{\textsc{Emb}}}(v_k))$. We can now show the correctness of procedure ${\ensuremath{\textsc{Emb}}}$.
\[lem:correctEmb\] For trees $P$ and $T$ and node $v\in V(P)$, ${\ensuremath{\textsc{Emb}}}(v)$ computes the set of deep occurrences of $P(v)$ in $T$.
By induction on the size of the subtree $P(v)$. If $v$ is a leaf, ${\ensuremath{\mathrm{emb}}}(v,T)$ is the deep set of nodes in $T$ with label ${\ensuremath{\mathrm{label}}}(v)$. It immediately follows that ${\ensuremath{\mathrm{emb}}}(v,T) = {\ensuremath{\textsc{Fl}}}(L(T),
{\ensuremath{\mathrm{label}}}(v))$ and thus case 1 follows.
Suppose that $v$ is an internal node with $k\geq1$ children $v_1, \ldots, v_k$. We show that ${\ensuremath{\mathrm{emb}}}(P(v),T) = {\ensuremath{\textsc{Emb}}}(v)$. Consider cases 2 and 3 of the algorithm.
For $k=1$ we have that $w\in {\ensuremath{\textsc{Emb}}}(v)$ implies that ${\ensuremath{\mathrm{label}}}(w) =
{\ensuremath{\mathrm{label}}}(v)$ and there is a node $w_1 \in {\ensuremath{\textsc{Emb}}}(v_1)$ such that ${\ensuremath{\mathrm{fl}}}({\ensuremath{\mathrm{parent}}}(w_1), {\ensuremath{\mathrm{label}}}(v)) = w$, that is, no node on the path between $w_1$ and $w$ is labeled ${\ensuremath{\mathrm{label}}}(v)$. By induction ${\ensuremath{\textsc{Emb}}}(v_1) = {\ensuremath{\mathrm{emb}}}(P(v_1),T)$ and therefore $w$ is the root of an embedding of $P(v)$ in $T$. Since ${\ensuremath{\textsc{Emb}}}(v)$ is the deep set of all such nodes it follows that $w \in {\ensuremath{\mathrm{emb}}}(P(v),T)$. Conversely, if $w \in {\ensuremath{\mathrm{emb}}}(P(v),T)$ then ${\ensuremath{\mathrm{label}}}(w) = {\ensuremath{\mathrm{label}}}(v)$, there is a node $w_1 \in {\ensuremath{\mathrm{emb}}}(P(v_1),T)$ such that $w \prec w_1$, and no node on the path between $w$ and $w_1$ is labeled ${\ensuremath{\mathrm{label}}}(v)$, that is, ${\ensuremath{\mathrm{fl}}}(w_1, {\ensuremath{\mathrm{label}}}(v)) = w$. Hence, $w \in {\ensuremath{\textsc{Emb}}}(v)$.
Next consider the case $k>1$. By Proposition \[prop:Mop\] and the induction hypothesis $$U_1={\ensuremath{\mathrm{mop}}}({\ensuremath{\mathrm{emb}}}(P(v_1),T),\ldots,{\ensuremath{\mathrm{emb}}}(P(v_k),T))\;.$$ We first show that $w \in
{\ensuremath{\mathrm{emb}}}(P(v),T)$ implies that $w\in {\ensuremath{\textsc{Emb}}}(v)$. By Proposition \[prop:embchildren\] there exists a pair of nodes $(w_1,w_k)\in {\ensuremath{\mathrm{mop}}}({\ensuremath{\mathrm{emb}}}(P(v_1),T),\ldots,{\ensuremath{\mathrm{emb}}}(P(v_k),T))$ such that $w={\ensuremath{\mathrm{fl}}}({\ensuremath{\mathrm{nca}}}(w_1,w_k),{\ensuremath{\mathrm{label}}}(v))$. We have $(w_1,w_k)\in U_1$ and it follows directly from the implementation that $w\in {\ensuremath{\textsc{Emb}}}(v)$. To see that we do not loose $w$Êby taking [$\textsc{Deep}$]{} of ${\ensuremath{\textsc{Nca}}}(U_1)$ assume that $w'={\ensuremath{\mathrm{nca}}}(w_1,w_k)$ is removed from the set in this step. This means there is a node $u$Êin ${\ensuremath{\textsc{Nca}}}(U_1)$ which is a descendant of $w'$ and which is still in the set. Since $w$Êis the root of a *deep* occurrence we must have $w={\ensuremath{\mathrm{fl}}}(w',{\ensuremath{\mathrm{label}}}(v))={\ensuremath{\mathrm{fl}}}(u,{\ensuremath{\mathrm{label}}}(v))$.
Let $w \in {\ensuremath{\textsc{Emb}}}(v)$. Then $w$Êis the first ancestor with label ${\ensuremath{\mathrm{label}}}(v)$ of a nearest common ancestor of a pair in $U_1$. That is, ${\ensuremath{\mathrm{label}}}(w) = {\ensuremath{\mathrm{label}}}(v)$ and there exists nodes $(w_1, w_k) \in
{\ensuremath{\mathrm{mop}}}({\ensuremath{\mathrm{emb}}}(P(v_1),T), \ldots, {\ensuremath{\mathrm{emb}}}(P(v_k),T))$ such that $w={\ensuremath{\mathrm{fl}}}({\ensuremath{\mathrm{nca}}}(w_1, w_k), {\ensuremath{\mathrm{label}}}(v))$. Clearly, $w$ is the root of an embedding of $P(v)$ in $T$. Assume for contradiction that $w$ is not a deep embedding, that is, $w \prec u$ for some node $u \in
{\ensuremath{\mathrm{emb}}}(P(v),T)$. We have just shown that this implies $u\in {\ensuremath{\textsc{Emb}}}(v)$. Since ${\ensuremath{\textsc{Emb}}}(v)$Êis a deep set this contradicts $w \in {\ensuremath{\textsc{Emb}}}(v)$.
The set $L(T)$ is deep and in all three cases of ${\ensuremath{\textsc{Emb}}}(V)$ the returned set is also deep. By induction it follows that the input to ${\ensuremath{\textsc{Parent}}}$, ${\ensuremath{\textsc{Fl}}}$, ${\ensuremath{\textsc{Nca}}}$, and ${\ensuremath{\textsc{MopRight}}}$ is always deep. We will use this fact to our advantage in the following algorithms.
A Simple Tree Inclusion Algorithm {#simple}
=================================
In this section we a present a simple implementation of the set procedures which leads to an efficient tree inclusion algorithm. Subsequently, we modify one of the procedures to obtain a family of tree inclusion algorithms where the complexities depend on the solution to a well-studied problem known as the *tree color problem*.
Preprocessing {#sec:simplepreprocessing}
-------------
To compute deep embeddings we require a data structure for $T$ which allows us, for any $v,w \in V(T)$, to compute ${\ensuremath{\mathrm{nca}}}_T(v,w)$ and determine if $v \prec w$ or $v \lhd w$. In linear time we can compute ${\ensuremath{\mathrm{pre}}}(v)$ and ${\ensuremath{\mathrm{post}}}(v)$ for all nodes $v \in V(T)$, and with these it is straightforward to test the two conditions. Furthermore,
For any tree $T$ there is a data structure using $O(n_T)$ space and preprocessing time which supports nearest common ancestor queries in $O(1)$ time.
Hence, our data structure uses linear preprocessing time and space (see also [@BFC2000; @AGKR2004] for more recent nearest common ancestor data structures).
Implementation of the Set Procedures {#implementationsimple}
------------------------------------
To answer tree inclusion queries we give an efficient implementation of the set procedures. The idea is to represent sets of nodes and sets of pairs of nodes in a left-to-right order using linked lists. For this purpose we introduce some helpful notation. Let $X = [x_1, \ldots, x_k]$ be a linked list of nodes. The *length* of $X$, denoted $|X|$, is the number of elements in $X$ and the list with no elements is written $[]$. The $i$th node of $X$, denoted $X[i]$, is $x_i$. Given any node $y$ the list obtained by *appending* $y$ to $X$, is the list $X \circ y =
[x_1, \ldots, x_k, y]$. If for all $i$, $1 \leq i \leq |X|-1$, $X[i] \lhd X[i+1]$ then $X$ is *ordered* and if $X[i] \unlhd
X[i+1]$ then $X$ is *semiordered*. Recall that $X[i] \unlhd X[i+1]$ means that we can have $X[i] \lhd X[i+1]$ or either of the nodes can be an ancestor of the other ($X[i] \lhd X[i+1]$ or $X[i] \preceq X[i+1]$ or $X[i] \succ X[i+1]$). A list $Y = [(x_1, z_k),
\ldots, (x_k, z_k)]$ is a *node pair list*. By analogy, we define length, append, etc. for $Y$. For a pair $Y[i] = (x_i,
z_i)$ define $Y[i]_1 = x_i$ and $Y[i]_2 = z_i$. If the lists $[Y[1]_1, \ldots, Y[k]_1]$ and $[Y[1]_2, \ldots, Y[k]_2]$ are both ordered or semiordered then $Y$ is *ordered* or *semiordered*, respectively.
The set procedures are implemented using node lists. All lists used in the procedures are either ordered or semiordered. As noted in Section \[sec:recursion\] we may assume that the input to all of the procedures, except ${\ensuremath{\textsc{Deep}}}$, represent a deep set, that is, the corresponding node list or node pair list is ordered. We assume that the input list given to ${\ensuremath{\textsc{Deep}}}$ is semiordered and the output, of course, is ordered. Hence, the output of all the other set procedures must be semiordered. In the following let $X$ be a node list, $Y$ a node pair list, and $\alpha$ a character in $\Sigma$. The detailed implementation of the set procedures is given below. We show the correctness in Section \[sec:simplecorrectness\] and discuss the complexity in Section \[sec:simplecomplexity\].
Return the list $[{\ensuremath{\mathrm{parent}}}(X[1]), \ldots,
{\ensuremath{\mathrm{parent}}}(X[{\ensuremath{| X |}}])]$.
Return the list $[{\ensuremath{\mathrm{nca}}}(Y[1]), \ldots, {\ensuremath{\mathrm{nca}}}(Y[{\ensuremath{| Y |}}])]$.
Initially, set $x := X[1]$ and $R := []$.
For $i:=2$ to $|X|$ do:
- Compare $x$ and $X[i]$. There are three cases:
1. $x \lhd X[i]$. Set $R:= R \circ x$ and $x:= X[i]$.
2. $x \prec X[i]$. Set $x := X[i]$.
3. $X[i] \preceq x$. Do nothing.
Return $R \circ x$.
The implementation of procedure [$\textsc{Deep}$]{} takes advantage of the fact that the input list is semiordered. In case 1 the node $X[i]$ is to the right of our “potential output node” $x$. Since any node that is a descendant of $x$ must also be to the left of $X[i]$ it cannot appear later in the list $X$ than $X[i]$. We can thus safely add $x$ to $R$ at this point. In case 2 the node $x$ is an ancestor of $X[i]$ and thus $x$ cannot be in ${\ensuremath{\textsc{Deep}}}(X)$. In case 3 the node $X[i]$ is an ancestor of $x$ and can therefore not be in ${\ensuremath{\textsc{Deep}}}(X)$.
Initially, set $R:=[]$.
Find the smallest $j$ such that $Y[1]_2 \lhd X[j]$ and set $y:=Y[1]_1$, $x:= X[j]$. If no such $j$ exists stop.
For $i :=2$ to ${\ensuremath{| Y |}}$ do:
- Until $Y[i]_2 \lhd X[j]$ or $j>{\ensuremath{| X |}}$ set $j:=j+1$. If $j> {\ensuremath{| X |}}$ stop and return $R := R \circ (y,x)$. Otherwise, compare $X[j]$ and $x$. There are two cases:
1. If $x \lhd X[j]$ set $R:=R\circ (y,x)$, $y:=Y[i]_1$, and $x:=X[j]$.
2. If $x=X[j]$ set $y:=Y[i]_1$.
Return $R := R \circ (y,x)$.
In procedure [$\textsc{MopRight}$]{} we have a “potential pair” $(y,x)$ where $y=Y[i']_1$ for some $i'$ and $Y[i']_2 \lhd x$. In case 1 we have $x \lhd X[j]$ and also $Y[i']_2 \lhd Y[i]_2$ since the input lists are ordered and $i'<i$ (see Figure \[fig:mopimplexample\](a)). Therefore, $(y,x)$ is inserted into $R$. In case 2 we have $x=X[j]$, i.e., $Y[i]_2\lhd
x$, and as before $Y[i']_2 \lhd Y[i]_2$ (see Figure \[fig:mopimplexample\](b)). Therefore $(y,x)$ cannot be in ${\ensuremath{\textsc{MopRight}}}(Y,X)$, and we set $(Y[i]_1,x)$ to be the new potential pair.
&&&& (0,.3)\
& (0,.3) &&& (-.3,.3) && (0,.3)\
& (0,-.3)[$Y[i']_2$]{} & (-.2,-.3) & (.6,0)[$Y[i]_2$]{} & & & & (0,-.3)[$X[j]$]{}&&&\
&& & (0,-.3)[$x$]{} & && (0,-.3)\
&&&&(0,0)[(a)]{}
&&&&\
& &&& &&\
& (0,-.3)[$Y[i']_2$]{} & & && & &\
&& & & (0,-.3)[$Y[i]_2$]{}&& (.3,-.3)[$x=X[j]$]{}\
&&&&(0,0)[(b)]{}
We can implement ${\ensuremath{\textsc{MopLeft}}}(X,Y)$ similarly to [$\textsc{MopRight}$]{} replacing smallest by largest, $\lhd$ by $\rhd$, and traversing the lists backwards:
Initially, set $R:=[]$.
Find the largest $j$ such that $Y[{\ensuremath{| Y |}}]_1 \rhd X[j]$ and set $y:=Y[{\ensuremath{| Y |}}]_2$ and $x:= X[j]$. If no such $j$ exists stop.
For $i :={\ensuremath{| Y |}}-1$ to $1$ do:
- Until $Y[i]_1 \rhd X[j]$ or $j<1$ set $j:=j-1$. If $j<1$ stop and return $R := R \circ (y,x)$. Otherwise, compare $X[j]$ and $x$. There are two cases:
1. If $x \rhd X[j]$ set $R:=(x,y) \circ R$, $y:=Y[i]_2$, and $x:=X[j]$.
2. If $x=X[j]$ set $y:=Y[i]_2$.
Return $R := R \circ (x,y)$.
Initially, set $L:=X$, $Z:=L$.
Repeat until $Z=[]$:
- For $i:=1$ to ${\ensuremath{| Z |}}$ do:
1. If ${\ensuremath{\mathrm{label}}}(Z[i])=\alpha$: Delete $Z[i]$ from $Z$ (but keep it in $L$).
2. If ${\ensuremath{\mathrm{label}}}(Z[i])\neq \alpha$ and ${\ensuremath{\mathrm{parent}}}(Z[i])\neq \bot$: Replace $Z[i]$ with ${\ensuremath{\mathrm{parent}}}(Z[i])$ in both $Z$ and $L$.
3. If ${\ensuremath{\mathrm{label}}}(Z[i])\neq \alpha$ and ${\ensuremath{\mathrm{parent}}}(Z[i])= \bot$: Delete $Z[i]$ from both $Z$ and $L$.
- Set $Z:={\ensuremath{\textsc{Deep}}}^*(Z,L)$.
Return $L$.
The procedure [$\textsc{Fl}$]{} computes the set ${\ensuremath{\textsc{Deep}}}(\{{\ensuremath{\mathrm{fl}}}(x,\alpha)| x\in X \})$ bottom up. The list $Z$Êcontains ancestors of the elements of $X$ for which we have not yet found an ancestor with label $\alpha$. In each step it considers each node $z$Êin the list $Z$. If it has the right label then $x \in {\ensuremath{\textsc{Fl}}}(X,\alpha)$ and we remove it from $Z$ but keep it in $L$. Otherwise we replace it with its parent (unless it is the root). Thus $L$Êcontains both the elements in $Z$Êand the part of ${\ensuremath{\textsc{Fl}}}(X,\alpha)$ found until now.
To keep the running time down we wish to maintain the invariant that $L$Êis deep at the beginning of each iteration of the outer loop. To do this procedure ${\ensuremath{\textsc{Fl}}}$ calls an auxiliary procedure ${\ensuremath{\textsc{Deep}}}^*(Z,L)$ which takes two ordered lists $Z$Êand $L$, where $Z\subseteq L$, and returns the ordered list representing the set ${\ensuremath{\textsc{Deep}}}(L) \cap Z$, i.e., ${\ensuremath{\textsc{Deep}}}^*(Z,L)=[z\in Z|\nexists
x \in L:z\prec x]$. If we use the procedure [$\textsc{Deep}$]{} to calculate ${\ensuremath{\textsc{Deep}}}^*$ it takes time $O({\ensuremath{| Z |}}+{\ensuremath{| L |}})=O({\ensuremath{| L |}})$. Instead we will show how to calculate it in time $O({\ensuremath{| Z |}})$ using a linked list representation for $Z$Êand $L$. We will need this in the proof of Lemma \[lem:fl\], which shows that the total running time of $\emph{all}$Êcalls to [$\textsc{Fl}$]{} from [$\textsc{Emb}$]{} takes time $O(n_T)$. Below we describe in more detail how to implement [$\textsc{Fl}$]{} together with the auxiliary procedures.
We use a doubly linked list to represent $L$ and extra pointers in this list to represent $Z$. Each element in the list has pointers [$\mathsf{Succ}_L$]{} and to its predecessor and successor in $L$. Similarly, each element in $Z$ has pointers [$\mathsf{Succ}_Z$]{} and to its predecessor and successor in $Z$ (right after the initialization these are equal to [$\mathsf{Succ}_L$]{} and [$\mathsf{Pred}_L$]{}). In the for loop we use the [$\mathsf{Succ}_Z$]{} pointers to find the next element in $Z$. To delete $Z[i]$Êfrom $Z$ in case 1 we set ${\ensuremath{\mathsf{Succ}_Z}}({\ensuremath{\mathsf{Pred}_Z}}(Z[i]))={\ensuremath{\mathsf{Succ}_Z}}(Z[i])$ and ${\ensuremath{\mathsf{Pred}_Z}}({\ensuremath{\mathsf{Succ}_Z}}(Z[i]))={\ensuremath{\mathsf{Pred}_Z}}(Z[i])$. The $L$ pointers stay the same. In case 2 we simply replace $Z[i]$ with its parent in the linked list. The [$\mathsf{Succ}$]{} and [$\mathsf{Pred}$]{} pointers stay the same. To delete $Z[i]$ from both $Z$ and $L$Êin case 3 we set ${\ensuremath{\mathsf{Succ}}}_j({\ensuremath{\mathsf{Pred}}}_j(Z[i]))={\ensuremath{\mathsf{Succ}}}_j(Z[i])$ and ${\ensuremath{\mathsf{Pred}}}_j({\ensuremath{\mathsf{Succ}}}_j(Z[i]))={\ensuremath{\mathsf{Pred}}}_j(Z[i])$ for $j \in \{Z,L\}$. Finally, to compute ${\ensuremath{\textsc{Deep}}}^*(Z,L)$ walk through $Z$ following the [$\mathsf{Succ}_Z$]{} pointers. At each node $z$ compare ${\ensuremath{\mathsf{Pred}_L}}(z)$ and ${\ensuremath{\mathsf{Succ}_L}}(z)$ with $z$. If one of them is a descendant of $z$ remove $z$ from the doubly linked list $Z$ and $L$ as in case 3. Note that instead of calling ${\ensuremath{\textsc{Deep}}}^*(Z,L)$ this comparison can also be done directly in step 2, which is the only place where we insert nodes that might be an ancestor of another node in $L$. We will show in the next section that it is enough to compare $z$ to its neighbors in the list $L$.
Correctness of the Set Procedures {#sec:simplecorrectness}
---------------------------------
Clearly, [$\textsc{Parent}$]{} and [$\textsc{Nca}$]{} are correct. The following lemmas show that ${\ensuremath{\textsc{Deep}}}$, ${\ensuremath{\textsc{Fl}}}$, and ${\ensuremath{\textsc{MopRight}}}$ are also correctly implemented. For notational convenience we write $x\in X$, for a list $X$, if $x = X[i]$ for some $i$, $1 \leq i \leq {\ensuremath{| X |}}$.
\[lem:deep\] Procedure ${\ensuremath{\textsc{Deep}}}(X)$ is correct.
Let $x$Êbe the variable in the procedure. We will first prove the following invariant on $x$: Let $i$Êbe the current index. Right before increasing $i$Êin line 2 of the code (“For $i:=2$ to $|X|$ do:”), we have $x \not\prec X[j]$ for any $1\leq j \leq i$.
Induction on $i$. The base case $i=1$Êis obviously correct. Let $i\geq 2$. Let $x'$ denote the value of the $x$Êvariable before the last iteration. If $x=x'$Êthen by the induction hypothesis $x \not\prec X[j]$ for any $1\leq j \leq i-1$. Since $x$Êwas not changed we have $X[i]\preceq x$ (case 3 of the procedure) and thus $x \not\prec X[j]$ for any $1\leq j \leq i$. If $x\neq x'$ then $x=X[i]$ and by the induction hypothesis $x' \not\prec X[j]$ for any $1\leq j \leq i-1$. Since $x'\neq x$Êeither $x'\lhd x$ (case 1 of the procedure) or $x' \prec x$ (case 2 of the procedure) holds. For $x'\prec x$Êit follows immediately from the induction hypothesis that $x \not\prec X[j]$ for any $1\leq j \leq i$, since all descendants of $x$ also are descendants of $x'$. For $x'\lhd x$ we note that $X[i]$ is the first node $x''$ occurring after $x'$Êin $X$ such that $x'\lhd x''$ (otherwise $x$Êwould have been reset in case 1 in the iteration where $X[i]=x''$). Since $X$ is semiordered no node $X[j]$ with smaller index than $i$ can be to the right of $x'$. Thus no node $X[j]$, $1\leq j < i$, can be to the right of $x'$. Since all descendants of $x$ must be to the right of $x'$ we have $x \not\prec X[j]$ for any $1\leq j \leq i$.
We are now ready to prove that $y\in {\ensuremath{\textsc{Deep}}}(X)$ iff there exists no $z\in X$ such that $y\prec z$. We first argue that if $y\in {\ensuremath{\textsc{Deep}}}(X)$ then $\nexists z\in X$ such that $y\prec z$. Let $y$Êbe an element in ${\ensuremath{\textsc{Deep}}}(X)$. Only elements that have been assigned to $x$Êduring the procedure are in the output. Consider the iteration where $x = y$ is appended to $R$. This only happens in case 1 of the procedure and thus $y=x \lhd X[i]$. Since $X$ is semiordered this implies that $x \lhd X[j]$ for $i \leq j \leq {\ensuremath{| X |}}$, and therefore $y=x \not \prec X[j]$ for $i \leq j \leq {\ensuremath{| X |}}$. By the above invariant it follows that $y=x \not \prec X[j]$ for $1 \leq j \leq i-1$. Thus if $y\in {\ensuremath{\textsc{Deep}}}(X)$ then $\nexists z\in X$ such that $y\prec z$.
Let $y\in X$ be an element such that $X\cap V(T(y))=\{y\}$. Let $j$Êbe the smallest index such that $X[j]=y$. When comparing $y$ and $x$Êduring the iteration where $i=j$Êwe are in case 1 or 2, since $j$Êis the smallest index such that $X[j]=y$ (implying $x\neq y$) and $X\cap V(T(y))=\{y\}$ (implying $y\not \prec x$). In either case $x$Êis set to $y$. Since there are no descendants of $y$ in $X$, the variable $x$Êremains equal to $y$ until added to $R$. If $y$ occurs several times in $X$ we will have $x=y$ each time we meet a copy of $y$ (except the first) and it follows from the implementation that $y$ will occur exactly once in $R$.
To show that the implementation of [$\textsc{MopRight}$]{} is correct we will use the following proposition.
\[prop:mopInvariant\] Before the first iteration of the for loop we have $y=Y[1]_1$, $x=X[j]$ and either $X[j-1] \lhd Y[1]_2 \lhd X[j]$ (if $j>1$) or $Y[1]_2 \lhd X[j]$ if ($j=1$).
At the end of each iteration of the for loop then, unless $Y[i]_2 \not\!\!\lhd X[|X|]$, we have $y=Y[i]_1$, $x=X[j]$ and either $X[j-1] \lhd Y[i]_2 \lhd X[j]$ (if $j>1$) or $Y[i]_2 \lhd X[j]$ if ($j=1$).
The first statement ($y=Y[1]_1$, $x=X[j]$ and either $X[j-1] \lhd Y[1]_2 \lhd X[j]$ (if $j>1$) or $Y[1]_2 \lhd X[j]$ if ($j=1$)) follows immediately from the implementation of the procedure line 2 and 3 (“Find the smallest $j$ such that $\ldots$”).
We prove the second statement by induction on $i$. Base case $i=2$. By the first statement we have $y=Y[1]_1$, $x=X[j]$ and either $X[j-1] \lhd Y[1]_2 \lhd X[j]$ (if $j>1$) or $Y[1]_2 \lhd X[j]$ if ($j=1$) before this iteration. Let $j'$ be the value of $j$ before this iteration. It follows immediately from the implementation that $y=Y[2]_1$ since $y$ is set to this in both case 1 and 2. If $Y[2]_2 \lhd X[j']$ then $j=j'$. Since $Y$ is ordered it follows that $X[j-1] \lhd Y[1]_2\lhd Y[2]_2 \lhd X[j]$ (if $j>1$) or $Y[2]_2 \lhd X[j]$ if ($j=1$). If $Y[2]_2 \not\!\!\lhd X[j']$ then $j$ is increased until $Y[2]_2 \lhd X[j]$ implying $X[j-1] \lhd Y[2]_2 \lhd X[j]$ unless $j>{\ensuremath{| X |}}$, since $X$ is ordered.
Induction step $i >2$. It follows immediately from the implementation that $y=Y[i]_1$ since $y$ is set to this in both case 1 and 2. By the induction hypothesis we have $y=Y[i]_1$, $x=X[j]$ and $Y[i]_2 \lhd X[j]$ right before this iteration. Let $j'$ be the value of $j$ before this iteration. If $Y[i]_2 \lhd X[j']$ then $j=j'$. Since $Y$ is ordered it follows that $X[j-1] \lhd Y[i-1]_2\lhd Y[i]_2 \lhd X[j]$ (if $j>1$) or $Y[i]_2 \lhd X[j]$ if ($j=1$). If $Y[i]_2 \not\!\!\lhd X[j']$ then $j$ is increased until $Y[i]_2 \lhd X[j]$ implying $X[j-1] \lhd Y[i]_2 \lhd X[j]$ unless $j>{\ensuremath{| X |}}$.
\[lem:nnm\] Procedure ${\ensuremath{\textsc{MopRight}}}(Y,X)$ is correct.
We want to show that for any $1\leq i' \leq {\ensuremath{| Y |}}$, $1 \leq j' \leq
{\ensuremath{| X |}}$: $$(Y[i']_1,X[j'])\in R \quad \Leftrightarrow \quad (Y[i']_2,X[j'])
\in {\ensuremath{\mathrm{mop}}}(Y|_2,X)\;.$$
Since ${\ensuremath{\mathop{Y|}_{2}}}$ and $X$ are ordered lists we have $$\label{eq:mop}
(Y[i']_2,X[j']) \in {\ensuremath{\mathrm{mop}}}(Y|_2,X) \quad \Leftrightarrow \quad
X[j'-1] \; \unlhd \; Y[i']_2 \; \lhd \; X[j'] \; \unlhd \; Y[i'+1]_2 \;,$$ for $i'< |Y|$Êand $j' \geq 2$. We have three *border cases* $$\label{eq:mopborder}
(Y[i']_2,X[j']) \in {\ensuremath{\mathrm{mop}}}(Y|_2,X) \quad \Leftrightarrow \quad
\begin{cases}
X[j'-1] \unlhd Y[i']_2 \lhd X[j'], & \textrm{for } i'= |Y|Ê\textrm{ and } j' \geq 2\\
Y[i']_2 \lhd X[j'] \unlhd Y[i'+1]_2, & \textrm{for } i'< |Y|Ê\textrm{ and } j' = 1 \\
Y[i']_2 \lhd X[j'], & \textrm{for } i'= |Y|Ê\textrm{ and } j'=1
\end{cases}$$ Thus to show $(Y[i']_1,X[j'])\in R \Leftrightarrow (Y[i']_2,X[j'])
\in {\ensuremath{\mathrm{mop}}}({\ensuremath{\mathop{Y|}_{2}}},X)$ it is enough to show $(Y[i']_1,X[j'])\in R \Leftrightarrow X[j'-1] \unlhd Y[i']_2 \lhd X[j'] \unlhd Y[i'+1]_2 $ (plus the border cases).
We start by showing that if the right hand side of Equation \[eq:mop\] or \[eq:mopborder\], i.e., $(Y[i']_2,X[j']) \in {\ensuremath{\mathrm{mop}}}(Y|_2,X)$, is satisfied then $(Y[i]_1,X[j])\in R$. We split the proof into two cases depending on the value of $j'$:
- $j'=1$: Since the list $Y|_2$ is ordered, $x$Êis set to $X[1]$ in line 2 (“Find the smallest $j$Êsuch that …”) and $y$Êis set to $Y[1]_1$. Now $x$ and remains the same when we get to the step in the iteration when $i=i'$ in the for loop. Thus $y$Êis set to $Y[i']_1$ (case 1). There are three subcases:
- $i'=1$: Then $y$Êis already set to $Y[i']_1$.
- $i'=|Y|$: Then we end the for loop after this iteration and return $R:=R \circ (y,x)=R \circ (Y[i']_1,X[1])$.
- $i' < |Y|$: Then $j$ is increased in the next iteration since $X[j] \unlhd Y[i'+1]_2$.
Therefore $(y,x)=(Y[i']_1,X[1])$ is added to $R$ (case 1 of the procedure or the stop condition in line 2 in the for loop (“If $j > {\ensuremath{| X |}}$ stop and return $R:= R\circ (y,x)$.”)).
- $j'\geq 2$: Since $X$ is ordered $x \unlhd X[j']$ after the execution of line 2 and 3 in the code (“Find the smallest $j$Êsuch that …”). There are two subcases:
- $i'=1$:ÊWe now have $x=X[j']$ and $y=Y[i']_1$.
- $i' >1$: Consider the step in the iteration when $i=i'$. Since $Y|_2$ and $X$ are both ordered $j\leq j'$. Now either $x=X[j']$Ê($j=j'$) in which case we set $y:=Y[i']_1$Ê(case 2), or $x \lhd X[j']$ ($j<j'$) in which case $j$ is increased to $j'$ since $X[j'-1]\lhd Y[i'] \lhd X[j']$. In the latter case we set $x:=X[j']$ and $y:=Y[i']_1$.
We have now argued that after the iteration where $i=i'$ we have $x:=X[j']$ and $y:=Y[i']_1$. If $|Y|=i'$ then we return $R=(y,x)$. If $i'<|Y|$ then $j$ is increased in the next iteration since $X[j'] \unlhd Y[i'+1]_2$. It follows from the code that when $j$Êis increased $(y,x)$ is added to $R$ (either case 1 of the procedure or the stop condition in line 2 in the for loop (“If $j > {\ensuremath{| X |}}$ stop and return $R:= R\circ (y,x)$.”)).
We will now show that $(Y[i']_1,X[j'])\in R$ implies one of the right hand sides of Equation \[eq:mop\] and \[eq:mopborder\]. Since $(Y[i']_1,X[j'])\in R$ we had $(y,x)=(Y[i']_1,X[j'])$ at some point during the execution. The pair $(y,x)$ can be added to $R$ only in the for loop before changing the values of $y$Êand $x$ or at the execution of the last line of the procedure. Therefore $(y,x)=(Y[i']_1,X[j'])$ at the beginning of some execution of the for loop, or after the last iteration ($i=|Y|$). It follows by Proposition \[prop:mopInvariant\] that $X[j-1] \lhd Y[i]_2 \lhd X[j]$ if $j>1$ or $Y[i]_2 \lhd X[j]$ if $j=1$. It remains to show that $X[j']\lhd Y[i'+1]_2$ for $i'<|Y|$. It follows from the implementation that $(y,x)$ only is added to $R$ inside the for loop if $j$ is increased. Thus $j$Êwas increased in the next iteration ($i=i'+1$) implying $X[j']\lhd Y[i'+1]_2$.
Procedure ${\ensuremath{\textsc{MopLeft}}}(X,Y)$ is correct.
Similar to the proof of Lemma \[lem:nnm\].
To show that [$\textsc{Fl}$]{} is correct we need the following proposition.
\[prop:deeplist\] Let $X$ be an ordered list and let $x$ be an ancestor of $X[i]$ for some $i \in \{1,\ldots,k\}$. If $x$ is an ancestor of some node in $X$ other than $X[i]$ then $x$ is an ancestor of $X[i-1]$ or $X[i+1]$.
Recall that $u\lhd v$ iff ${\ensuremath{\mathrm{pre}}}(u)<{\ensuremath{\mathrm{pre}}}(v)$ and ${\ensuremath{\mathrm{post}}}(u)\leq {\ensuremath{\mathrm{post}}}(v)$. Since $x\prec X[i]$ we have ${\ensuremath{\mathrm{pre}}}(x)<{\ensuremath{\mathrm{pre}}}(X[i])$ and ${\ensuremath{\mathrm{post}}}([X[i])<{\ensuremath{\mathrm{post}}}(x)$. Assume there exists a descendant $X[j]$ of $x$ such that $j\not \in \{i-1,i,i+1\}$. If $j<i-1$ we have $${\ensuremath{\mathrm{pre}}}(x)\leq {\ensuremath{\mathrm{pre}}}(X[j]) <{\ensuremath{\mathrm{pre}}}(X[i-1]),$$ where the first inequality follows from $x\prec X[j]$ and the second from $X$ being ordered. And $${\ensuremath{\mathrm{post}}}(X[i-1])<{\ensuremath{\mathrm{post}}}(X[i])\leq {\ensuremath{\mathrm{post}}}(x),$$ where the first inequality follows from $X$ being ordered and the second from $x\prec X[i]$. Thus $x \prec X[i-1]$.
Similarly, for $j> i+1$, we have ${\ensuremath{\mathrm{pre}}}(x)\leq {\ensuremath{\mathrm{pre}}}(X[i]) <{\ensuremath{\mathrm{pre}}}(X[i+1])$ and ${\ensuremath{\mathrm{post}}}(X[i+1])<{\ensuremath{\mathrm{post}}}(X[j])\leq {\ensuremath{\mathrm{post}}}(x)$ implying that $x\prec X[i+1]$.
Proposition \[prop:deeplist\] shows that the doubly linked list implementation of ${\ensuremath{\textsc{Deep}}}^*$ is correct. Since all changes to the list are either deletions or insertions of a parent in the place of its child, the list $L$ (and thus also $Z$) is ordered at the beginning of each iteration of the outer loop.
\[lem:flcorrect\] Procedure ${\ensuremath{\textsc{Fl}}}(X,\alpha)$ is correct.
Let $F=\{{\ensuremath{\mathrm{fl}}}(x,\alpha)\mid x \in X\}$. We first show that ${\ensuremath{\textsc{Fl}}}(X,\alpha) \subseteq F$. Consider a node $x \in {\ensuremath{\textsc{Fl}}}(X,\alpha)$. Since $x$ is in $L$Êafter the final iteration, $x$Êwas deleted from $Z$Êduring some iteration. Thus ${\ensuremath{\mathrm{label}}}(x)=\alpha$. For any $y\in X$ we follow the path from $y$Êto the root and stop the first time we meet a node with label $\alpha$ or even earlier since we keep the list deep. Thus $x\in F$.
It remains to show that ${\ensuremath{\textsc{Deep}}}(F)\subseteq {\ensuremath{\textsc{Fl}}}(X,\alpha)$. Let $x$ be a node in ${\ensuremath{\textsc{Deep}}}(F)$, let $z \in X$ be the node such that $x={\ensuremath{\mathrm{fl}}}(z,\alpha)$, and let $z=x_1,x_2,\ldots,x_k=x$ be the nodes on the path from $z$ to $x$. We will argue after each iteration of the algorithm we have $x_i \in L$ for some $i$. Since ${\ensuremath{\mathrm{label}}}(x_i)\neq \alpha$ for $i<k$ this is the same as $x_i\in Z$ for $i<k$.ÊBefore the first iteration we have $x_1 \in X =Z$. As long as $i<k$Êwe replace $x_i$ with $x_{i+1}$Êin case 1 of the for loop, since ${\ensuremath{\mathrm{label}}}(x_i)\neq \alpha$. When $i=k$Êwe remove $x_k$ from $Z$Êbut keep it in $L$. It remains to show that we do not delete $x_i$Êin the computation of ${\ensuremath{\textsc{Deep}}}^*(Z,L)$ in any iteration. If $x_i$Êis removed then there is a node $z\in L$ that is a descendant of $x_i$ and thus also a descendant of $x$. We argued above that $L\setminus Z\subseteq F$ and thus $z \in Z$ since $x \in {\ensuremath{\textsc{Deep}}}(F)$. But since $x \in {\ensuremath{\textsc{Deep}}}(F)$ no node on the path from $z$Êto $x$ can have label $\alpha$ and therefore $x_i$Êwill eventually be reinserted in $Z$.
Since ${\ensuremath{\textsc{Fl}}}(X,\alpha)$ is a deep set, we have ${\ensuremath{\textsc{Deep}}}(F)\subseteq {\ensuremath{\textsc{Fl}}}(X,\alpha) \subseteq(F) \Rightarrow {\ensuremath{\textsc{Deep}}}(F)={\ensuremath{\textsc{Fl}}}(X,\alpha)$.
Complexity of the Set Procedures {#sec:simplecomplexity}
--------------------------------
For the running time of the node list implementation observe that, given the data structure described in Section \[sec:simplepreprocessing\], all set procedures, except ${\ensuremath{\textsc{Fl}}}$, perform a single pass over the input using constant time at each step. Hence we have,
\[lem:auxprocedures\] For any tree $T$ there is a data structure using $O(n_T)$ space and preprocessing which supports each of the procedures ${\ensuremath{\textsc{Parent}}}$, ${\ensuremath{\textsc{Deep}}}$, ${\ensuremath{\textsc{MopRight}}}$, ${\ensuremath{\textsc{MopLeft}}}$, and ${\ensuremath{\textsc{Nca}}}$ in linear time (in the size of their input).
The running time of a single call to [$\textsc{Fl}$]{} might take time $O(n_T)$. Instead we will divide the calls to [$\textsc{Fl}$]{} into groups and analyze the total time used on such a group of calls. The intuition behind the division is that for a path in $P$ the calls made to [$\textsc{Fl}$]{} by [$\textsc{Emb}$]{} are done bottom up on disjoint lists of nodes in $T$.
\[lem:fl\] For disjoint ordered node lists $X_1, \ldots, X_k$ and labels $\alpha_1, \ldots, \alpha_k$, such that any node in $X_{i+1}$ is an ancestor of some node in ${\ensuremath{\textsc{Fl}}}_T(X_i, \alpha_i)$, $1 \leq
i < k$, all of ${\ensuremath{\textsc{Fl}}}_T(X_1, \alpha_1), \ldots , {\ensuremath{\textsc{Fl}}}_T(X_k,
\alpha_k)$ can be computed in $O(n_T)$ time.
Let $Z$ and $L$ be as in the implementation of the procedure. Since ${\ensuremath{\textsc{Deep}}}^*$ takes time $O({\ensuremath{| Z |}})$ and each of the steps in the for loop takes constant time, we only need to show that the total length of the lists $Z$—summed over all the calls—is $O(n_T)$ to analyze the total time usage. We will show that any node in $T$ can be in $Z$ at the beginning of the for loop at most twice during all calls to [$\textsc{Fl}$]{}.
Consider a single call to [$\textsc{Fl}$]{}. Except for the first iteration, a node can be in $Z$ only if one of its children were in $Z$ in the last iteration. Note that $Z$ is ordered at the beginning of each for loop. Thus if a node is in $Z$ at the beginning of the for loop none of its children are in $Z$Êand thus in one call to [$\textsc{Fl}$]{} a node can be in $Z$ only once.
Look at a node $z$ the first time it appears in $Z$ at the beginning of an execution of the for loop. Assume that this is in the call ${\ensuremath{\textsc{Fl}}}(X_i,\alpha_i)$. If $z \in X_i$ then $z$ cannot be in $Z$ in any later calls, since no node in $X_j$ where $j>i$ can be a descendant of a node in $X_i$. If ${\ensuremath{\mathrm{label}}}(z)\neq \alpha_i$ then $z$ cannot be in $Z$Êin any of the later calls. To see this consider the time when $z$ is removed from $Z$ (case 2 or case 3). Since the set $L$ is deep at the beginning of the for loop and $Z\subseteq L$, no descendant of $z$ will appear in $Z$ later in this call to [$\textsc{Fl}$]{}, and no node in the output from ${\ensuremath{\textsc{Fl}}}(X_i,\alpha_i)$ can be a descendant of $z$. Since any node in $X_j$, $j>i$, is an ancestor of some node in ${\ensuremath{\textsc{Fl}}}(X_i,\alpha_i)$ neither $z$ or any descendant of $z$ can be in any $X_j$, $j>i$. Thus $z$ cannot appear in $Z$ in any later calls to [$\textsc{Fl}$]{}. Now if ${\ensuremath{\mathrm{label}}}(z)=\alpha_i$ then we might have $z \in X_{i+1}$. In that case, $z$ will appear in $Z$ in the first iteration of the procedure call ${\ensuremath{\textsc{Fl}}}(X_{i+1},\alpha_i)$, but not in any later calls since the lists are disjoint, and since no node in $X_j$ where $j>i+1$ can be a descendant of a node in $X_{i+1}$. If ${\ensuremath{\mathrm{label}}}(z)=\alpha_i$ and $z \not
\in X_{i+1}$ then clearly $z$ cannot appear in $Z$ in any later call. Thus a node in $T$ is in $Z$ at the beginning of an execution of the for loop at most twice during all the calls.
Complexity of the Tree Inclusion Algorithm
------------------------------------------
Using the node list implementation of the set procedures we get:
\[lem:simpletime\] For trees $P$ and $T$ the tree inclusion problem can be solved in $O(l_Pn_T)$ time.
By Lemma \[lem:auxprocedures\] we can preprocess $T$ in $O(n_T)$ time and space. Let $g(n)$ denote the time used by ${\ensuremath{\textsc{Fl}}}$ on a list of length $n$. Consider the time used by ${\ensuremath{\textsc{Emb}}}({\ensuremath{\mathrm{root}}}(P))$. We bound the contribution for each node $v\in V(P)$. If $v$Êis a leaf we are in case 1 of [$\textsc{Emb}$]{}. The cost of computing ${\ensuremath{\textsc{Fl}}}(L(T), {\ensuremath{\mathrm{label}}}(v))$Êis $O(g(l_T))$, and by Lemma \[lem:fl\] (with $k=1$) we get $O(g(l_T))=O(n_T)$. Hence, the total cost of all leaves is $O(l_P n_T)$. If $v$ has a single child $w$ we are in case 2 of [$\textsc{Emb}$]{}, and by Lemma \[lem:auxprocedures\] the cost is $O(g(|{\ensuremath{\textsc{Emb}}}(w)|))$. If $v$ has more than one child the cost of ${\ensuremath{\textsc{MopRight}}}$, ${\ensuremath{\textsc{Nca}}}$, and ${\ensuremath{\textsc{Deep}}}$ is bounded by $\sum_{w \in {\ensuremath{\mathrm{child}}}(v)} O(|{\ensuremath{\textsc{Emb}}}(w)|)$. Furthermore, since the length of the output of ${\ensuremath{\textsc{MopRight}}}$ (and thus ${\ensuremath{\textsc{Nca}}}$) is at most $z = \min_{w \in {\ensuremath{\mathrm{child}}}(v)}
|{\ensuremath{\textsc{Emb}}}(w)|$ the cost of ${\ensuremath{\textsc{Fl}}}$ is $O(g(z))$. Hence, the total cost for internal nodes is, $$\label{eq:internal} \sum_{v \in V(P)\backslash L(P)}
O\bigg(g(\min_{w \in {\ensuremath{\mathrm{child}}}(v)} |{\ensuremath{\textsc{Emb}}}(w)|) + \sum_{w \in
{\ensuremath{\mathrm{child}}}(v)} |{\ensuremath{\textsc{Emb}}}(w)|\bigg) = \sum_{v \in V(P)}
O(g(|{\ensuremath{\textsc{Emb}}}(v)|))\;.$$ Next we bound (\[eq:internal\]). For any $w \in {\ensuremath{\mathrm{child}}}(v)$ we have that ${\ensuremath{\textsc{Emb}}}(w)$ and ${\ensuremath{\textsc{Emb}}}(v)$ are disjoint ordered lists. Furthermore we have that any node in ${\ensuremath{\textsc{Emb}}}(v)$ must be an ancestor of some node in ${\ensuremath{\textsc{Fl}}}({\ensuremath{\textsc{Emb}}}(w), {\ensuremath{\mathrm{label}}}(v))$. Hence, by Lemma \[lem:fl\], for any leaf to root path $\delta = v_1, \ldots, v_k$ in $P$, we have that $\sum_{u \in \delta} g(|{\ensuremath{\textsc{Emb}}}(u)|) = O(n_T)$. Let $\Delta$ denote the set of all root to leaf paths in $P$. It follows that, $$\sum_{v \in V(T)} g(|{\ensuremath{\textsc{Emb}}}(v)|) \leq \sum_{p \in \Delta} \sum_{u
\in p} g(|{\ensuremath{\textsc{Emb}}}(u)|) = O(l_Pn_T)\;.$$ Since this time is the same as the time spent at the leaves the time bound follows.
To analyze the space used by the algorithm we first bound the size of ${\ensuremath{\textsc{Emb}}}(v)$ for each node $v \in V(P)$. We then use this to bound the total the size of embeddings stored in the recursion stack in the computation of ${\ensuremath{\textsc{Emb}}}({\ensuremath{\mathrm{root}}}(P))$, i.e., the total size of embeddings stored by recursive calls during the computation.
\[lem:emb\_size\] For any tree $P$ we have $\forall v \in V(P)$: $${\ensuremath{| {\ensuremath{\textsc{Emb}}}_T(v) |}} \leq \frac{l_T}{l_{P(v)}}\;.$$
By Lemma \[lem:correctEmb\] ${\ensuremath{\textsc{Emb}}}(v)$ is the set of deep occurrences of $P(v)$ in $T$. By the definition of deep the occurrences are disjoint and no node in one occurrence can be an ancestor of a node in another occurrence. Each occurrence has at least $l_{P(v)}$ descendant leaves and each of these leaves is an ancestor of at least one distinct leaf in $T$ (see also Figure \[fig:MopSubtree\](b)). Thus the number of occurrences is bounded by $l_T/l_{P(v)}$.
\[lem:lin\_space\] The total size of saved embeddings on the recursion stack at any time during the computation of ${\ensuremath{\textsc{Emb}}}({\ensuremath{\mathrm{root}}}(P))$ is at most $O(l_T)$.
Let node $v$ be the node for which we are currently computing [$\textsc{Emb}$]{}. Let $p$ be the path from the root to $v$ and let $w_0,\ldots,w_\ell$ be the light nodes on this path. We have $\ell={\ensuremath{\mathrm{ldepth}}}(v)$. There is one embedding on the stack for each light node on the path (see Figure \[fig:heavypath\]): For the heavy nodes on the path there can be no saved embeddings in the recursion as the algorithm always recurses on the heavy child first. For each light node $w_i$Êon the path $p$ except the root $w_0$Êthe stack will contain either ${\ensuremath{\textsc{Emb}}}({\ensuremath{\mathrm{heavy}}}({\ensuremath{\mathrm{parent}}}(w_i)))$, or $U_j={\ensuremath{\textsc{MopRight}}}(U_{j-1},R_j)$, where $v_j$Êis $w_i$’s left sibling, or $U_j={\ensuremath{\textsc{MopLeft}}}(U_{j-1},R_j)$, where $v_j$Êis $w_i$’s right sibling. The computation of $U_j$ is a series of [$\textsc{MopRight}$]{} (or [$\textsc{MopLeft}$]{}) computations that started with the pair of node lists $({\ensuremath{\textsc{Emb}}}({\ensuremath{\mathrm{heavy}}}({\ensuremath{\mathrm{parent}}}(w_i))),{\ensuremath{\textsc{Emb}}}({\ensuremath{\mathrm{heavy}}}({\ensuremath{\mathrm{parent}}}(w_i))))$ as the first argument to [$\textsc{MopRight}$]{} (or [$\textsc{MopLeft}$]{}). As the output of [$\textsc{MopRight}$]{} (or [$\textsc{MopLeft}$]{}) can be no larger than the input to the procedure we have ${\ensuremath{| U_j |}}=O({\ensuremath{| {\ensuremath{\textsc{Emb}}}({\ensuremath{\mathrm{heavy}}}({\ensuremath{\mathrm{parent}}}(w_i)))) |}}$ and thus the total space used at any time during the recursion is $$O\big (\sum_{i=1}^{{\ensuremath{\mathrm{ldepth}}}(v)} {\ensuremath{| {\ensuremath{\textsc{Emb}}}({\ensuremath{\mathrm{heavy}}}({\ensuremath{\mathrm{parent}}}(w_i))) |}}\big )\;.$$ By Lemma \[lem:emb\_size\] we have $${\ensuremath{| {\ensuremath{\textsc{Emb}}}({\ensuremath{\mathrm{heavy}}}({\ensuremath{\mathrm{parent}}}(w_i))) |}} \leq \frac{l_T}{l_{P({\ensuremath{\mathrm{heavy}}}({\ensuremath{\mathrm{parent}}}(w_i)))}}\;,$$ and thus $$\label{eq:totalspace}
\sum_{i=1}^{{\ensuremath{\mathrm{ldepth}}}(v)} {\ensuremath{| {\ensuremath{\textsc{Emb}}}({\ensuremath{\mathrm{heavy}}}({\ensuremath{\mathrm{parent}}}(w_i))) |}}\leq l_T\sum_{i=1}^{{\ensuremath{\mathrm{ldepth}}}(v)}\frac{1}{l_{P({\ensuremath{\mathrm{heavy}}}({\ensuremath{\mathrm{parent}}}(w_i)))}}\;.$$ By the definition of heavy the node ${\ensuremath{\mathrm{heavy}}}({\ensuremath{\mathrm{parent}}}(w_i))$ has more leaves in its subtree than $w_i$, i.e., $$\label{eq:sibling}
l_{P(w_i)} \leq l_{P({\ensuremath{\mathrm{heavy}}}({\ensuremath{\mathrm{parent}}}(w_i)))}\;.$$ Obviously, ${\ensuremath{\mathrm{heavy}}}({\ensuremath{\mathrm{parent}}}(w_i))$ has no more leaves in its subtree than its parent, i.e., $$\label{eq:heavy_parent}
l_{P({\ensuremath{\mathrm{heavy}}}({\ensuremath{\mathrm{parent}}}(w_i)))}\leq l_{P({\ensuremath{\mathrm{parent}}}(w_i))}\;.$$ Since $w_i$ is light it has at most half the number of leaves in its subtree as its parent, that is $$\label{eq:half_parent}
l_{P(w_i)}\leq l_{P({\ensuremath{\mathrm{parent}}}(w_i))}/2\;.$$ Combining this with the fact that $w_i$Êis an ancestor of $w_{i+1}$ and ${\ensuremath{\mathrm{heavy}}}({\ensuremath{\mathrm{parent}}}(w_{i+1}))$ we get, $$\begin{aligned}
{2}
l_{P({\ensuremath{\mathrm{heavy}}}({\ensuremath{\mathrm{parent}}}(w_j))} &\leq l_{P({\ensuremath{\mathrm{parent}}}(w_j))} &\quad& \textrm{by } \eqref{eq:heavy_parent}\\
&\leq l_{P(w_{j-1})} && \textrm{since } w_{j-1} \textrm{ is an ancestor of } w_j \\
&\leq l_{P({\ensuremath{\mathrm{parent}}}(w_{j-1}))}/2 && \textrm{by } \eqref{eq:half_parent} \\
&\leq l_{P(w_{j-2})}/2 && \textrm{since } w_{j-2} \textrm{ is an ancestor of } {\ensuremath{\mathrm{parent}}}(w_{j-1}) \\
&\leq l_{P({\ensuremath{\mathrm{heavy}}}({\ensuremath{\mathrm{parent}}}(w_{j-2})))}/2, && \textrm{by } \eqref{eq:sibling}\end{aligned}$$ for any $2<j\leq {\ensuremath{\mathrm{ldepth}}}(v)$. Let $l_i=l_{P({\ensuremath{\mathrm{heavy}}}({\ensuremath{\mathrm{parent}}}(w_i)))}$ for all $i$. To bound the sum in we will use that $l_i \leq l_{i-2}/2$, $l_i < l_{i-1}$, and $l_{{\ensuremath{\mathrm{ldepth}}}(v)} \geq 1$. We have $$\sum_{i=1}^{{\ensuremath{\mathrm{ldepth}}}(v)}\frac{1}{l_i}\leq 2\sum_{i=2, i \textrm{ odd}}^{{\ensuremath{\mathrm{ldepth}}}(v)}\frac{1}{l_i} \leq 2 \cdot 2=4,$$ since the $l_i$’s in the last sum is decreasing with a factor of 2. Combining this with Equation we get $$\sum_{i=1}^{{\ensuremath{\mathrm{ldepth}}}(v)} {\ensuremath{| {\ensuremath{\textsc{Emb}}}({\ensuremath{\mathrm{heavy}}}({\ensuremath{\mathrm{parent}}}(w_i))) |}}\leq l_T\sum_{i=1}^{{\ensuremath{\mathrm{ldepth}}}(v)}\frac{1}{l_{P({\ensuremath{\mathrm{heavy}}}({\ensuremath{\mathrm{parent}}}(w_i)))}} \leq 4 l_T \;.$$
& &&&&&& (0,.3)[$w_0$]{}\
&&&(-1.6,0)[${\ensuremath{\mathrm{heavy}}}({\ensuremath{\mathrm{parent}}}(w_1))$]{}& &&&&& (.4,0)[$w_1$]{}\
& &&&&&& &&&&\
&&&&& &&&& & &&\
&&& &&&&& &&\
&(-1.6,0)[${\ensuremath{\mathrm{heavy}}}({\ensuremath{\mathrm{parent}}}(w_2))$]{} &&&&(.4,0)[$w_2$]{}\
&&&(.4,0)[$w_3$]{} &&&&(1.6,0)[${\ensuremath{\mathrm{heavy}}}({\ensuremath{\mathrm{parent}}}(w_3))$]{}\
& &&&&\
&&&(-1.6,0)[${\ensuremath{\mathrm{heavy}}}({\ensuremath{\mathrm{parent}}}(w_4))$]{} &&&&(.8,0)[$v=w_4$]{}
\[thm:simple\] For trees $P$ and $T$ the tree inclusion problem can be solved in $O(l_Pn_T)$ time and $O(n_T)$ space.
The time bound follows from Lemma \[lem:simpletime\]. Next consider the space used by ${\ensuremath{\textsc{Emb}}}({\ensuremath{\mathrm{root}}}(P))$. The preprocessing of Section \[sec:simplepreprocessing\] uses only $O(n_T)$ space. By Lemma \[lem:lin\_space\] the space used for the saved embeddings is $O(l_T) = O(n_T)$.
An Alternative Algorithm {#sec:alt}
------------------------
In this section we present an alternative algorithm. Since the time complexity of the algorithm in the previous section is dominated by the time used by ${\ensuremath{\textsc{Fl}}}$, we present an implementation of this procedure which leads to a different complexity. Define a *firstlabel data structure* as a data structure supporting queries of the form ${\ensuremath{\mathrm{fl}}}(v, \alpha)$, $v\in V(T)$, $\alpha \in
\Sigma$. Maintaining such a data structure is known as the *tree color problem*. This is a well-studied problem, see e.g. [@Die89; @MM1996; @FM1996; @AHR1998]. With such a data structure available we can compute ${\ensuremath{\textsc{Fl}}}$ as follows,
Return the list ${\ensuremath{\textsc{Deep}}}([{\ensuremath{\mathrm{fl}}}(X[1], \alpha),
\ldots, {\ensuremath{\mathrm{fl}}}(X[{\ensuremath{| X |}}], \alpha)])$.
\[thm:simple2\] Let $P$ and $T$ be trees. Given a firstlabel data structure using $s(n_T)$ space, $p(n_T)$ preprocessing time, and $q(n_T)$ time for queries, the tree inclusion problem can be solved in $O(p(n_T) +
l_Pl_T\cdot q(n_T))$ time and $O(s(n_T) + n_T)$ space.
Constructing the firstlabel data structures uses $O(s(n_T))$ space and $O(p(n_T))$ time. The total cost of the leaves is bounded by $O(l_p l_T\cdot q(n_T))$, since the cost of a single leaf is $O(l_T\cdot q(n_T))$. As in the proof of Theorem \[thm:simple\] we have that the total time used by the internal nodes is bounded by $\sum_{v \in V(P)} g(|{\ensuremath{\textsc{Emb}}}(v)|)$, where $g(n)$ is the time used by ${\ensuremath{\textsc{Fl}}}$ on a list of length $n$, that is, $g(n)\leq n\cdot q(n_T)$. By Lemma \[lem:auxprocedures\] and Lemma \[lem:lin\_space\] for any leaf to root path $\delta = v_1, \ldots, v_k$ in $P$, we have that $\sum_{u \in \delta} |{\ensuremath{\textsc{Emb}}}(u)| \leq O(l_T)$. Let $\Delta$ denote the set of all root to leaf paths in $P$. It follows that, $$\sum_{v \in V(P)} g(|{\ensuremath{\textsc{Emb}}}(v)|) \leq \sum_{p \in \Delta} \sum_{u
\in p} g(|{\ensuremath{\textsc{Emb}}}(u)|) \leq \sum_{p \in \Delta}O(l_T \cdot q(n_T))\leq O(l_Pl_T \cdot q(n_T)) .$$ Since this time is the same as the time spent at the leaves the time bound follows.
Several firstlabel data structures are available, for instance, if we want to maintain linear space we have,
\[lem:dietz\] For any tree $T$ there is a data structure using $O(n_T)$ space, $O(n_T)$ expected preprocessing time which supports firstlabel queries in $O(\log \log n_T)$ time.
The expectation in the preprocessing time is due to perfect hashing. Since our data structure does not need to support efficient updates we can remove the expectation by using the deterministic dictionary of Hagerup et al. [@HMP2001]. This gives a worst-case preprocessing time of $O(n_T \log n_T)$. However, using a simple two-level approach this can be reduced to $O(n_T)$ (see e.g. [@Thorup2003]). Plugging in this data structure we obtain,
\[cor:simple\] For trees $P$ and $T$ the tree inclusion problem can be solved in $O(l_Pl_T\log\log n_T + n_T)$ time and $O(n_T)$ space.
A Faster Tree Inclusion Algorithm {#micromacro}
=================================
In this section we present a new tree inclusion algorithm which has a worst-case subquadratic running time. As discussed in the introduction the general idea is to divide $T$ into clusters of logarithmic size which we can efficiently preprocess and then use this to speed up the computation with a logarithmic factor.
Clustering
----------
In this section we describe how to divide $T$ into clusters and how the macro tree is created. For simplicity in the presentation we assume that $T$ is a binary tree. If this is not the case it is straightforward to construct a binary tree $B$, where $n_{B} \leq
2n_T$, and a mapping $g : V(T) \rightarrow V(B)$ such that for any pair of nodes $v,w \in V(T)$, ${\ensuremath{\mathrm{label}}}(v) = {\ensuremath{\mathrm{label}}}(g(v))$, $v \prec w$ iff $g(v) \prec g(w)$, and $v \lhd w$ iff $g(v) \lhd g(w)$. The nodes in the set $U = V(B)\backslash \{g(v) \mid v \in V(T)\}$ are assigned a special label $\beta \not\in \Sigma$. It follows that for any tree $P$, $P \sqsubseteq T$ iff $P \sqsubseteq B$.
Let $C$ be a connected subgraph of $T$. A node in $V(C)$ adjacent to a node in $V(T)\backslash V(C)$ is a *boundary* node. The boundary nodes of $C$ are denoted by $\delta C$. We have ${\ensuremath{\mathrm{root}}}(T) \in \delta C$ if ${\ensuremath{\mathrm{root}}}(T) \in
V(C)$. A *cluster* of $C$ is a connected subgraph of $C$ with at most two boundary nodes. A set of clusters $CS$ is a *cluster partition* of $T$ iff $V(T) = \cup_{C\in CS} V(C)$, $E(T) = \cup_{C\in CS} E(C)$, and for any $C_1 ,C_2 \in CS$, $E(C_1) \cap E(C_2) = \emptyset$, $|E(C_1)| \geq 1$. If $|\delta C| = 1$ we call $C$ a *leaf cluster* and otherwise an *internal cluster*.
We use the following recursive procedure ${\ensuremath{\textsc{Cluster}}}_T(v,s)$, adopted from [@AR2002c], which creates a cluster partition $CS$ of the tree $T(v)$ with the property that $|CS| = O(s)$ and $|V(C)| \leq {\left\lceil{n_T/s}\right\rceil}$ for each $C \in CS$. A similar cluster partitioning achieving the same result follows from [@AHT2000; @AHLT1997; @Frederickson1997].
For each child $u$ of $v$ there are two cases:
1. $|V(T(u))| + 1 \leq {\left\lceil{n_T/s}\right\rceil}$. Let the nodes $\{v\} \cup V(T(u))$ be a leaf cluster with boundary node $v$.
2. $|V(T(u))| \geq {\left\lceil{n_T/s}\right\rceil}$. Pick a node $w \in V(T(u))$ of maximum depth such that $|V(T(u))| + 2 - |V(T(w))| \leq {\left\lceil{n_T/s}\right\rceil}$. Let the nodes $V(T(u)) \backslash V(T(w)) \cup \{v,w\}$ be an internal cluster with boundary nodes $v$ and $w$. Recursively, compute ${\ensuremath{\textsc{Cluster}}}_T(w, s)$.
\[lem:clustering\] Given a tree $T$ with $n_T>1$ nodes, and a parameter $s$, where ${\left\lceil{n_T/s}\right\rceil} \geq 2$, we can build a cluster partition $CS$ in $O(n_T)$ time, such that $|CS| = O(s)$ and $|V(C)| \leq
{\left\lceil{n_T/s}\right\rceil}$ for any $C \in CS$.
The procedure ${\ensuremath{\textsc{Cluster}}}_T({\ensuremath{\mathrm{root}}}(T), s)$ clearly creates a cluster partition of $T$ and it is straightforward to implement in $O(n_T)$ time. Consider the size of the clusters created. There are two cases for $u$. In case $1$, $|V(T(u))| + 1 \leq
{\left\lceil{n_T/s}\right\rceil}$ and hence the cluster $C = \{v\} \cup V(T(u))$ has size $|V(C)| \leq {\left\lceil{n_T/s}\right\rceil}$. In case $2$, $|V(T(u))| + 2 -
|V(T(w))| \leq {\left\lceil{n_T/s}\right\rceil}$ and hence the cluster $C = V(T(u))
\backslash V(T(w)) \cup \{v,w\}$ has size $|V(C)| \leq
{\left\lceil{n_T/s}\right\rceil}$.
Next consider the size of the cluster partition. Let $c=
{\left\lceil{n_T/s}\right\rceil}$. We say that a cluster $C$ is *bad* if ${\ensuremath{| V(C) |}} \leq c/2$ and *good* otherwise. We will show that at least a constant fraction of the clusters in the cluster partition are good. It is easy to verify that the cluster partition created by procedure [$\textsc{Cluster}$]{} has the following properties:
- Let $C$ be a bad internal cluster with boundary nodes $v$ and $w$ ($v \prec
w$). Then $w$ has two children with at least $c/2$ descendants each.
- Let $C$ be a bad leaf cluster with boundary node $v$. Then the boundary node $v$ is contained in a good cluster.
By (ii) the number of bad leaf cluster is at most twice the number of good internal clusters and by (i) each bad internal cluster has two child clusters. Therefore, the number of bad internal clusters is bounded by the number of leaf clusters. Let $b_i$ and $g_i$ denote the number of bad and good internal clusters, respectively, and let $b_l$ and $g_l$ denote the number of bad and good leaf clusters, respectively. We have $$b_i \leq b_l + g_l \leq 2 g_i + g_l,$$ and therefore the number of bad clusters is bounded by $$b_l + b_i \leq 2 g_i + g_l + 2 g_i = 4g_i + g_l \;.$$ Thus the number of bad clusters is at most 4 times the number of good clusters, and therefore at most a constant fraction of the total number of clusters. Since a good cluster is of size more than $c/2$, there can be at most $2s$ good clusters and thus ${\ensuremath{| CS |}}=O(s)$.
Let $C\in CS$ be an internal cluster with $v, w \in \delta C$. The *spine path* of $C$ is the path between $v,w$ excluding $v$ and $w$. A node on the spine path is a *spine node*. A node to the left and right of $v$ or of any node on the spine path is a *left node* and *right node*, respectively. If $C$ is a leaf cluster with $v \in \delta C$ then any proper descendant of $v$ is a *leaf node*.
&&& (0,0)(.2,.2)(0,.4)[$v$]{} &&&&&&&&& (0,.4)[$v$]{}\
&& & &&&&&&\
& & (-.3,0)(.6,1.2) & (0,0)(.3,1.2) (0,0)(2,1.45) & (.3,-.3)(.6,.8)&&&&&&&& (.6,.3)[$s(v,w)$]{}\
& & & & & &&&&& (-.2,.4)[$l(v,w)$]{} &&&& (.2,.4)[$r(v,w)$]{}\
& && (0,0)(.2,.2)(0,-.4)[$w$]{} &&&&&&&&& (0,-.4)[$w$]{}\
&&& (0,-.3)[(a)]{} &&&&&&&&& (0,-.3)[(b)]{}\
\
&&& (0,0)(.2,.2)(0,.4)[$v$]{} &&&&&&&&& (0,.4)[$v$]{}\
&& & (0,-.3)(1,.8)(0,-.3)(1.5,1.06) &&&&&&\
&& & & &&&&&&&& (0,-.4)[$l(v)$]{}\
&&& (0,-.4)[(c)]{} &&&&&&&&& (0,-.4)[(d)]{}\
Let $CS$ be a cluster partition of $T$ as described in Lemma \[lem:clustering\]. We define an ordered *macro tree* $M$. Our definition of $M$ may be viewed as an ”ordered” version of the macro tree defined in [@AR2002c]. The node set $V(M)$ consists of the boundary nodes in $CS$. Additionally, for each internal cluster $C \in CS$ with $v,w \in \delta C$, $v \prec w$, we have the nodes $s(v,w)$, $l(v,w)$ and $r(v,w)$ and edges $(v, s(v,w)), (s(v,w),l(v,w)), (s(v,w), w)$, and $(s(v,w),r(v,w))$. That is, the nodes $l(v,w)$, $r(v,w)$ and $w$ are all children of $s(v,w)$. The nodes are ordered so that $l(v,w) \lhd w \lhd r(v,w)$. For each leaf cluster $C$, $v \in \delta C$, we have the node $l(v)$ and edge $(v,l(v))$. Since ${\ensuremath{\mathrm{root}}}(T)$ is a boundary node, $M$ is rooted at ${\ensuremath{\mathrm{root}}}(T)$. Figure \[clusterexample\] illustrates these definitions.
With each node $v \in V(T)$ we associate a unique macro node denoted $c(v)$. Let $u \in V(C)$, where $C \in CS$. $$c(u) =
\begin{cases}
u & \text{if $u$ is boundary node}, \\
l(v) & \text{if $u$ is a leaf node and $v \in \delta C$}, \\
s(v,w) & \text{if $u$ is a spine node, $v,w \in \delta C$, and $v\prec w$}, \\
l(v,w) & \text{if $u$ is a left node, $v,w \in \delta C$, and $v\prec w$}, \\
r(v,w) & \text{if $u$ is a right node, $v,w \in \delta C$, and $v\prec w$}.
\end{cases}$$
Conversely, for any macro node $i \in V(M)$ define the *micro forest*, denoted $C(i)$, as the induced subgraph of $T$ of the set of nodes $\{v \mid v \in V(T), i =
c(v)\}$. We also assign a *set* of labels to $i$ given by ${\ensuremath{\mathrm{label}}}(i) = \{{\ensuremath{\mathrm{label}}}(v) \mid v \in V(C(i))\}$. If $i$ is a spine node or a boundary node the unique node in $V(C(i))$ of greatest depth is denoted by ${\ensuremath{\mathrm{first}}}(i)$. Finally, for any set of nodes $\{i_1,
\ldots, i_k\} \subseteq V(M)$ we define $C(i_1, \ldots, i_k)$ as the induced subgraph of the set of nodes $V(C(i_1)) \cup \cdots \cup V(C(i_k))$.
The following propositions state useful properties of ancestors, nearest common ancestor, and the left-to-right ordering in the micro forests and in $T$. The propositions follow directly from the definition of the clustering. See also Figure \[fig:propositions\].
\[lem:ancestorlemma\] For any pair of nodes $v, w \in V(T)$, the following hold
- If $c(v) = c(w)$ then $v \prec_T w$ iff $v \prec_{C(c(v))} w$.
- If $c(v) \neq c(w)$, and for some boundary nodes $v',w'$ we have $c(v) =s(v',w')$, and $c(w)
\in \{l(v', w'), r(v',w')\}$, then $v \prec_T w$ iff $v \prec_{C(c(w),s(v',w'), v')} w$.
- In all other cases, $v \prec_T w$ iff $c(v) \prec_{M} c(w)$.
Case (i) says that if $v$ and $w$ belong to the same macro node then $v$ is an ancestor of $w$ iff $v$ is an ancestor of $w$ in the micro forest for that macro node. Case (ii) says that if $v$ is a spine node and $w$ is a left or right node in the same cluster then $v$ is an ancestor of $w$ iff $v$ is an ancestor of $w$ in the micro tree induced by that cluster (Figure \[fig:propositions\](a)). Case (iii) says that in all other cases $v$ is an ancestor of $w$ iff the macro node $v$ belongs to is an ancestor of the macro node $w$ belongs to in the macro tree.
&&& (-.3,0)[$v'$]{}\
&& & (.2,0)[$v$]{}\
& & (-.29,0)(.68,1.2) & (0,.1)(.4,1.2) (-.63,.25)(1.28,1.8) &\
& & (-.3,0)[$w$]{} & & &\
& && (0,-.4)[$w'$]{}\
&&& (0,-.3)[(a)]{}\
\
&&& (-.3,0)[$v'$]{}\
&& &\
& (.2,0)[$v$]{}& (-.29,0)(.68,1.2) & (-.63,.25)(1.28,1.8) &\
& & (-.3,0)[$w$]{} & & &\
& && (0,-.4)[$w'$]{}\
&&& (0,-.3)[(b)]{}\
\
&&& (-.3,0)[$v'$]{}\
&&(.3,0)[$v$]{} & (0,-.35)(1.1,1)(0,-.35)(1.4,1.6)\
&& & (.3,0)[$w$]{} & &\
\
\
&&& (0,-.4)[(c)]{}\
\
&&& (-.3,0)[$v'$]{}\
&& (-.3,0)[$v$]{} &\
& & (-.29,0)(.68,1.2) & (.2,.1)[$w$]{} (0,.1)(.4,1.2) (-.63,.25)(1.28,1.8) &\
& & & & &\
& && (0,-.4)[$w'$]{}\
&&& (0,-.3)[(d)]{}\
\
&&& (.3,0)[$v'$]{}\
&& &\
& & & (0,.1)(.4,1.2) (.7,.25)(1.32,1.85) & (.4,-.34)(.7,.8) (.3,0)[$w$]{}\
& & & (.15,.15)[$v$]{} & & &&\
& && (0,-.4)[$w'$]{}\
&&& (0,-.3)[(e)]{}\
&&& (.3,0)[$v'$]{}\
&& &\
& & (-.29,0.1)(.68,1.2) & (0,0.2)(2.3,1.6) & (.4,-.3)(.7,.8) (.3,0)[$w$]{}\
& & (-.3,0)[$v$]{}& & & &&\
& && (0,-0.4)[$w'$]{}\
&&& (0,-.3)[(f)]{}\
&&& (-.3,0)[$v'$]{}\
&& (-.3,0)[$v$]{} &\
& & (-.29,0.05)(.68,1.1) & (-.63,.35)(1.28,1.6) &\
& & & (-.3,0)[$w'$]{}\
& & & (.3,0)[$w$]{}\
&& &&\
&&& (0,-.3)[(g)]{}\
\[lem:orderlemma\] For any pair of nodes $v, w \in V(T)$, the following hold
- If $c(v) = c(w) \in \{r(v',w'),l(v',w')\}$ for some boundary nodes $v',w'$, then $v \lhd w$ iff $v \lhd_{C(c(v),v',s(v',w'))}
w$.
- If $c(v) = c(w)=l(v')$ for some boundary node $v'$, then $v \lhd w$ iff $v \lhd_{C(c(v),v')} w$.
- If $c(v) = l(v', w')$ and $c(w)=s(v',w')$ for some boundary nodes $v',w'$, then $v \lhd w$ iff $v \lhd_{C(c(v), c(w), v')} w$.
- If $c(v)=s(v',w')$ and $c(w) = r(v', w')$ for some boundary nodes $v',w'$, then $v \lhd w$ iff $v \lhd_{C(c(v), c(w), v')} w$.
- In all other cases, $v \lhd w$ iff $c(v) \lhd_{M} c(w)$.
Case (i) says that if $v$ and $w$ are both either left or right nodes in the same cluster then $v$ is to the left of $w$ iff $v$ is to the left of $w$ in the micro tree induced by their macro node together with the spine and top boundary node of the cluster (Figure \[fig:propositions\](b)). Case (ii) says that if $v$ and $w$ are both leaf nodes in the same cluster then $v$ is to the left of $w$ iff $v$ is to the left of $w$ in the micro tree induced by that leaf cluster (Figure \[fig:propositions\](c)). Case (iii) says that if $v$ is a left node and $w$ is a spine node in the same cluster then $v$ is to the left of $w$ iff $v$ is to the left of $w$ in the micro tree induced by their two macro nodes and the top boundary node of the cluster (Figure \[fig:propositions\](d)). Case (iv) says that if $v$ is a spine node and $w$ is a right node in the same cluster then $v$ is to the left of $w$ iff $v$ is to the left of $w$ in the micro tree induced by their two macro nodes and the top boundary node of the cluster (Figure \[fig:propositions\](e)). In all other cases $v$ is to the left of $w$ if the macro node $v$ belongs to is to the left of the macro node of $w$ in the macro tree (Case (v)).
\[lem:ncalemma\] For any pair of nodes $v, w \in V(T)$, the following hold
- If $c(v) = c(w)=l(v')$ for some boundary node $v'$, then ${\ensuremath{\mathrm{nca}}}_T(v,w)={\ensuremath{\mathrm{nca}}}_{C(c(v),v')}(v,w)$.
- If $c(v) = c(w)\in\{l(v',w'),r(v',w')\}$ for some boundary nodes $v',w'$, then\
${\ensuremath{\mathrm{nca}}}_T(v,w)={\ensuremath{\mathrm{nca}}}_{C(c(v), s(v',w'), v')}(v,w)$.
- If $c(v) = c(w)=s(v',w')$ for some boundary nodes $v',w'$, then ${\ensuremath{\mathrm{nca}}}_T(v,w)={\ensuremath{\mathrm{nca}}}_{C(c(v))}(v,w)$.
- If $c(v) \neq c(w)$ and $c(v),c(w)\in\{l(v',w'),r(v',w'),s(v',w')\}$ for some boundary nodes $v',w'$, then\
${\ensuremath{\mathrm{nca}}}_T(v,w)={\ensuremath{\mathrm{nca}}}_{C(c(v),c(w),s(v',w'),v')}(v,w)$.
- If $c(v) \neq c(w)$, $c(v)\in\{l(v',w'),r(v',w'),s(v',w')\}$, and $w' \preceq_{M}
c(w)$ for some boundary nodes $v',w'$, then ${\ensuremath{\mathrm{nca}}}_T(v,w)={\ensuremath{\mathrm{nca}}}_{C(c(v),s(v',w'),v',w')}(v,w')$.
- In all other cases, ${\ensuremath{\mathrm{nca}}}_T(v,w)={\ensuremath{\mathrm{nca}}}_{M}(c(v),c(w))$.
Case (i) says that if $v$ and $w$ are leaf nodes in the same cluster then the nearest common ancestor of $v$ and $w$ is the nearest common ancestor of $v$ and $w$ in the micro tree induced by that leaf cluster (Figure \[fig:propositions\](c)). Case (ii) says that if $v$ and $w$ are both either left nodes or right nodes then the nearest common ancestor of $v$ and $w$ is the nearest common ancestor in the micro tree induced by their macro node together with the spine and top boundary node of the cluster (Figure \[fig:propositions\](b)). Case (iii) says that if $v$ and $w$ are both spine nodes in the same cluster then the nearest common ancestor of $v$ and $w$ is the nearest common ancestor of $v$ and $w$ in the micro tree induced by their macro node. Case (iv) says that if $v$ and $w$ are in different macro nodes but are right, left, or spine nodes in the same cluster then the nearest common ancestor of $v$ and $w$ is the nearest common ancestor of $v$ and $w$ in the micro tree induced by that cluster (we can omit the bottom boundary node) (Figure \[fig:propositions\](f)). Case (v) says that if $v$ is a left, right, or spine node, and the bottom boundary node $w'$ of $v$’s cluster is an ancestor in the macro tree of the macro node containing $w$, then the nearest common ancestor of $v$ and $w$ is the nearest common ancestor of $v$ and $w'$ in the micro tree induced by the macro node of $v$, the spine node, and the top and bottom boundary nodes of $v$’s cluster (Figure \[fig:propositions\](g)). In all other cases the nearest common ancestor of $v$ and $w$ is the nearest common ancestor of their macro nodes in the macro tree (Case (vi)).
Preprocessing {#sec:preprocessing}
-------------
In this section we describe how to preprocess $T$. First build a cluster partition $CS$ of the tree $T$ with clusters of size $s$, to be fixed later, and the corresponding macro tree $M$ in $O(n_T)$ time. The macro tree is preprocessed as in Section \[sec:simplepreprocessing\]. However, since nodes in $M$ contain a set of labels, we now store a dictionary for ${\ensuremath{\mathrm{label}}}(v)$ for each node $v \in V(M)$. Using the deterministic dictionary of Hagerup et al. [@HMP2001] all these dictionaries can be constructed in $O(n_{T}\log n_{T})$ time and $O(n_{T})$ space. Furthermore, we extend the definition of ${\ensuremath{\mathrm{fl}}}$ such that ${\ensuremath{\mathrm{fl}}}_{M}(v, \alpha)$ is the nearest ancestor $w$ of $v$ such that $\alpha \in {\ensuremath{\mathrm{label}}}(w)$.
Next we show how to preprocess the micro forests. For any cluster $C\in CS$, deep sets $X, Y, Z\subseteq V(C)$, $i \in \mathbb{N}$, and $\alpha \in \Sigma$ define the following procedures.
Return the number of nodes in $X$.
Return the leftmost $i$ nodes in $X$.
Return the rightmost $i$ nodes in $X$.
Return all nodes of $X$ to the left of the leftmost node in $Y$.
Return all nodes of $X$ to the right of the rightmost node in $Y$.
where $X=\{m_1 \lhd \cdots \lhd m_k\}$, $Y=\{v_1 \lhd \cdots \lhd v_k\}$, and $Z \subseteq Y$. Return $R:=\{m_{j} \mid v_j \in Z \}$.
Return the pair $(R_1,R_2)$, where $R_1={\ensuremath{\mathop{{\ensuremath{\mathrm{mop}}}(X,Y)|}_{1}}}$ and $R_2={\ensuremath{\mathop{{\ensuremath{\mathrm{mop}}}(X,Y)|}_{2}}}$.
If we want to specify that a procedure applies to a certain cluster $C$ we add the subscript $C$. In addition to these procedures we also define the set procedures on clusters, that is, ${\ensuremath{\textsc{parent}}}$, ${\ensuremath{\textsc{nca}}}$, ${\ensuremath{\textsc{deep}}}$, and ${\ensuremath{\textsc{fl}}}$, as in Section \[sec:recursion\]. Collectively, we will call these the *cluster procedures*. We represent the input and output sets in the procedures as bit strings indexed by preorder numbers. Specifically, a subset $X$ in a cluster $C$ is given by a bit string $b_{1} \ldots b_{s}$, such that $b_{i} = 1$ iff the $i$th node in a preorder traversal of $C$ is in $X$. If $C$ contains fewer than $s$ nodes we leave the remaining values undefined.
The procedure ${\ensuremath{\textsc{size}}}(X)$ is the number of ones in the bit string. The procedure ${\ensuremath{\textsc{left}}}(i,X)$ corresponds to setting all bits in $X$ larger than the $i$th set bit to zero. Similarly, ${\ensuremath{\textsc{right}}}(i,X)$ corresponds to setting all bits smaller than the $i$th largest set bit to zero. Similarly, the procedures ${\ensuremath{\textsc{leftof}}}(X,Y)$, ${\ensuremath{\textsc{rightof}}}(X,Y)$, ${\ensuremath{\textsc{mop}}}(X,Y)$, and ${\ensuremath{\textsc{match}}}(X,Y,Z)$ only depend on the preorder of the nodes and thus only on the bit string and not any other information about the cluster.
Next we show how to implement the cluster procedures efficiently. We precompute the value of all procedures, except ${\ensuremath{\textsc{fl}}}$, for all possible inputs and clusters. By definition, these procedures do not depend on any specific labeling of the nodes in the cluster. Hence, it suffices to precompute the value for all rooted, ordered trees with at most $s$ nodes. The total number of these is less than $2^{2s}$ (consider e.g. an encoding using balanced parenthesis). Furthermore, the number of possible input sets is at most $2^{s}$. Since at most $3$ sets are given as input to a cluster procedure, it follows that we can tabulate all solutions using less than $2^{3s}\cdot 2^{2s} = 2^{5s}$ bits of memory. Hence, choosing $s \leq 1/10\log n$ we use $O(2^{\frac{1}{2}\log n}) = O(\sqrt{n})$ bits. Using standard bit wise operations each solution is easily implemented in $O(s)$ time giving a total time of $O(\sqrt{n}\log n)$.
Since the procedure ${\ensuremath{\textsc{fl}}}$ depends on the alphabet, which may be of size $n_{T}$, we cannot efficiently apply the same trick as above. Instead define for any cluster $C \in CS$, $X \subseteq V(C)$, and $\alpha \in \Sigma$:
Return the set $\{x \mid \text{$x$ is an ancestor of a node in $X$}\}$.
Return the set $\{x \mid x \in V(C), {\ensuremath{\mathrm{label}}}(x) = \alpha\}$.
Clearly, ${\ensuremath{\textsc{ancestor}}}$ can be implemented as above. For ${\ensuremath{\textsc{eq}}}_{C}$ note that the total number of distinct labels in $C$ is at most $s$. Hence, ${\ensuremath{\textsc{eq}}}_{C}$ can be stored in a dictionary with at most $s$ entries each of which is a bit string of length $s$. Thus, (using again the result of [@HMP2001]) the total time to build all such dictionaries is $O(n_{T}\log n_{T})$.
By the definition of ${\ensuremath{\textsc{fl}}}$ we have that, $${\ensuremath{\textsc{fl}}}_C(X,\alpha) = {\ensuremath{\textsc{deep}}}_C({\ensuremath{\textsc{ancestor}}}_C(X) \cap {\ensuremath{\textsc{eq}}}_C(\alpha)).$$ Since intersection can be implemented using a binary *and*-operation, ${\ensuremath{\textsc{fl}}}_{C}(X, \alpha)$ can be computed in constant time. Later, we will also need to compute union of sets represented as bit strings and we note that this can be done using a binary *or*-operation.
To implement the set procedures in the following section we often need to “restrict” the cluster procedures to work on a subtree of a cluster. Specifically, for any set of macro nodes $\{i_{1}, \ldots, i_{k}\}$ in the *same* cluster $C$ (hence, $k \leq 5$), we will replace the subscript $C$ with $C(i_{1}, \ldots, i_{k})$. For instance, ${\ensuremath{\textsc{parent}}}_{C(s(v,w), l(v,w))}(X) = \{{\ensuremath{\mathrm{parent}}}(x) \mid x\in X \cap V(C(s(v,w), l(v,w))\} \cap V(C(s(v,w), l(v,w))$. To implement all restricted versions of the cluster procedures, we compute for each cluster $C \in CS$ a bit string representing the set of nodes in each micro forest. Clearly, this can be done in $O(n_{T})$ time. Since there are at most $5$ micro forests in each cluster it follows that we can compute any restricted version using an additional constant number of and-operations.
Note that the total preprocessing time and space is dominated by the construction of deterministic dictionaries which use $O(n_{T}\log n_{T})$ time and $O(n_T)$ space.
Implementation of the Set Procedures {#implementation-of-the-set-procedures}
------------------------------------
Using the preprocessing from the previous section we show how to implement the set procedures in sublinear time. First we define a compact representation of node sets. Let $T$ be a tree with macro tree $M$. For simplicity, we identify nodes in $M$ with a number almost equal to their preorder number, which we denote their *macro tree number*: All nodes nodes except spine and left nodes are identified with their preorder number. Spine nodes are identified with their preorder number + 1 if they have a left node as a child and with their preorder number otherwise, and left nodes are identified with their preorder number - 1. Hence, we swap the order of left and spine nodes in the macro tree numbering. We will explain the reason for using macro tree numbers below.
Let $S \subseteq V(T)$ be any subset of nodes of $T$. A *micro-macro node array* (abbreviated node array) $X$ representing $S$ is an array of size $n_{M}$. The $i$th entry, denoted $X[i]$, represents the subset of nodes in $C(i)$, that is, $X[i] = V(C(i)) \cap S$. The set $X[i]$ is encoded using the same bit representation as in Section \[sec:preprocessing\]. By our choice of parameter in the clustering the space used for this representation is $O(n_{T}/\log n_{T})$.
We can now explain the reason for using macro tree numbers to identify the nodes instead of preorder numbers. Consider a node array representing a deep set. If a left node and the corresponding spine node both are non-empty, then all nodes in the left node are to the left of the node in the spine node. Formally,
\[prop:macrotreeorder\] Consider a node array $X$ representing a deep set ${\ensuremath{\mathcal{X}}}$. For any pair of nodes $v, w \in {\ensuremath{\mathcal{X}}}$, such that $v\in X[i]$ and $w\in X[j]$, $i\neq j$, we have $$v \lhd w \Leftrightarrow i<j \;.$$
By Proposition \[lem:orderlemma\](v) the claim is true for $i \lhd j$. The remaining cases are $i=l(v',w')$Êand $j=s(v',w')$ (Proposition \[lem:orderlemma\] (iii)) and $i=s(v',w')$Êand $j=r(v',w')$ (Proposition \[lem:orderlemma\](iv)). In both cases $i<j$ and it follows immediately that $v \lhd w \Rightarrow i <j$. For the other direction, it follows from the structure of the macro tree that in both cases either $v \lhd w$Êor $w \prec v$. But ${\ensuremath{\mathcal{X}}}$ is deep and thus $v \lhd w$.
Thus, by using macro tree numbers we encounter the nodes in $X$ according to their preorder number in the original tree $T$. This simplifies the implementation of all the procedures except [$\textsc{Deep}$]{}, since they all get deep sets as input.
We now present the detailed implementation of the set procedures on node arrays. As in Section \[simple\] we assume that the input to all of the procedures, except [$\textsc{Deep}$]{}, represent a deep set. Let $X$ be a node array.
Initialize an empty node array $R$ of size $n_{M}$ ($R[i]:=\emptyset$ for $i=1,\ldots n_M$) and set $i:=1$.
Repeat until $i>n_M$:
- While $X[i]= \emptyset$ do $i:=i+1$.
There are three cases depending on the type of $i$:
1. $i\in \{l(v,w), r(v,w)\}$. Compute $$N := {\ensuremath{\textsc{parent}}}_{C(i, s(v,w), v)}(X[i])\;.$$ For each $j \in \{i, s(v,w), v\}$, set $R[j] := R[j] \cup (N \cap V(C(j)))$.
2. $i = l(v)$. Compute $$N := {\ensuremath{\textsc{parent}}}_{C(i, v)}(X[i])\;.$$
For each $j \in \{i, v\}$, set $R[j] :=R[j] \cup (N \cap V(C(j)))$.
3. $i \not\in \{l(v,w), r(v,w), l(v)\}$. Compute $$N := {\ensuremath{\textsc{parent}}}_{C(i)}(X[i])\;.$$
If $N \neq \emptyset$ set $R[i]:=R[i] \cup N$. Otherwise, if $j := {\ensuremath{\mathrm{parent}}}_{M}(i) \neq \bot$ set $R[j] := R[j] \cup \{{\ensuremath{\mathrm{first}}}(j)\}$.
Set $i:=i+1$.
Return $R$.
Procedure ${\ensuremath{\textsc{Parent}}}$ has three cases. Case 1 handles the fact that left or right nodes may have a node on a spine or the top boundary node as parent. Since no left or right nodes can have their parent outside their cluster there is no need to compute parents in the macro tree. Case 2 handles the fact that a leaf node may have the boundary node as parent. Since no leaf node can have its parent outside its cluster there is no need to compute parents in the macro tree. Case 3 handles boundary and spine nodes. In this case there is either a parent within the micro forest or we can use the macro tree to compute the parent of the root of the micro tree. Since the input to [$\textsc{Parent}$]{} is deep we only need to do one of the two things. If the computation of parent in the macro tree returns a nonempty set, this set is added to the output. Otherwise (the returned set is empty), we compute parent of $i$Êin the macro tree. If the computation of parent in the macro tree returns a node $j$, this will either be a spine node or a boundary node. To take care of the case where $j$ is a spine node, we add the lowest node (${\ensuremath{\mathrm{first}}}(j)$) in $j$ to the output. If $j$Êis a boundary node this is just $j$Êitself. We now give the implementation of procedure [$\textsc{Nca}$]{}. The input to procedure [$\textsc{Nca}$]{} is two node arrays $X$ and $Y$ representing two subsets ${\ensuremath{\mathcal{X}}}, {\ensuremath{\mathcal{Y}}} \subseteq V(T)$, ${\ensuremath{| {\ensuremath{\mathcal{X}}} |}}={\ensuremath{| {\ensuremath{\mathcal{Y}}} |}}=k$. The output is a node array $R$ representing the set ${\ensuremath{\textsc{Deep}}}(\{{\ensuremath{\mathrm{nca}}}({\ensuremath{\mathcal{X}}}_i,{\ensuremath{\mathcal{Y}}}_i) \mid 1\leq i\leq k\})$, where ${\ensuremath{\mathcal{X}}}_i$ and ${\ensuremath{\mathcal{Y}}}_i$ is the $i$th element of ${\ensuremath{\mathcal{X}}}$ and ${\ensuremath{\mathcal{Y}}}$, wrt. their preorder number in the tree, respectively. We also assume that we have ${\ensuremath{\mathcal{X}}}_i \lhd {\ensuremath{\mathcal{Y}}}_i$ for all $i$ (since ${\ensuremath{\textsc{Nca}}}$ is always called on a set of minimum ordered pairs). Note, that $\mathcal{X}_l $ and $\mathcal{Y}_l$ can belong to different clusters/nodes in the macro tree, i.e., we might have $\mathcal{X}_l \in X[i]$ and $\mathcal{Y}_l \in Y[j]$ where $i\neq j$.
Initialize an empty node array $R$ of size $n_{M}$, set $i:=1$ and $j:=1$.
Repeat until $i>n_M$ or $j>n_M$:
- While $X[i]= \emptyset$ do $i:=i+1$. While $Y[j] = \emptyset$ do $j:=j+1$. Set $n:=\min({\ensuremath{\textsc{size}(X[i])}},{\ensuremath{\textsc{size}(Y[j])}})$, $X_i:={\ensuremath{\textsc{left}}}(n,X[i])$, and $Y_j:={\ensuremath{\textsc{left}}}(n,Y[j])$.
Compare $i$ and $j$. There are two cases:
1. $i=j$. Set $$S:=
\begin{cases}
C(i,v), & \textrm{if } i=l(v),\\
C(i, s(v,w), v), & \textrm{if } i\in\{l(v,w),r(v,w)\}.
\end{cases}$$ Compute $N:={\ensuremath{\textsc{nca}}}_S(X_i,Y_j)$.
For each macro node $h$ in $S$ set $R[h]:=R[h] \cup (N\cap V(C(h)))$.
2. $i \neq j$. Compute $h:={\ensuremath{\textsc{nca}}}_M(i,j)$. There are two subcases:
1. $h$ is a boundary node. Set $R[h]:=1$.
2. $h$ is a spine node $s(v,w)$. There are three cases:
1. $i \in \{l(v,w),s(v,w)\}$ and $j\in \{s(v,w),r(v,w)\}$.
Compute $N:={\ensuremath{\textsc{nca}}}_{C(i,j,s(v,w),v)}(X_i,Y_j)$.
2. $i=l(v,w)$ and $w \preceq j$.
Compute $N:={\ensuremath{\textsc{nca}}}_{C(i,s(v,w),v,w)}({\ensuremath{\textsc{right}}}(1,X_i),w)$.
3. $j=r(v,w)$ and $w \preceq i$.
Compute $N:={\ensuremath{\textsc{nca}}}_{C(j,s(v,w),w,v)}(w,{\ensuremath{\textsc{left}}}(1,Y_j))$.
Set $R[h]:=R[h]\cup (N\cap V(C(h)))$ and $R[v]:=R[v]\cup (N\cap V(C(v)))$.
Set $X[i]:=X[i]\setminus X_i$ and $Y[j]:=Y[j]\setminus Y_j$.
Return ${\ensuremath{\textsc{Deep}}}(R)$.
In procedure [$\textsc{Nca}$]{} we first find the next non-empty entries in the node arrays $X[i]$Êand $Y[j]$. We then compare the sizes of $X[i]$Êand $Y[j]$ and construct two sets of equal sizes $X_i$Êand $Y_j$Êconsisting of the $\min({\ensuremath{\textsc{size}(X[i])}},{\ensuremath{\textsc{size}(Y[j])}})$ leftmost nodes from $X[i]$ and $Y[j]$. In Section \[sec:correctsetmacro\] we prove the following invariant on $X_i$ and $Y_j$ $${\ensuremath{\textsc{left}}}(1,X_i)=\mathcal{X}_l \textrm{ and } {\ensuremath{\textsc{left}}}(1,Y_j)=\mathcal{Y}_l \textrm { for some } l \;.$$ The procedure has two main cases. If $i=j$ then $i$ is either a leaf, left, or right node due to the invariant and the assumption on the input that ${\ensuremath{\mathcal{X}}}_l \lhd {\ensuremath{\mathcal{Y}}}_l$ (for a formal proof see Section \[sec:correctsetmacro\]). If $i$Êis a leaf node the nearest common ancestors of all pairs in $X_i$ and $Y_j$Êare in the leaf node or the boundary node. If $i$Êis a left or right node the nearest common ancestors of all the pairs are in $i$, on the spine, or in the top boundary node. In Case 1 we compute ${\ensuremath{\textsc{nca}}}$Êin the appropriate cluster depending on the type of $i$.
If $i\neq j$ we first compute the nearest common ancestor $h$Êof $i$Êand $j$ in the macro tree (Case 2). Due to the structure of the macro tree $h$Êis either a spine node or a boundary node (left, right, and leaf nodes have no descendants). If $h$Êis a boundary node all pairs in $X_i$Êand $Y_j$Êhave the same nearest common ancestor, namely $h$ (Case 2(a)). If $h$Êis a spine node there are three cases depending on the types of $i$Êand $j$. In Case 2(b)i $i=l(v,w)$ and $j\in \{s(v,w), r(v,w)\}$ (see Figure \[fig:propositions\](d) and (f)), or $i=s(v,w)$ and $j=r(v,w)$ (see Figure \[fig:propositions\](e)). In this case we compute ${\ensuremath{\textsc{nca}}}$ in the cluster containing $i, j, s(v,w), v$. In case 2(b)ii $i$ is a left node $l(v,w)$ and $j$Êis a (not necessarily proper) descendant of $w$ (see Figure \[fig:propositions\](g)). In this case we compute ${\ensuremath{\textsc{nca}}}$ on the rightmost node in $X_i$ and $w$Êin the cluster containing $i, v, w, s(v,w)$. We can restrict the computation to ${\ensuremath{\textsc{right}}}(1,X_i)$ because we always run [$\textsc{Deep}$]{} on the output from [$\textsc{Nca}$]{} before using it in any other computation and all nearest common ancestors of the pairs in $X_i$Êand $Y_j$Êwill be on the spine, and the deepest one will be the nearest common ancestor of the rightmost nodes in $X_i$Êand $Y_j$ (see Section \[sec:correctsetmacro\] for a formal proof). Case 2(b)iii is similar to Case 2(b)ii.
In the end of the iteration we have computed the nearest common ancestors of all the pairs in $X_i$Êand $Y_j$Êand the nodes from these are removed from $X[i]$Êand $Y[j]$.\
\
The implementation of [$\textsc{Deep}$]{} resembles the previous implementation, but takes advantage of the fact that the input list is in macro tree order.
Initialize an empty node array $R$ of size $n_{M}$.
Find the smallest $j$ such that $X[j]\neq \emptyset$. If no such $j$Êexists stop. Set $i:=j+1$.
Repeat until $i>n_M$:
- While $X[i]= \emptyset$ set $i:=i+1$.
Compare $j$ and $i$. There are three cases:
1. $j \lhd i$. Set $$S:=
\begin{cases}
C(j,v), & \textrm{if } j=l(v),\\
C(j, s(v,w), v), & \textrm{if } j\in\{l(v,w),r(v,w)\}, \\
C(j), & \textrm{otherwise}.
\end{cases}$$ Set $R[j] := {\ensuremath{\textsc{deep}}}_S (X[j])$.
2. $j \prec i$. There are two cases.
1. $j=s(v,w)$ and $i=r(v,w)$. Compute $$N := {\ensuremath{\textsc{deep}}}_{C(r(v,w), s(v,w), v)}(X[i] \cup X[j])\;.$$ Set $R[j]:=X[j]\cap N$.
2. All other cases. Do nothing.
3. $i \prec j$ (can happen if $i=s(v,w)$ and $j=l(v,w)$). Compute $$N := {\ensuremath{\textsc{deep}}}_{C(l(v,w), s(v,w), v)}(X[i] \cup X[j])\;.$$ Set $R[j]:=X[j]\cap N$, $X[i]:=X[i] \cap N$.
Set $j:=i$ and $i:=i+1$.
Set $R[j] := {\ensuremath{\textsc{deep}}}_S (X[j])$, where $S$Êis set as in Case 1.
Return $R$.
The procedure [$\textsc{Deep}$]{} has three cases. In case 1 node $i$ is to the right of our “potential output node” $j$. Since any node $l$ that is a descendant of $j$ must be to the left of $i$ ($l<i$) it cannot appear later in the list $X$ than $i$. We can thus safely add ${\ensuremath{\textsc{deep}}}_S(X[j])$ to $R$ at this point. To ensure that the cluster we compute [$\textsc{Deep}$]{} on is a tree we include the top boundary node if $j$Êis a leaf node and the top and spine node if $j$Êis a left or right node. We add the result to $R$ and set $i$ to be our new potential output node. In case 2 node $j$ is an ancestor of $i$ and therefore no node from $C(j)$ can be in the output list unless $j$ is a spine node and $i$Êis the corresponding right node. If this is the case we compute [$\textsc{Deep}$]{} of $X[j]$ and $X[i]$ in the cluster containing $i$Êand $j$ and add the result for $j$Êto the output and set $i$ to be our new potential output node. In case 3 node $i$Êis an ancestor of $j$. This can only happen if $j$ is a left node and $i$Êthe corresponding spine node. We compute [$\textsc{Deep}$]{} of $X[j]$ and $X[i]$ in the cluster containing $i$Êand $j$ and add the result for $j$Êto the output. We also set $X[i]$Êto result and let $i$Êbe our potential output node. The reason for this is that there might be nodes later in the input that are descendants of $i$, but since a left node has no children we can safely add the result for $j$Êto the output. After scanning the whole node array $X$ we add the last potential node $j$Êto the output after computing [$\textsc{Deep}$]{} of it as in Case 1.\
\
We now give the implementation of procedure [$\textsc{MopRight}$]{}. Procedure ${\ensuremath{\textsc{MopRight}}}$ takes a pair of node arrays $(X,Y)$ and another node array $Z$ as input. The pair $(X,Y)$ represents a set of minimum ordered pairs, where the first coordinates are in $X$ and the second coordinates are in $Y$. To simplify the implementation of procedure ${\ensuremath{\textsc{MopRight}}}$ it calls two auxiliary procedures ${\ensuremath{\textsc{MopSim}}}$ and ${\ensuremath{\textsc{Match}}}$ defined below. Procedure ${\ensuremath{\textsc{MopSim}}}$ computes ${\ensuremath{\mathrm{mop}}}$ of $Y$ and $Z$, and procedure ${\ensuremath{\textsc{Match}}}$ takes care of finding the first-coordinates from $X$ corresponding to the first coordinates from the minimum ordered pairs from $M$.
Compute $M:={\ensuremath{\textsc{MopSim}}}(Y,Z)$.
Compute $R:={\ensuremath{\textsc{Match}}}(X,Y,{\ensuremath{\mathop{M|}_{1}}})$.
Return $(R,{\ensuremath{\mathop{M|}_{2}}})$.
Procedure ${\ensuremath{\textsc{MopSim}}}$ takes two node arrays as input and computes ${\ensuremath{\mathrm{mop}}}$ of these.
Initialize two empty node arrays $R$ and $S$ of size $n_M$, set $i:=1$, $j:=1$, $(r_1,r_2):=(0,\emptyset)$, $(s_1,s_2):=(0,\emptyset)$. Repeat the following until $i>n_M$ or $j>n_M$:
- While $X[i] = \emptyset$ set $i:=i+1$. There are 4 cases:
1. $i=l(v,w)$ for some $v, w$. Until $Y[j]\neq \emptyset$ and either $i \lhd j$, $i=j$, or $j=s(v,w)$: set $j:=j+1$.
2. $i=s(v,w)$ for some $v, w$. Until $Y[j]\neq \emptyset$ and either $i \lhd j$ or $j=r(v,w)$: set $j:=j+1$.
3. $i \in \{r(v,w), l(v)\}$ for some $v, w$. Until $Y[j]\neq \emptyset$ and either $i \lhd j$ or $i=j$: set $j:=j+1$.
4. $i$ is a boundary node. Until $Y[j]\neq \emptyset$ and $i \lhd j$: set $j:=j+1$
Compare $i$ and $j$. There are two cases:
1. $i\lhd j$: Compare $s_1$ and $j$. If $s_1 < j$ set $R[r_1]:=R[r_1] \cup r_2$, $S[s_1]:=S[s_1] \cup s_2$, and $(s_1,s_2):= (j, {\ensuremath{\textsc{left}}}_{C(j)}(1,Y[j]))$.
Set $(r_1,r_2):=(i, {\ensuremath{\textsc{right}}}_{C(i)}(1,X[i]))$ and $i=i+1$.
2. Otherwise compute $(r,s):={\ensuremath{\textsc{mop}}}_{C(i,j,v)}(X[i],Y[j])$, where $v$Êis the top boundary node in the cluster $i$ and $j$Êbelong to.
If $r\neq \emptyset$ do:
- Compare $s_1$ and $j$. If $s_1 < j$ or if $s_1=j$ and ${\ensuremath{\textsc{leftof}}}_{C(i,j)}(X[i],s_2)=\emptyset$ then set $R[r_1]:=R[r_1] \cup r_2$, $S[s_1]:=S[s_1] \cup s_2$.
- Set $(r_1,r_2):=(i,r)$ and $(s_1,s_2):=(j,s)$.
There are two subcases:
1. $i=j$ or $i=l(v,w)$ and $j=s(v,w)$. Set $X[i]:={\ensuremath{\textsc{right}}}_{C(i)}(1,{\ensuremath{\textsc{rightof}}}_{C(i)}(X[i],r))$ and $j:=j+1$.
2. $i=s(v,w)$ and $j=r(v,w)$. If $r=\emptyset$ set $j:=j+1$ otherwise set $i:=j$.
Set $R[r_1]:=R[r_1] \cup r_2$ and $S[s_1]:=S[s_1] \cup s_2$.
Return $(R,S)$.
Procedure ${\ensuremath{\textsc{MopSim}}}$ is somewhat similar to the previous implementation of the procedure ${\ensuremath{\textsc{MopRight}}}$ from Section \[implementationsimple\]. As in the previous implementation we have a “potential pair” $((r_1,r_2), (s_1,s_2))$, where $r_1$ and $s_1$ are macro nodes, $r_2 \subseteq X[r_1]$, $s_2 \subseteq Y[s_1]$, where $r_2=\{r^1 \lhd \cdots \lhd r^k \}$ and $s_2=\{s^1\lhd \cdots \lhd s^k \}$ such that $r^l\lhd s^l$ for $l=1,\ldots k$. Furthermore, for any $l$Êthere exists no node $y \in Y[j]$, for $j<s_1$, such that $r^l \lhd y \lhd s^l$ and no node $x \in X[i]$, for $i<r_1$, such that $r^l \lhd x \lhd s^l$.
We have the following invariant at the beginning of each iteration: $$\nexists x \in X[i], \textrm{ such that } x \unlhd x', \textrm{ for any } x' \in r_2.$$
We first find the next non-empty macro node $i$. We then have 4 cases depending on which kind of node $i$ is. In Case I $i$ is a left node. Due to Proposition \[lem:orderlemma\] we can have ${\ensuremath{\mathrm{mop}}}$ in $i$ (case (i)), in the spine (case (iii)), or in a node to the right of $i$ (case(v)). In Case II $i$ is a spine node. Due to Proposition \[lem:orderlemma\] we can have ${\ensuremath{\mathrm{mop}}}$ in the right node (case (iv)) or in a node to the right of $i$ (case(v)). In Case III $i$ is a right node or a leaf node. Due to Proposition \[lem:orderlemma\] we can have ${\ensuremath{\mathrm{mop}}}$ in $i$ (case (i) and (ii)) or in a node to the right of $i$ (case(v)). In the last case (Case IV) $i$ must be a boundary node and ${\ensuremath{\mathrm{mop}}}$Êmust be in a node to the right of $i$.
We then compare $i$Êand $j$. The case where $i \lhd j$Êis similar to the previous implementation of the procedure. We compare $j$ with our potential pair. If $s_1 < j$ then $s_1 \lhd j$ since the input is deep, and we can insert $r_2$ and $s_2$ into our output node arrays $R$ and $S$, respectively. We also set $s_1$ to $j$ and $s_2$ to the leftmost node in $Y[j]$ (if $s_1=j$ we already have $(s_1,s_2)=(j, {\ensuremath{\textsc{left}}}_{C(j)}(1,Y[j]))$). Then—both if $s_1 \lhd j$ or $s_1=j$—we set $r_1$ to $i$ and $r_2$ to the rightmost node in $X[i]$. That we only need the rightmost node in $X[i]$ and the leftmost node in $Y[j]$ follows from the definition of ${\ensuremath{\mathrm{mop}}}$ and the structure of the macro tree.
Case 2 ($i \ntriangleleft j$) is more complicated. In this case we first compute ${\ensuremath{\mathrm{mop}}}$ in the cluster $i$Êand $j$ belong to. If this results in any minimum ordered pairs ($r \neq \emptyset$) we must update our potential pair. Otherwise we leave the potential pair as it is and only update $i$Êand $j$. If $r \neq \emptyset$ we compare $s_1$ and $j$. As in Case 1 of the procedure we add our potential pair to the output and update the potential pair with $r$ and $s$ if $s_1 < j$, since this implies $s_1 \lhd j$. If $s_1=j$Êand no nodes in $X[i]$ are to the left of the leftmost node in $s_2$ we also add the potential pair to the output and update it. We show in the next section that in this case $|s_2|=1$. Therefore we can safely add the potential pair to the output. In all other cases the pair $(r,s)\neq (\emptyset,\emptyset)$Êshows a contradiction to our potential pair and we update the potential pair without adding anything to the output.
Finally, we update $X[i]$, $i$, and $j$. There are two cases depending on $i$Êand $j$. In Case (a) either $i=j$ or $i$Êis a left node and $j$Êis the corresponding spine node. In both cases we can have nodes in $X[i]$ that are not to the left of any node in $Y[j]$. These nodes could be in a minimum ordered pair with nodes from another macro node. We show in the next section that this can only be true for the rightmost node in $X[i]$. $X[i]$Êis updated accordingly. After this update all nodes in $Y[j]$Êare to the left of all nodes in $X[i]$ in the next iteration and therefore $j$ is incremented. In Case (b) $i$Êis a spine node and $j$ is the corresponding right node. Since the input lists are deep, there is only one node in $X[i]$. If $r= \emptyset$ then no node in $Y[j]$Êis to the right of the single node in $X[i]$. Since the input arrays are deep, no node later in the array $X$ can be to the left of any node in $Y[j]$ and we therefore increment $j$. If $r \neq \emptyset$ then $(r_1,r_2)=(i,X[i])$ and we update $i$. Instead of incrementing $i$Êby one we set $i:=j$, this is correct since all macro nodes with macro node number between $i$Êand $j$Êare descendants of $i$, and thus contains no nodes from $X$, since $X$ is deep.
When reaching the end of one of the arrays we add our potential pair to the output and return.
As in Section \[implementationsimple\] we can implement [$\textsc{MopLeft}$]{} similarly to [$\textsc{MopRight}$]{}.\
\
Procedure [$\textsc{Match}$]{} takes three node arrays $X$, $Y$, and $Y'$ representing deep sets ${\ensuremath{\mathcal{X}}}$, ${\ensuremath{\mathcal{Y}}}$, and ${\ensuremath{\mathcal{Y}}}'$, where ${\ensuremath{| {\ensuremath{\mathcal{X}}} |}}={\ensuremath{| {\ensuremath{\mathcal{Y}}} |}}$, and ${\ensuremath{\mathcal{Y}}}' \subseteq {\ensuremath{\mathcal{Y}}}$. The output is a node array representing the set $\{{\ensuremath{\mathcal{X}}}_j \mid {\ensuremath{\mathcal{Y}}}_j \in {\ensuremath{\mathcal{Y'}}}\}$.
Initialize an empty node array $R$ of size $n_M$, set $X_L:=\emptyset$, $Y_L:=\emptyset$, $Y_L':=\emptyset$, $x:=0$, $y:=0$, $i:=1$ and $j:=1$.
Repeat until $i>n_M$ or $j>n_M$:
- Until $X[i] \neq \emptyset$ set $i:=i+1$. Set $x:={\ensuremath{\textsc{size}(X[i])}}$.
Until $Y[j] \neq \emptyset$ set $j:=j+1$. Set $y:={\ensuremath{\textsc{size}(Y[j])}}$.
Compare $Y[j]$ and $Y'[j]$. There are two cases:
1. $Y[j]=Y'[j]$. Compare $x$ and $y$. There are three cases:
1. $x=y$. Set $R[i]:=R[i] \cup X[i]$, $i:=i+1$, and $j:=j+1$.
2. $x < y$. Set $R[i]:=R[i] \cup X[i]$, $
Y[j]:={\ensuremath{\textsc{right}}}(y-x,Y[j])$, $Y'[j]:=Y[j]$, and $i:=i+1$.
3. $x > y$. Set $X_L:={\ensuremath{\textsc{left}}}(y,X[i])$, $R[i]:=R[i] \cup X_L$, $X[i]:=X[i]\setminus X_L$, and $j:=j+1$.
2. $Y[j]\neq Y'[j]$. Compare $x$ and $y$. There are three cases:
1. $x=y$. Set $R[i]:=R[i] \cup {\ensuremath{\textsc{match}}}(X[i],Y[j],Y'[j])$, $i:=i+1$, and $j:=j+1$.
2. $x < y$. Set $Y_L:={\ensuremath{\textsc{left}}}(x,Y[j])$, $Y'_L:=Y'[j] \cap Y_L$,
$R[i]:=R[i] \cup {\ensuremath{\textsc{match}}}(X[i],Y_L,Y'_L)$, $Y[j]:=Y[j]\setminus Y_L$, $Y'[j]:=Y'[j] \setminus Y'_L$, and $i:=i+1$.
3. $x > y$. Set $X_L:={\ensuremath{\textsc{left}}}(y,X[i])$, $R[i]:=R[i] \cup {\ensuremath{\textsc{match}}}(X_L,Y[j],Y'[j])$, $X[i]:=X[i]\setminus X_L$, and $j:=j+1$.
Return $R$.
Procedure ${\ensuremath{\textsc{Match}}}$ proceeds as follows. First we find the first non-empty entries in the two node arrays $X[i]$ and $Y[j]$. We then compare $Y[j]$ and $Y'[j]$.
If they are equal we keep all nodes in $X$ with the same rank as the nodes in $Y[j]$. We do this by splitting into three cases. If there are the same number of nodes $X[i]$ and $Y[j]$ we add all nodes in $X[i]$ to the output and increment $i$ and $j$. If there are more nodes in $Y[j]$Êthan in $X[i]$ we add all nodes in $X[i]$ to the output and update $Y[j]$ and $Y'[j]$ to contain only the $y-x$ leftmost nodes in $Y[j]$. We then increment $i$ and iterate. If there are more nodes in $X[i]$Êthan in $Y[j]$ we add the first $y$ nodes in $X[i]$Êto the output, increment $j$, and update $X[i]$Êto contain only the nodes we did not add to the output.
If $Y[j] \neq Y'[j]$ we call the cluster procedure [$\textsc{match}$]{}. Again we split into three cases depending on the number of nodes in $X[i]$ and $Y[j]$. If they have the same number of nodes we can just call [$\textsc{match}$]{} on $X[i]$, $Y[j]$, and $Y'[j]$ and increment $i$ and $j$. If ${\ensuremath{\textsc{size}(Y[j])}} > {\ensuremath{\textsc{size}(X[i])}}$ we call match with $X[i]$ the leftmost ${\ensuremath{\textsc{size}(X[i])}}$ nodes of $Y[j]$ and with the part of $Y'[j]$ that are a subset of these leftmost ${\ensuremath{\textsc{size}(X[i])}}$ nodes of $Y[j]$. We then update $Y[j]$ and $Y'[j]$ to contain only the nodes we did not use in the call to [$\textsc{match}$]{} and increment $i$. If ${\ensuremath{\textsc{size}(Y[j])}} < {\ensuremath{\textsc{size}(X[i])}}$Êwe call [$\textsc{match}$]{} with the leftmost ${\ensuremath{\textsc{size}(Y[j])}}$ nodes of $X[i]$, $Y[j]$, and $Y'[j]$. We then update $X[i]$ to contain only the nodes we did not use in the call to [$\textsc{match}$]{} and increment $j$.
Initialize an empty node array $R$ of size $n_M$ and two node lists $L$ and $S$.
Repeat until $i>n_M$:
- Until $X[i] \neq \emptyset$ set $i:=i+1$.
There are three cases depending on the type of $i$:
1. $i \in \{l(v,w), r(v,w)\}$. Compute $$N := {\ensuremath{\textsc{fl}}}_{C(i, s(v,w), v)}(X[i], \alpha)\;.$$
If $N \neq \emptyset$, set $R[j] = R[j] \cup (N \cap V(C(j)))$ for each $j \in \{i, s(v,w), v\}$.
Otherwise, set $L := L \circ {\ensuremath{\mathrm{parent}}}_{{M}}(v)$.
2. $i = l(v)$. Compute $$N := {\ensuremath{\textsc{fl}}}_{C(i, v)}(X[i],\alpha)\;.$$
If $N \neq \emptyset$, set $R[j] :=R[j] \cup (N \cap V(C(j)))$ for each $j \in \{i, v\}$.
Otherwise, set $L := L \circ {\ensuremath{\mathrm{parent}}}_{{M}}(v)$.
3. $i \not\in \{l(v,w), r(v,w),l(v)\}$. Compute $$N := {\ensuremath{\textsc{fl}}}_{C(i)}(X[i], \alpha)\;.$$
If $N \neq \emptyset$, set $R[i] := R[i] \cup N$.
Otherwise set $L := L \circ {\ensuremath{\mathrm{parent}}}_{{M}}(i)$.
Subsequently, compute the list $S := {\ensuremath{\textsc{fl}}}_{M}(L, \alpha)$. For each node $i \in S$ set $R[i] := R[i] \cup {\ensuremath{\textsc{fl}}}_{C(i)}({\ensuremath{\mathrm{first}}}(i), \alpha))$.
Return ${\ensuremath{\textsc{Deep}}}(R)$.
The ${\ensuremath{\textsc{Fl}}}$ procedure is similar to ${\ensuremath{\textsc{Parent}}}$. The cases 1, 2 and 3 compute ${\ensuremath{\textsc{Fl}}}$ on a micro forest. If the result is within the micro tree we add it to $R$ and otherwise we store in a node list $L$ the node in the macro tree which contains the parent of the root of the micro forest. Since we always call [$\textsc{Deep}$]{}on the output from ${\ensuremath{\textsc{Fl}}}(X,\alpha)$ there is no need to compute [$\textsc{Fl}$]{} in the macro tree if $N$ is nonempty. We then compute ${\ensuremath{\textsc{Fl}}}$ in the macro tree on the list $L$, store the results in a list $S$, and use this to compute the final result.
Consider the cases of procedure [$\textsc{Fl}$]{}. In Case 1 $i$ is a left or right node. Due to Proposition \[lem:ancestorlemma\] case (i) and (ii) ${\ensuremath{\mathrm{fl}}}$ of a node in $i$ can be in $i$ or on the spine or in the top boundary node. If this is not the case it can be found by a computation of ${\ensuremath{\textsc{Fl}}}$ of the parent of the top boundary node of the $i$’s cluster in the macro tree (Proposition \[lem:ancestorlemma\] case (iii)). In Case 2 $i$Êis a leaf node. Then ${\ensuremath{\mathrm{fl}}}$ of a node in $i$ must either be in $i$, in the top boundary node, or can be found by a computation of ${\ensuremath{\textsc{Fl}}}$ of the parent of the top boundary node of the $i$’s cluster in the macro tree. If $i$ is a spine node or a boundary node, then ${\ensuremath{\mathrm{fl}}}$ of a node in $i$ is either in $i$Êor can be found by a computation of ${\ensuremath{\textsc{Fl}}}$ of the parent of $i$ in the macro tree.
Correctness of the Set Procedures {#sec:correctsetmacro}
---------------------------------
The following lemmas show that the set procedures are correctly implemented.
Procedure [$\textsc{Parent}$]{} is correctly implemented.
We will prove that in iteration $i$Êthe procedure correctly computes the parents of all nodes in $i$. There are four cases depending on the type of $i$. Consider the case $i\in\{l(v,w),r(v,w)\}$, i.e., $i$Êis a left or right node. For all nodes $x$ in $C(i)$, ${\ensuremath{\mathrm{parent}}}(x)$ is either in $C(i)$, on the spine $s(v,w)$, or is the top boundary node $v$. The parents of all input nodes in $C(i)$ is thus in $N$ computed in Case 1 in the procedure. The last line in Case 1 (“For each $j\in \{i, s(v,w),v\},\ldots$”) adds the set of parents to the right macro node in the output array.
If $i$ is a leaf node $l(v)$ then for any node $x \in C(i)$, ${\ensuremath{\mathrm{parent}}}(x)$ is either in $C(i)$ or is the boundary node $v$. The parents of all input nodes in $C(i)$ is thus in $N$ computed in Case 2 in the procedure. The last line in Case 2 (“For each $j\in \{i,v\}, \ldots$”) adds the set of parents to the right macro node in the output array.
If $i$Êis a spine node $s(v,w)$ then the input contains at most one node in $C(i)$, since the input to the procedure is deep. For any $x\in C(i)$, ${\ensuremath{\mathrm{parent}}}(x)$ is either a node on the spine or the top boundary node $v$. This is handled by Case 3 in the procedure. Let $x$Êbe the node in $X[i]$. If ${\ensuremath{\mathrm{parent}}}(x)=v$, then $N=\emptyset$, and we compute $j$, which is the parent $v$Êof $i$Êin the macro tree, and add $j$Êto the output array (since $j=v$Êis a boundary node ${\ensuremath{\mathrm{first}}}(j)=v$). If ${\ensuremath{\mathrm{parent}}}(x)$ is another node $y$Êon the spine, then $N=\{y\} \neq \emptyset$ and $y$Êis added to the output array.
If $i$ is a boundary node $v$, then ${\ensuremath{\mathrm{parent}}}(v)$ is either another boundary node $v'$, the bottom node on a spine, or $\bot$Êif $v$Êis the root. This is handled by Case 3 in the procedure. In all three cases $N=\emptyset$ and we compute the parent $j$ of $i$Êin the macro tree. If $i$Êis the root, then $j=\bot$ and we do nothing. Otherwise, we add ${\ensuremath{\mathrm{first}}}(j)$ to the output. If $parent(v)$Êis a boundary node then ${\ensuremath{\mathrm{first}}}(j)=j$. If $j$Êis a spine node then ${\ensuremath{\mathrm{first}}}(j)$ is the bottom node on $j$.
In each iteration of the procedure we might add nodes to the output, but we never delete anything written to the output in earlier iterations. Procedure [$\textsc{Parent}$]{} thus correctly computes the parents of all nodes in $X$.
Before proving the correctness of procedure [$\textsc{Nca}$]{} we will prove the following invariant on $X_i$Êand $Y_j$ in the procedure.
\[inv:nca\] In procedure [$\textsc{Nca}$]{} we have the following invariant of $X_i$ and $Y_j$: $${\ensuremath{\textsc{left}}}(1,X_i)=\mathcal{X}_l \textrm{ and } {\ensuremath{\textsc{left}}}(1,Y_j)=\mathcal{Y}_l \textrm { for some } l \;.$$
The proof is by induction on the number of iterations of the outer loop. After the while loop on $X$ in the first iteration $i$Êis the smallest integer such that $X[i]\neq \emptyset$. Due to the macro tree order of the array $X$, $X[i]$ contains the first nodes from $X$ wrt. the preorder of the original tree (Proposition \[prop:macrotreeorder\]). Similarly, $Y[j]$Êcontains the leftmost node in ${\ensuremath{\mathcal{Y}}}$. The invariant now follows immediately from the assignment of $X_i$Êand $Y_j$.
For the induction step consider iteration $m$ and let $i'$Êand $j'$ be the values of $i$Êand $j$ after the while loops in the previous iteration. By the induction hypothesis ${\ensuremath{\textsc{left}}}(1,X_{i'})=\mathcal{X}_l \textrm{ and } {\ensuremath{\textsc{left}}}(1,Y_{j'})=\mathcal{Y}_l \textrm { for some } l$. Let $n'=\min({\ensuremath{\textsc{size}(X[i'])}},{\ensuremath{\textsc{size}(Y[j'])}})$. Then $X_{i'}$Êcontains ${\ensuremath{\mathcal{X}}}_l, \ldots, {\ensuremath{\mathcal{X}}}_{l+n'}$ and $Y_{j'}$Êcontains ${\ensuremath{\mathcal{Y}}}_l, \ldots, {\ensuremath{\mathcal{Y}}}_{l+n'}$. We will show that ${\ensuremath{\textsc{left}}}(1,X_i)={\ensuremath{\mathcal{X}}}_{l+n'+1}$. In the end of the previous iteration we removed $X_{i'}$ from $X[i']$. There are two cases. If $X[i']\neq \emptyset$Êthen it clearly contains ${\ensuremath{\mathcal{X}}}_{l+n'+1}$ as its leftmost node. Since a spine node can only contain one node from ${\ensuremath{\mathcal{X}}}$, $i'$ cannot be a spine node. Thus $i=i'$ when we get to the 5th line in the current iteration (“Set $n:=\min({\ensuremath{\textsc{size}(X[i])}},{\ensuremath{\textsc{size}(Y[j])}}), \ldots $”). It follows that ${\ensuremath{\textsc{left}}}(1,X_i)={\ensuremath{\mathcal{X}}}_{l+n'+1}$. In the other case $X[i']=\emptyset$Êafter the removal of $X_{i'}$. It follows from the macro tree order of $X$Êthat $X[i]$ contains ${\ensuremath{\mathcal{X}}}_{l+n'+1}$ as its leftmost node. It follows by a similar argument that ${\ensuremath{\textsc{left}}}(1,Y_j)={\ensuremath{\mathcal{Y}}}_{l+n'+1}$.
Let $X$ and $Y$ be two node arrays representing the deep sets ${\ensuremath{\mathcal{X}}}$Êand ${\ensuremath{\mathcal{Y}}}$, ${\ensuremath{| {\ensuremath{\mathcal{X}}} |}}={\ensuremath{| {\ensuremath{\mathcal{Y}}} |}}=k$, and let ${\ensuremath{\mathcal{X}}}_i$ and ${\ensuremath{\mathcal{Y}}}_i$Êdenote the $i$th element of ${\ensuremath{\mathcal{X}}}$ and ${\ensuremath{\mathcal{Y}}}$, wrt. to their preorder number in the tree, respectively. For all $i=1,\ldots , k$, assume ${\ensuremath{\mathcal{X}}}_i \lhd {\ensuremath{\mathcal{Y}}}_i$. Procedure ${\ensuremath{\textsc{Nca}}}(X,Y)$ correctly computes ${\ensuremath{\textsc{Deep}}}(\{{\ensuremath{\mathrm{nca}}}({\ensuremath{\mathcal{X}}}_i,{\ensuremath{\mathcal{Y}}}_i) | 1\leq i\leq k\})$.
We are now ready to show that the procedure correctly takes care of all possible cases from Proposition \[lem:ncalemma\]. The proof is split into two parts. First we will argue that some of the cases from the proposition cannot occur during an iteration of the outer loop of [$\textsc{Nca}$]{}. Afterwards we prove that the procedure takes care of all the cases that can occur.
Case (iii) cannot happen since if $i=j$ is a spine nodeÊthen ${\ensuremath{\mathcal{X}}}_l$ is either a descendant or an ancestor of ${\ensuremath{\mathcal{Y}}}_l$ contradicting the assumption on the input that $\mathcal{X}_l \lhd \mathcal{Y}_l$. Case (vi) can only happen if $i\neq j$: If $i=j$ and we are in case (vi) then $i=j$Êis a boundary node, and this would imply that $C(i)$ only consists of one node, i.e., $\mathcal{X}_l= X[i]=Y[j]=\mathcal{Y}_l$ contradicting the assumption on the input that $\mathcal{X}_l \lhd \mathcal{Y}_l$. Due to this assumption on the input we also have that in case (iv) of the proposition $i$Êis either a left node or a spine node and $j$ is a spine node or a right node. For case (v) either $i$Êis a left node and $j$Êis a descendant of the bottom boundary node of $i$’s cluster or $j$Êis a right node and $i$Êis a descendant of the bottom boundary node of $j$’s cluster. All the other cases from case (v) would contradict the assumption that $\mathcal{X}_l \lhd \mathcal{Y}_l$.
The procedure first constructs two sets $X_i$Êand $Y_j$Êcontaining the elements ${\ensuremath{\mathcal{X}}}_l,\ldots,{\ensuremath{\mathcal{X}}}_{l+n}$ and${\ensuremath{\mathcal{Y}}}_l,\ldots,{\ensuremath{\mathcal{Y}}}_{l+n}$ for some $l$, respectively, where $n=\min({\ensuremath{\textsc{size}(X[i])}},{\ensuremath{\textsc{size}(Y[j])}})$. The procedure [$\textsc{Nca}$]{} has two main cases depending on whether $i=j$ or not. Case 1 ($i=j$) takes care of cases (i)–(ii) from Proposition \[lem:ncalemma\]. Case 2 ($i\neq j$) takes care of the remaining cases from Proposition \[lem:ncalemma\] (iv)–(vi) that can occur.
First consider Case 1: We compute nearest common ancestors $N$Êof the $n$Ênodes in $X_i$Êand $Y_j$ in a cluster $S$ depending on what kind of node $i$ is. We need to show that Case 1 handles Case (i) and (ii) from the Proposition correctly.
- Case (i). $i=j$Êis a leaf node. By the Proposition the nearest common ancestors of the pairs in $({\ensuremath{\mathcal{X}}}_l,{\ensuremath{\mathcal{Y}}}_l), \ldots,({\ensuremath{\mathcal{X}}}_{l+n},{\ensuremath{\mathcal{Y}}}_{l+n})$ from $X_i$ and $Y_j$ is either in $c(i)$Êor in the boundary node, i.e., in $C(i,v)$.
- Case (ii). $i=j$ is a left or right node. By the Proposition the nearest common ancestors of the pairs in $\{({\ensuremath{\mathcal{X}}}_l,{\ensuremath{\mathcal{Y}}}_l), \ldots,({\ensuremath{\mathcal{X}}}_{l+n},{\ensuremath{\mathcal{Y}}}_{l+n})\}$ from $X_i$ and $Y_j$ is either in $c(i)$, on the spine, or in the top boundary node, i.e., in $C(i,s(v,w),v)$.
Thus $S$ is correctly set in both cases. After the computation of $N$Êthe output is then added to the entries in the output array $R$Êfor each of the macro nodes in $S$. Case 1 thus handles Case (i)-(ii) (and only these two cases) from Proposition \[lem:ncalemma\].
Next consider Case 2 ($i \neq j$). We first compute the nearest common ancestor $h$Êof $i$Êand $j$Êin the macro tree. The macro node $h$ is either a boundary node or a spine node due to the structure of the macro tree (see also Proposition \[lem:ncalemma\]). We will show that Case 2 takes care of the remaining cases.
- Case (iv). From the above discussion it follows that we have one of the three following cases. $i=l(v,w)$ and $j=s(v,w)$, $i=l(v,w)$ and $j=r(v,w)$, or $i=s(v,w)$ and $j=r(v,w)$. All three cases are handled in Case 2(b)i of the procedure. It follows from the proposition that [$\textsc{nca}$]{} is computed in the correct cluster.
- Case (v). It follows from the discussion above that either $i=l(v,w)$ and $w\preceq j$, or $j=r(v,w)$ and $w\preceq i$. These two cases are handled by Case 2(b)ii and 2(b)iii of the procedure. It follows from the Proposition that [$\textsc{nca}$]{} is computed in the correct cluster. We need to argue that we can restrict the computation of [$\textsc{nca}$]{} to the pair $({\ensuremath{\textsc{right}}}(1,X_i),w)$ instead of computing [$\textsc{nca}$]{} for all nodes in $\{{\ensuremath{\mathcal{X}}}_l, \ldots,{\ensuremath{\mathcal{X}}}_{l+n}\}$. Consider the case where $i=l(v,w)$ and $w\preceq j$ (Case 2(b)ii of the procedure). Since $w \preceq{\ensuremath{\mathcal{Y}}}_r$Êfor all $r=l,\ldots l+n$, and ${\ensuremath{\mathcal{X}}}_l \lhd {\ensuremath{\mathcal{X}}}_{l+1} \lhd \ldots \lhd {\ensuremath{\mathcal{X}}}_{l+n}$, then ${\ensuremath{\mathrm{nca}}}({\ensuremath{\mathcal{X}}}_{r},{\ensuremath{\mathcal{Y}}}_{r})\preceq {\ensuremath{\mathrm{nca}}}({\ensuremath{\mathcal{X}}}_{l+n},{\ensuremath{\mathcal{Y}}}_{l+n})$ for all $r=l,\ldots l+n$. Thus we do not need to compute ${\ensuremath{\mathrm{nca}}}({\ensuremath{\mathcal{X}}}_{r},{\ensuremath{\mathcal{Y}}}_{r})$ for $r \neq n+l$, since the output of the procedure is ${\ensuremath{\textsc{Deep}}}(\{{\ensuremath{\mathrm{nca}}}({\ensuremath{\mathcal{X}}}_i,{\ensuremath{\mathcal{Y}}}_i) | 1\leq i\leq k\})$. A similarl argument shows that we can restrict the computation to $(w,{\ensuremath{\textsc{left}}}(1,Y_j))$ in Case 2(b)iii.
- Case (vi). It follows from the discussion above and the proposition that $i\neq j$Êand $i$Êand $j$Êare in different clusters, and we are not in any of the cases from (iv) and (v). Thus $h$Êmust be a boundary node and all the pairs $\{({\ensuremath{\mathcal{X}}}_l,{\ensuremath{\mathcal{Y}}}_l), \ldots,({\ensuremath{\mathcal{X}}}_{l+n},{\ensuremath{\mathcal{Y}}}_{l+n})\}$ have the same nearest common ancestor, namely $h$. This is handled by Case 2(a).
We have now argued that the procedure correctly takes care of all possible cases from Proposition \[lem:ncalemma\]. It remains to show that all pairs from $\{{\ensuremath{\mathrm{nca}}}({\ensuremath{\mathcal{X}}}_i,{\ensuremath{\mathcal{Y}}}_i) | 1\leq i\leq k\}$ are considered during the computation. It follows from the invariant that we only consider pairs from the input. In the last lines we remove the nodes from the input that we have computed the [$\mathrm{nca}$]{}s of in this iteration. It follows from the proof of the invariant that no entry in the input arrays is left nonempty. Thus all pairs are taken care of.
To prove that procedure [$\textsc{Deep}$]{} is correctly implemented we will use the following fact about preorder and postorder numbers in the macro tree.
\[prop:preordermacro\] Let $i$ and $j$Êbe nodes in the macro tree identified by their macro tree number such that $i<j$. For all $x \in C(i), y \in C(j)$ we have
1. ${\ensuremath{\mathrm{pre}}}(x) < {\ensuremath{\mathrm{pre}}}(y)$ unless $i=l(v,w)$ and $j=s(v,w)$.
2. ${\ensuremath{\mathrm{post}}}(y) > {\ensuremath{\mathrm{post}}}(x)$ unless $i=s(v,w)$ and $j=r(v,w)$.
\[prop:preordermacrotree\] Let $x_1,\ldots,x_n$ be nodes from the macro tree associated with their macro tree number such that $x_1 < x_2 < \cdots < x_n$. If $x_i \lhd x_j$ for some $i$Êand $j$Êthen $x_i \lhd x_k$ for all $x_k>x_j$.
From $x_i \lhd x_j$ we have ${\ensuremath{\mathrm{pre}}}(x_i)<{\ensuremath{\mathrm{pre}}}(x_j)$ and ${\ensuremath{\mathrm{post}}}(x_i) < {\ensuremath{\mathrm{post}}}(x_j)$. Since $x_k>x_j$Êwe have ${\ensuremath{\mathrm{pre}}}(x_j)<{\ensuremath{\mathrm{pre}}}(x_k)$Êunless $x_k=s(v,w)$ and $x_j=l(v,w)$. In that case, ${\ensuremath{\mathrm{pre}}}(x_k)+1={\ensuremath{\mathrm{pre}}}(x_j)>{\ensuremath{\mathrm{pre}}}(x_i)$. Since $x_i \lhd x_j$ we have $x_i\neq x_j$Êand thus ${\ensuremath{\mathrm{pre}}}(x_k)>{\ensuremath{\mathrm{pre}}}(x_i)$.
It remains to show that ${\ensuremath{\mathrm{post}}}(x_i) < {\ensuremath{\mathrm{post}}}(x_k)$. Assume for the sake of contradiction that ${\ensuremath{\mathrm{post}}}(x_k)<{\ensuremath{\mathrm{post}}}(x_i) <{\ensuremath{\mathrm{post}}}(x_j)$. This implies $x_i \prec x_k$ and $x_j \prec x_k$ contradicting $x_i \lhd x_j$.
We will first prove the following invariants on $i$Êand $j$ in procedure [$\textsc{Deep}$]{}.
\[inv:deepi\] After the while loop in line 4 of procedure [$\textsc{Deep}$]{} (“While $X[i]=\emptyset$ set $i:=i+1$.”) we have the following invariant on $i$Êand $j$: For all $l$Êsuch that $j < l < i$ we have $X[l]=\emptyset$.
Let $i'$ be the value of $i$ in the iteration before the previous one. Then $i$Êis the smallest index greater than $i'$ such that the corresponding entry in $X$ is nonempty. This is true since $i$Êis always incremented by one in the end of an iteration, and at the beginning of the next iteration we start with the while loop incrementing $i$Êuntil we find a nonempty entry. Since $j=i'$,Ê$i$ is the first nonempty entry greater than $j$ and the claim follows.
\[inv:deep\] Right before each iteration of the main loop of procedure [$\textsc{Deep}$]{} (“Repeat until $i>n_M$”) we have the following invariant on $j$: For all nodes $x\in X[j]$ and $y \in X[l]$, where $1 \leq l < j$, we have $x \not \prec y$.Ê
Recall that $x \prec y \Leftrightarrow {\ensuremath{\mathrm{pre}}}(x)<{\ensuremath{\mathrm{pre}}}(y) \textrm{ and } {\ensuremath{\mathrm{post}}}(y) < {\ensuremath{\mathrm{post}}}(x)$. By Proposition \[prop:preordermacro\] the only case where we can have ${\ensuremath{\mathrm{pre}}}(x) < {\ensuremath{\mathrm{pre}}}(y)$ is if $l=l(v,w)$ and $j=s(v,w)$Êfor some $v,w$. Assume this is the case. If $X[l]=\emptyset$ the claim follows trivially. Otherwise, let $i'$ and $j'$Êbe the values of $i$Êand $j$Êin the previous iteration, respectively (since $l<j$Êand $X[l]\neq \emptyset$ there must be such an iteration). We have $j=l+1$, $i'=j=s(v,w)$Êand $j'=l=l(v,w)$. Thus in the previous iteration the procedure entered case 3, where $X[i']$ was set to $X[i']\cap {\ensuremath{\textsc{deep}}}_{C(l(v,w),s(v,w),v)}(X[i']\cup X[j'])$, and thus $X[j]$ contains no nodes that are ancestors of nodes in $X[j']=X[l]$.Ê
Procedure [$\textsc{Deep}$]{} is correctly implemented.
We will prove that $x\in {\ensuremath{\textsc{Deep}}}(X)$ iff $x\in X$ and there exists no $y\in X$ such that $x\prec y$.
Assume $x \in {\ensuremath{\textsc{Deep}}}(X)$. Consider the iteration when $x$ is assigned to the output. There are three cases depending on which case we are in when $x$ is added to the input. If $j \lhd i$Ê(Case 1 of the procedure) then $x \in {\ensuremath{\textsc{Deep}}}_S(X[j])$ and it follows from the invariant on $j$ (Lemma \[inv:deep\]) that $x$Êhas no descendants in any nodes $y \in X[l]$, $l<j$. For $j<l<i$ the claim follows directly from Lemma \[inv:deepi\]. It remains to show that $x$Êhas no descendants in $X[l]$ for $l\geq i$. By Proposition \[prop:preordermacrotree\] we have $j\lhd l$Êfor all $l>i$ and the claim follows from Proposition \[lem:ancestorlemma\].
If $j \prec i$ (Case 2 of the procedure) then $j$ is a spine node $s(v,w)$ and $i$ is the corresponding right node $r(v,w)$, and we compute $N:= {\ensuremath{\textsc{Deep}}}_{C(r(v,w),s(v,w),v)}(X[i]\cup X[j])$. Since $x \in {\ensuremath{\textsc{Deep}}}(X)$ we have $x \in R[j]=X[j]\cap N$. It follows from the invariant (Lemma \[inv:deep\]) and the computation of $N$ that $x$Êhas no descendants in $X[l]$Êfor any $l\leq j$. For $l>j$ it follows from the structure of the macro tree that for any $l>i$ we have $j \lhd l$. For $j<l<i$ the claim follows directly from Lemma \[inv:deepi\]. The claim follows from Proposition \[lem:ancestorlemma\]. For $j<l<i$ the claim follows directly from Lemma \[inv:deepi\].
If $i \prec j$ (Case 3 of the procedure) then $i$Êis a spine node $s(v,w)$ and $j$ is the corresponding left node $l(v,w)$, and we compute $N:= {\ensuremath{\textsc{Deep}}}_{C(l(v,w),s(v,w),v)}(X[i]\cup X[j])$. Since $x \in {\ensuremath{\textsc{Deep}}}(X)$ we have $x \in R[j]=X[j]\cap N$. It follows from the computation of $N$Êthat $x$Êhas no descendants in $X[i]\cup X[j]$. Since $l(v,w)$ has no descendants in the macro tree it follows from Proposition \[lem:ancestorlemma\] that $x$Êhas no descendants in $X[l]$Êfor any $l \neq j$.
If $x$ is assigned to the output in the line next to last (“Set $R[j]:= {\ensuremath{\textsc{Deep}}}_S(X[j])$, where $S$Êis set as in Case 1”) then it follows from the invariant on $j$ (Lemma \[inv:deep\]) and the computation of ${\ensuremath{\textsc{Deep}}}_S(X[j])$ that $x$Êhas no descendants in $X$.
For the other direction let $x \in X$ be a node such that $X\cap V(T(x))=\{x\}$. Let $l$Êbe the index such that $x\in X[l]$. All nonempty entries in $X$ are $i$Êin the second line of the main loop (“Compare $i$Êand $j$..”) at some iteration. Consider the iteration when $i=l$. Unless $i=l(v,w)$Êand $j=s(v,w)$ (Case 3 of the procedure) $X[i]$Êis not changed in this iteration. If we are in Case 3, then $N$ is computed and $X[i]$ is set to $X[i]\cap N$. Since $x$ has no descendants in $X$Êwe have $x \in N$ and thus $x \in X[i]$ after the assignment. At the end of this iteration $j$ is set to $i$. Consider the next iteration when $j=l$. If $j \lhd i$ or $i>n_M$ then $x \in {\ensuremath{\textsc{Deep}}}_S(X[j])=R[j]$. If $j \prec i$ we have $j=s(v,w)$ and $i = r(v,w)$ since $x$Êhas no descendants in $X$. For the same reason we have $x\in N$ and thus $x\in X[j]\cap N=R[j]$. If $i \prec j$ we have $i=s(v,w)$ and $j=l(v,w)$. Again $x\in N$ and thus $x\in X[j]\cap N=R[j]$.
We now consider procedures [$\textsc{MopSim}$]{} and [$\textsc{Match}$]{}.
\[lem:mopInv1\] Let $((r_1,r_2), (s_1,s_2))$ be as defined in procedure [$\textsc{MopSim}$]{}. Then $r_1$ and $s_1$ are macro nodes, $r_2 \subseteq X[r_1]$, $s_2 \subseteq Y[s_1]$, where $r_2=\{r^1 \lhd \cdots \lhd r^k \}$ and $s_2=\{s^1\lhd \cdots \lhd s^k \}$. For any $l=1,\ldots, k$ we have
1. $r^l\lhd s^l$,
2. for all $j\leq s_1$Êthere exists no node $y \in Y[j]$ such that $r^l \lhd y \lhd s^l$,
3. for all $i \leq r_1$ there exists no node $x \in X[i]$ such that $r^l \lhd x \lhd s^l$.
It follows immediately from the code that $r_1$ and $s_1$ are macro nodes and that $r_2 \subseteq X[r_1]$, $s_2 \subseteq Y[s_1]$, where $r_2=\{r^1 \lhd \cdots \lhd r^{k_1} \}$ and $s_2=\{s^1\lhd \cdots \lhd s^{k_2} \}$. Due to the macro tree order of the tree and the fact that $X$ represents a deep set, no node in $X[i]$ can be to the right of any node in $X[r_1]$ for $i<r_1$. To prove condition 3 it is thus enough to prove it for $i=r_1$. We proceed by induction on the number $k$Êof iterations of the outer loop. We consider the time right after the $k$th iteration of the loop, i.e., right before the $(k+1)$th iteration. The base case ($k=0$) is trivially satisfied.
For the induction step let $r_i'$ and $s_i'$ for $i=1,2$ be the values of $r_i$ and $s_i$, respectively,Êafter the $(k-1)$th round. There are 3 cases:
1. $r_2'=r_2$ and $s_2'=s_2$.
2. $r_2' \neq r_2$ and $s_2'=s_2$.
3. $r_2' \neq r_2$ and $s_2' \neq s_2$
Let $i^*$Êand $j^*$Êbe the values of $i$ and $j$ at line 11 (“Compare $i$Êand $j$”) in iteration $k$.
Case 1: the claim follows directly from the induction hypothesis.
Case 2: condition 2 from the lemma follows directly from the induction hypothesis. Since $s_2$ and thus also $s_1$ were not changed, $r_2$ was set in case 1 of the procedure and $j^*=s_1$. Therefore, $i^*\lhd j^*$, $r_1=i^*$, and ${\ensuremath{| r_2 |}}=1$. Let $r_2=\{r^1\}$ and $s_2=\{s^1\}$. We have $r_1=i^* \lhd j^* = s_1$ and thus $r^1\lhd s^1$ satisfying condition 1 from the lemma. To prove condition 3 is satisfied we only have to consider the case $i=r_1$. Since $r_2$ was set in case 1 of the procedure, $r^1$ is the rightmost node in $X[r_1]$ and it follows immediately that there exists no node $x \in X[r_1]$ such that $r^1 \lhd x \lhd s^1$.
Case 3: Let $y \in Y[j]$, for $j \leq s_1$, be a node such that $y \not \in s_2$. We will show that $r^l$ is not to the left of $y$. Assume $j=s_1$. Since $s_2' \neq s_2$ then there are two cases depending on which case of the procedure the potential pair was set in. If it is in case 2 of the procedure the claim follows from the correctness of the implementation of [$\mathrm{mop}$]{} and the computation $(r,s)={\ensuremath{\mathrm{mop}}}_{C(i,j,v)}(X[i^*],Y[j^*])={\ensuremath{\mathrm{mop}}}_{C(i,j,v)}(X[i^*],Y[s_1])$. If it is in case 1, then $r_1 =i^* \lhd j^*=s_1$. Since $s_2$ is the leftmost node in $Y[s_1]$ (line 2 in case 1 of the procedure: “Set $(r_1,r_2) := (i, {\ensuremath{\textsc{right}}}_{C(i)}(1,X[i]))$”) the claim follows. Now assume $j < s_1$. We will use that we just proved the claim for $j=s_1$. Assume for the sake of contradiction that there exists a $y \in Y[j]$ such that $r^l \lhd y$. Since $Y$ is representing a deep set and due to the macro tree order of $Y$ this implies $x \lhd y \lhd y'$ for all $y' \in Y[s_1]$ contradicting that the claim is true for $j=s_1$.
\[lem:mopInv2\] We have the following invariant at the beginning of each iteration of [$\textsc{MopSim}$]{}: $$\nexists x \in X[i], \textrm{ such that } x \unlhd x', \textrm{ for any } x' \in r_2.$$
By induction on the number of iterations of the outer loop. In the base case $r_2=\emptyset$ and the condition is trivially satisfied. Note that $X$ is representing a deep set and thus either $x\lhd x'$ or $x' \lhd x$ for all $x \in X[i]$. For the induction step let $i'$, $j'$, and $r_2'$Êbe the values of $i$, $j$, and $r_2$Êrespectively in the iteration before this. By the induction hypothesis $x'\lhd x$ for all $x \in X[i']$Êand $x' \in r_2'$. Due to the macro tree order of $X$ and the fact that $X$ represents a deep set, all nodes in $X[i']$ are to the left of all nodes in $X[i]$. Thus, if $r_2=r_2'$ it follows from the induction hypotheses that $x'\lhd x$ for all $x \in X[i]$Êand $x' \in r_2'=r_2$. For $r_2'\neq r_2$ there are two cases: If $i' \lhd j'$ then $r_2={\ensuremath{\textsc{right}}}_{C(i')}(1,X[i'])$ and $i>i'$ and thus the condition is satisfied. Otherwise $r_2$Êwas set in case 2 of the procedure. Since $r_2\neq r_2'$Êwe have $r_2=r \subseteq X[i']$ and $r\neq \emptyset$. There are two subcases: If $i=j$ or $i=l(v,w)$ and $j=s(v,w)$ (Case 2(a) of the procedure) then $X[i]$Êeither contains a single node, which is the rightmost of the nodes in $X[i']$ that are to the right of all nodes in $r_2$Êor if there are no such nodes $X[i]=\emptyset$. In both cases the condition is satisfied. If $i=s(v,w)$ and $j=r(v,w)$ then $i>i'$ and the condition is satisfied.
\[lem:mopsim\] Procedure [$\textsc{MopSim}$]{} is correctly implemented.
Let $\mathcal{X}$ and $\mathcal{Y}$ be the sets represented by $X$ and $Y$, respectively. We want to show that $$(x,y)\in {\ensuremath{\textsc{MopSim}}}(X,Y) \Leftrightarrow (x,y)\in {\ensuremath{\mathrm{mop}}}(\mathcal{X},\mathcal{Y})\,.$$ Assume $(x,y)\in {\ensuremath{\textsc{MopSim}}}(X,Y)$. Consider the the round where $x$ and $y$Êwere added to $R$ and $S$, respectively. We have $x = r^l \in r_2$ and $y = s^l \in s_2$. We want to show that there is no node $x' \in X[i]$ for any $i$ such that $x \lhd x' \lhd y$ and no node $y' \in Y[j]$ for any $j$ such that $x \lhd y' \lhd y$. By Lemma \[lem:mopInv1\] this is true for $i\leq r_1$ and $j \leq s_1$. By the macro tree order of $Y$ we have that $y \lhd y'$ for any $y' \in Y[j]$ when $j>s_1$. Let $i'$ be the value of $i$ in the round where $x$ and $y$ is added to the output. We will show that no node in $X[i']$ is to the left of any node in $s_2$. Due to the macro tree order of $X$ this implies that no node in $X[i]$ is to the left of any node in $s_2$ for any $i \geq i'$. If $i'=r_1$ then it follows directly from Lemma \[lem:mopInv1\]. If $i'>r_1$ it follows from the implementation of the procedure that $i'$ is the first non-empty entry in $X$ greater than $r_1$. Thus the claim follows for any $j$. We now return to show that no node in $X[i']$ is to the left of any node in $s_2$. There are two cases depending on whether $j=s_1$ or $j>s_1$. If $j > s_1$ then $j$ was changed either in one of the four cases I–IV, or in the previous iteration in case 2. If $j$ was equal to $s_1$ at the beginning of this iteration then $j$ was incremented in one of the four cases I–IV. Thus none of the cases applied to $s_1$. By Proposition \[lem:orderlemma\] no node in $X[i']$ can be to the left of a node in $X[s_1]$. Since $s_2 \subseteq X[s_1]$ the claim follows. If $j=s_1$ it follows from case 2 that ${\ensuremath{\textsc{left}}}(X[i'],s_2)=\emptyset$ (otherwise the potential pairs would not have been added to the output in this iteration) and the claim follows immediately.
Now assume $(x,y)\in {\ensuremath{\mathrm{mop}}}(\mathcal{X},\mathcal{Y})$. We will deal with each of the cases from Proposition \[lem:orderlemma\] separately.
1. Case (i): $c(x)=c(y)=r(v,w)$.
2. Case (i): $c(x)=c(y)=l(v,w)$.
3. Case (ii): $c(x)=c(y)=l(v)$.
4. Case (iii): $c(x)=l(v,w)$ and $c(y)=s(v,w)$.
5. Case (iv): $c(x)=s(v,w)$ and $c(y)=r(v,w)$.
6. Case (v): $c(x)=l(v,w)$ and $c(y)=r(v,w)$.
7. Case (v): $c(x)\lhd c(y)$ and $c(x)$ and $c(y)$ belong to different clusters.
Note that if $c(x)\lhd c(y)$ then $x$ is the rightmost node in $X[c(x)]$ and $y$ is the leftmost node in $Y[c(y)]$. We first show that in all cases we will have $x= r^l \in r_2$ and $y=s^l \in s_2$ for some $l$ at some iteration. Consider the first iteration where either $x\in X[i]$ or $y\in Y[j]$. Let $i'$ and $j'$ be the values of $i$ and $j$, respectively, in this iteration. There are three cases:
- $x \in X[i']$ and $y \in Y[j']$. For case 1–5 the procedure goes into case 2. From the correctness of [$\textsc{mop}$]{} we get $x \in r$ and $y \in s$. Thus $r\neq \emptyset$ and we set $(r_1,r_2)=(i',r)$ and $(s_1,s_2)=(j',s)$ and the claim follows. For case 6–7 the procedure goes into case 1. Since this iteration is the first where $y\in Y[j]$ we have $j'>s_1$ and we set $(r_1,r_2)=(i',{\ensuremath{\textsc{right}}}_{C(i')}(1,X[i']))$ and $(s_1,s_2)=(j',{\ensuremath{\textsc{left}}}_{C(j')}(1,Y[j']))$. Since $x$ is the rightmost node in $X[i']$ and $y$ is the leftmost node in $Y[c(j')]$ the claim follows.
- $x \in X[i']$ and $y \not \in Y[j']$. Since $(x,y)\in {\ensuremath{\mathrm{mop}}}(\mathcal{X}, \mathcal{Y})$ this implies $j'<c(y)$ and there exists no node $y' \in Y[j']$ such that $x \lhd y'$. Assume that there existed such a $y'$. Then $x \lhd y' \lhd y$ due to the macro tree order of $Y$ contradicting $(x,y)\in {\ensuremath{\mathrm{mop}}}(\mathcal{X}, \mathcal{Y})$. Thus $i' \not\!\!\lhd\; j'$. From case I–IV of the procedure it follows that either $i'=j'$, $i'=l(v,w)$ and $j'=s(v,w)$, or $i'=s(v,w)$ and $j'=r(v,w)$. From this and $j'<c(y)$ it follows that we are in case 4 or 7 from above.
The procedure enters case 2 in this iteration. If we are in case 4 then $i'=l(v,w)=j'$ and $c(y)=s(v,w)$. If $r=\emptyset$ then $i=i'$, $X[i']$ is unchanged, and $j=j'+1=s(v,w)=c(y)$ at the end of this iteration. If $r\neq \emptyset$ then $x$ must be to the right of all nodes in $x'\in r$. Assume that there is a $x' \in r$ such that $x \lhd x'$. Since $x' \in r$ there exists a node $y' \in s$ such that $x' \lhd y' \lhd y$. That $y' \lhd y$ follows from $y' \in l(v,w)$ and $y\in s(v,w)$ and the assumption that $\mathcal{Y}$ is deep. Thus $x \lhd x' \lhd y' \lhd y$ contradicting that $(x,y)\in {\ensuremath{\mathrm{mop}}}(\mathcal{X}, \mathcal{Y})$. Therefore, $i=i'$, $x \in X[i']$ and $j=j'+1=s(v,w)=c(y)$ at the end of this iteration. From case I of the procedure and the analysis of case (a) it follows that $x= r^l \in r_2$ and $y=s^l \in s_2$ for some $l$.
Now assume we are in case 7. By the same argument as before $i=i'$, $x \in X[i]$, and $j >j'$ at the end of this iteration. Unless $i'=l(v,w)=j'$ this implies that $i \lhd j$ at line 11 (“Compare $i$ and $j$”) in the next iteration. If $i'=l(v,w)=j'$ then either $i \lhd j$ after the first loop in the next iteration, and the claim follows as before, or $i=l(v,w)$ and $j=s(v,w)$. In the last case we get into case (b) again, but it follows from the analysis that in the iteration after the next we will have $i \lhd j=c(y)$. The claim follows from the analysis of case (a).
- $x \not \in X[i']$ and $y \in Y[j']$. It follows by inspection of the cases that unless we are in case 1 we have $i' \lhd j'$. If we are in case 1 ($j'=c(x)=r(v,w)=c(y)$) we have either $i' \lhd j'$ or $i'=s(v,w)$. First we consider the cases 2–7. Since $i' \lhd j'$ the procedure enters case 1 in this iteration. Thus $i$ is incremented and $j$ stays the same. This happens until $i=c(x)$. Now consider case 1. If $i' \lhd j'$ the procedure enters case 1 in this iteration. Thus $i$ is incremented and $j$ stays the same. In the next iteration either the same happens or $i'=s(v,w)$. If $i'=s(v,w)$ the procedure enters case 2. Since $i'$ is a spine node and $\mathcal{X}$ is deep, $X[i]$ contains only one node $x'$. By the structure of the macro tree and the assumption that $\mathcal{X}$ is deep $x' \lhd x$. Since $x \lhd y \in Y[j']$ this implies $r \neq \emptyset$. It follows from case 2(b) of the procedure that $i$ is incremented while $j$ stays the same. At line 11 (“Compare $i$ and $j$”) in the next iteration we will have $j=i=r(v,w)$ since all entries in $X$ between $j'$ and $r(v,w)$ are empty due to the assumption that $\mathcal{X}$ is deep. The claim follows from the analysis in case (a).
It remains to show that once $x=r^l \in r_2$ and $y=s^l \in s_2$ they will stay this way until added to the output. Consider the iteration where $x$ and $y$ are assigned to $r_2$ and $s_2$. At the end of this iteration either $i$ or $j$ or both are incremented. Assume $j$ is incremented while the potential pairs are still unchanged. Since $j$ is incremented we have $s_1 < j$ until $s_1$ is changed. It follows from case 1 and 2 of the procedure that in this case $(r_1,r_2)$ is only changed if at the same time $(s_1,s_2)$ are changed and right before that $(r_1,r_2)$ and $(s_1,s_2)$ are added to the output.
Consider first case 1–3. If $i$ is incremented then $j$ is incremented in one the cases I–V in the next iteration since $i'=j'$. By the above argument $x$ and $y$ are added to the output. For case 4 $j$ is incremented (case 2(a) of the procedure) and the claim follows as before. For case 5–7 first note that $r_2$ and $s_2$ contain only one node each, i.e., $x= r_2$ and $y= s_2$. For case 5 $i$ is incremented (case 2(b) of the procedure). Since $\mathcal{X}$ is deep we have $i\geq r(v,w)=j'$ at line 11 (“Compare $i$ and $j$”) in the next iteration. If $i > r(v,w)$ then $j>j'$ and the claim follows. If $i=r(v,w)$ the procedure enters case 2. If $r = \emptyset$ then $j$ is incremented and the claim follows. If $r \neq \emptyset$ then $s_1 =j$ and $(x,y) \in {\ensuremath{\mathrm{mop}}}(\mathcal{X},\mathcal{Y})$ implies ${\ensuremath{\textsc{left}}}_{C(i,j)}(X[i], s_2) ={\ensuremath{\textsc{leftof}}}_{C(i,j)}(X[i], y)= \emptyset$. Thus $(r_1,r_2)$ and $(s_1,s_2)$ are added to the output. If we are in case 6 and 7, $i$ is incremented. Consider case 6. Since $(x,y) \in {\ensuremath{\mathrm{mop}}}(\mathcal{X},\mathcal{Y})$ all entries in $X$ between $l(v,w)$ and $r(v,w)$ are empty. Thus at line 11 (“Compare $i$ and $j$”) in the next iteration $i\geq r(v,w)$. The proof is equivalent to the one for case 5. Consider case 7. If $j$ is a boundary node then all entries in $X$ between $c(x)$ and $j$ are empty. Thus $j$ is incremented in the second loop of the next iteration. For all other cases for $j$ the proof is similar to the proof of case 5.
\[inv:match\] In procedure [$\textsc{Match}$]{} we have the following invariant of $X[i]$ and $Y[j]$ after line 2 in the main loop (“Repeat until”): $${\ensuremath{\textsc{left}}}(1,X[i])=\mathcal{X}_l \textrm{ and } {\ensuremath{\textsc{left}}}(1,Y[j])=\mathcal{Y}_l \textrm { for some } l \;.$$
Induction on the number of iterations of the outer loop. Base case: In the first iteration $X[i]$ and $Y[j]$ are the first nonempty entries in $X$ and $Y$ and thus ${\ensuremath{\textsc{left}}}(1,X[i])=\mathcal{X}_1 \textrm{ and } {\ensuremath{\textsc{left}}}(1,Y[j])=\mathcal{Y}_1$. For the induction step let $i'$Êand $j'$Êbe the values of $i$Êand $j$ in the previous iteration. By the induction hypothesis ${\ensuremath{\textsc{left}}}(1,X[i'])=\mathcal{X}_{l'} \textrm{ and } {\ensuremath{\textsc{left}}}(1,Y[j'])=\mathcal{Y}_{l'}$. If $x=|X[i']|= |Y[j']|$ both $i$ and $j$Êwere incremented and ${\ensuremath{\textsc{left}}}(1,X[i])=\mathcal{X}_{l'+x}$ and ${\ensuremath{\textsc{left}}}(1,Y[j])=\mathcal{Y}_{l'+x}$. If $x=|X[i']|< |Y[j']|$ then $i$ was incremented implying ${\ensuremath{\textsc{left}}}(1,X[i])=\mathcal{X}_{l'+x}$. In that case $j=j'$ and $Y[j]={\ensuremath{\textsc{left}}}(x,Y[j'])$ implying ${\ensuremath{\textsc{left}}}(1,Y[j])=\mathcal{Y}_{l'+x}$. Similarly, if $|X[i']|> |Y[j']|=y$ we have ${\ensuremath{\textsc{left}}}(1,X[i])=\mathcal{X}_{l'+y}$. In that case $j=j'$ and ${\ensuremath{\textsc{left}}}(1,Y[j])=\mathcal{Y}_{l'+y}$.
\[lem:matchmacro\] Procedure [$\textsc{Match}$]{} is correctly implemented.
We need to show that for all $1\leq k\leq |\mathcal{X}|$: $\mathcal{X}_{k} \in {\ensuremath{\textsc{Match}}}(X,Y,Y') \Leftrightarrow \mathcal{X}_{k} \in \{\mathcal{X}_j | \mathcal{Y}_j \in \mathcal{Y}\}$. Consider the iteration where $\mathcal{X}_{k} \in X[i]$ and $\mathcal{Y}_{k} \in Y[j]$. By Lemma \[inv:match\]Ê such an iteration exists. If $Y[j]=Y[j']$ then $\mathcal{Y}_{k} \in \mathcal{Y'}$ implying $\mathcal{X}_{k} \in \{\mathcal{X}_j | \mathcal{Y}_j \in \mathcal{Y}\}$. It follows from the implementation of case 1(a) and 1(b) that if $x \leq y$ all nodes in $X[i]$Êare added to the output and thus $\mathcal{X}_{k} \in {\ensuremath{\textsc{Match}}}(X,Y,Y')$. If $x >y$ then $\mathcal{X}_{k} \in {\ensuremath{\textsc{left}}}(y,X[i])$ since $\mathcal{Y}_{k} \in Y[j]$ and thus $\mathcal{X}_{k} \in {\ensuremath{\textsc{Match}}}(X,Y,Y')$.
If $Y[j]\neq Y'[j]$ the procedure calls ${\ensuremath{\textsc{match}}}$ with some subset of $X[i]$, $Y[j]$,Êand $Y'[j]$ depending on the size of $x$Êand $y$. By Lemma \[inv:match\]Êand the correctness of [$\textsc{match}$]{} it follows that $\mathcal{X}_{k} \in {\ensuremath{\textsc{Match}}}(X,Y,Y') \Leftrightarrow \mathcal{X}_{k} \in \{\mathcal{X}_j | \mathcal{Y}_j \in \mathcal{Y}\}$.
Procedure [$\textsc{MopRight}$]{} is correctly implemented.
Follows from the correctness of [$\textsc{MopSim}$]{} (Lemma \[lem:mopsim\]) and [$\textsc{Match}$]{} (Lemma \[lem:matchmacro\]).
Finally, we consider correctness of the [$\textsc{Fl}$]{} procedure.
Procedure [$\textsc{Fl}$]{} is correctly implemented.
Let $\mathcal{X}$ denote the set represented by $X$Êand let $F=\{{\ensuremath{\mathrm{fl}}}(x,\alpha)| x\in \mathcal{X} \}$. To show ${\ensuremath{\textsc{Fl}}}(X,\alpha) \subseteq F$ we will first show that for any node $x$ added to $R$ during the computation $x \in F$. Consider a node $x \in R[i]$ for some $i$. Either $x$Êwas added directly to $R$ after a computation of $N$ in one of the three cases of the procedure or it was added after the computation of $S$. In the first case $x \in F$ follows from the correctness of ${\ensuremath{\textsc{Fl}}}_C$. If $x$Êwas added after the computation of $S$ it follows from the correctness of ${\ensuremath{\textsc{Fl}}}_M$ that $x \in C(i)$ for some $i \in S$. Due to the correctness of ${\ensuremath{\textsc{Fl}}}_C$ we have $x \in F$.
To show ${\ensuremath{\textsc{Deep}}}(F) \subseteq {\ensuremath{\textsc{Fl}}}(X,\alpha)$ we use Proposition \[lem:ancestorlemma\]. Let $x$Êbe a node in ${\ensuremath{\textsc{Deep}}}(F)$ and let $x'$ be a node in $X$ such that ${\ensuremath{\mathrm{fl}}}(x',\alpha)=x$. We have $x' \in X[i]$Êfor some $i$. If $i$ is a left or right node then according to Proposition \[lem:ancestorlemma\] $x$ can be in $i$ (case (i)), on the spine (case (ii)), in the top boundary node (case (ii)), or in an ancestor of $i$ in the macro tree (case (iii)). If $x$ is in the same cluster as $x'$ then it follows from the correctness of ${\ensuremath{\textsc{Fl}}}_C$ that $x \in N$.ÊThus $x$ is added to $R$ and due to the correctness of ${\ensuremath{\textsc{Deep}}}$ we have $x \in {\ensuremath{\textsc{Fl}}}(X,\alpha)$. If $c(x)$ is in a different cluster than $c(x')$ then $c(x)$ is an ancestor of $c(x')$ in the macro tree due to Proposition \[lem:ancestorlemma\]. Since $x \in {\ensuremath{\textsc{Deep}}}(F)$ we have $N=\emptyset$ and thus ${\ensuremath{\mathrm{parent}}}(v) \prec_M c(x')$ is added to $L$. It follows from the correctness of ${\ensuremath{\textsc{Fl}}}_M$ that $c(x) \in S$. Due to the structure of the macro tree $c(x)$ is either a boundary node or a spine node and thus $x={\ensuremath{\mathrm{fl}}}_{C(c(x))}({\ensuremath{\mathrm{first}}}(c(x)),\alpha)={\ensuremath{\textsc{Fl}}}_{C(c(x))}({\ensuremath{\mathrm{first}}}(c(x)),\alpha)$. The last equality follows from the correctness of ${\ensuremath{\textsc{Fl}}}_C$. That $x \in {\ensuremath{\textsc{Fl}}}(X,\alpha)$ now follows from the above analysis showing that only nodes from $F$ are added to $R$ and the correctness of [$\textsc{Deep}$]{}.
If $i$Êis a leaf node then $x$ can be in $i$ (case (i)), in the top boundary node (case (iii)), or in an ancestor of $i$Êin the macro tree (case (iii)). The correctness follows by an analysis similar to the one for the previous case. If $i$Êis a spine node or a boundary node, then $x$Êis either in $i$ (case (i)) or in an ancestor of $i$Ê in the macro tree (case (iii)). The correctness follows by an analysis similar to the one for the first case.
Complexity of the Tree Inclusion Algorithm
------------------------------------------
To analyze the complexity of the node array implementation we first bound the running time of the above implementation of the set procedures. All procedures scan the input from left-to-right while gradually producing the output. In addition to this procedure ${\ensuremath{\textsc{Fl}}}$ needs a call to a node list implementation of ${\ensuremath{\textsc{Fl}}}$ on the macro tree. Given the data structure described in Section \[sec:preprocessing\] it is easy to check that each step in the scan can be performed in $O(1)$ time giving a total of $O(n_{T}/\log n_T)$ time. Since the number of nodes in the macro tree is $O(n_{T}/\log n_T)$, the call to the node list implementation of ${\ensuremath{\textsc{Fl}}}$ is easily done within the same time. Hence, we have the following lemma.
\[lem:auxmacro\] For any tree $T$ there is a data structure using $O(n_T)$ space and $O(n_T\log n_{T})$ preprocessing time which supports all of the set procedures in $O(n_T/\log n_T)$ time.
Next consider computing the deep occurrences of $P$ in $T$ using the procedure ${\ensuremath{\textsc{Emb}}}$ of Section \[sec:recursion\] and Lemma \[lem:auxmacro\]. The following lemma bounds the space usage.
\[lem:lin\_space\_wc\] The total size of the saved embeddings at any time during the computation of ${\ensuremath{\textsc{Emb}}}({\ensuremath{\mathrm{root}}}(P))$ is $O(n_T)$.
Let $v$Êbe the node for which we are currently computing [$\textsc{Emb}$]{}. Let $p$Êbe the path from the root to $v$ and let $w_0,\ldots, w_l$Êbe the light nodes on this path. We have $l={\ensuremath{\mathrm{ldepth}}}(v)$. As in the proof of Lemma \[lem:lin\_space\] it suffices to bound ${\ensuremath{| {\ensuremath{\textsc{Emb}}}({\ensuremath{\mathrm{heavy}}}({\ensuremath{\mathrm{parent}}}(w_i))) |}}$ for all $i$. Assume that $l_P \leq l_T$ (otherwise we can check this in linear time and conclude that $P$ cannot be included in $T$). Each of the node arrays use $O(n_T/\log n_T)$ space and therefore by Corollary \[cor:ldepth\] we have that $\sum_{i=1}^{l}{\ensuremath{| {\ensuremath{\textsc{Emb}}}({\ensuremath{\mathrm{heavy}}}({\ensuremath{\mathrm{parent}}}(w_i))) |}}= O(n/\log n_T
\cdot \log l_P) = O(n_T)$.
For the time complexity note that during the computation of ${\ensuremath{\textsc{Emb}}}({\ensuremath{\mathrm{root}}}(P))$ each node $v \in V(P)$ contributes a constant number of calls to the set procedures. Hence, the total time used by the algorithm is $O(n_Pn_T/\log n_T + n_{T}\log n_{T})$. Thus we have shown the following.
\[thm:complex\] For trees $P$ and $T$ the tree inclusion problem can be solved in $O(n_Pn_T/\log n_T + n_{T}\log n_{T})$ time and $O(n_T)$ space.
Combining the results in Theorems \[thm:simple\], \[thm:complex\] and Corollary \[cor:simple\] we have the main result of Theorem \[thm:main\].
Conclusion
==========
We have presented three algorithms for the tree inclusion problem, which match or improve the best known time complexities while using only linear space. We believe that some of the new ideas are likely to be of both practical and theoretical value in future work. From a practical perspective, space is a common bottleneck for processing large data sets and hence reducing the space can significantly improve performance in practice. From a theoretical perspective, we have introduced several non-trivial algorithms to manipulate sets of nodes in trees that may have applications to other problems. For instance, the ${\ensuremath{\textsc{Nca}}}$ procedure from Section \[micromacro\] computes multiple nearest common ancestor queries in time *sublinear* in the size of input sets.
Acknowledgments {#acknowledgments .unnumbered}
===============
We would like to thank the anonymous reviewers of earlier drafts of this paper for many valuable comments that greatly improved the quality of the paper. We would also especially like to thank the reviewer who discovered the error in the space complexity of the original draft.
[^1]: Technical University of Denmark, Department of Informatics and Mathematical Modelling. This work is part of the DSSCV project supported by the IST Programme of the European Union (IST-2001-35443).
[^2]: Corresponding author: Technical University of Denmark, Department of Informatics and Mathematical Modelling, Building 322, Office 124, DK-2800 Kongens Lyngby, Denmark. Phone: (+45) 45 25 36 73. Fax: (+45) 45 88 26 73. Email: [[email protected]]{}.
[^3]: An extended abstract of this paper appeared in Proceedings of the 32nd International Colloquium on Automata, Languages and Programming, Lecture Notes in Computer Science, vol. 3580, pp. 66-77, Springer-Verlag, 2005.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'We investigate the effect of large-$x$ resummation on parton distributions by performing a fit of Deep Inelastic Scattering data from the NuTeV, BCDMS and NMC collaborations, using NLO and NLL soft-resummed coefficient functions. Our results show that soft resummation has a visible impact on quark densities at large $x$. Resummed parton fits would therefore be needed whenever high precision is required for cross sections evaluated near partonic threshold.'
author:
- 'G. Corcella'
- 'L. Magnea'
title: |
Impact of large-x resummation\
on parton distribution functions [^1]
---
CERN-PH-TH/2005-118
[ address=[CERN, Department of Physics, Theory Division, CH-1211 Geneva 23, Switzerland]{} ]{}
[ address=[ Università di Torino and INFN, Sezione di Torino, Via P. Giuria 1, I-10125, Torino, Italy]{} ]{}
A precise knowledge of parton distribution functions (PDF’s) at large $x$ is important to achieve the accuracy goals of the LHC and other high energy accelerators. We present a simple fit of Deep Inelastic Scattering (DIS) structure function data, and extract NLO and NLL-resummed quark densities, in order to establish qualitatively the effects of soft-gluon resummation.
Structure functions $F_i(x,Q^2)$ are given by the convolution of coefficient functions and PDF’s. Finite-order coefficient functions present logarithmic terms that are singular at $x = 1$, and originate from soft or collinear gluon radiation. These contributions need to be resummed to extend the validity of the perturbative prediction. Large-$x$ resummation for the DIS coefficient function was performed in [@Sterman:1986aj; @Catani:1989ne] in the massless approximation, and in [@Laenen:1998kp; @Corcella:2003ib] with the inclusion of quark-mass effects, relevant at small $Q^2$.
Soft resummation is naturally performed in moment space, where large-$x$ terms correspond, at ${\cal O}(\alpha_s)$, to single ($\alpha_s \ln N$) and double ($\alpha_s \ln^2 N$) logarithms of the Mellin variable $N$. In the following, we shall consider values of $Q^2$ sufficiently large to neglect quark-mass effects. Furthermore, we shall implement soft resummation in the next-to-leading logarithmic (NLL) approximation, which corresponds to keeping terms ${\cal O}
(\alpha_s^n \ln^{n+1} N )$ (LL) and ${\cal O} (\alpha_s^n \ln^n N)$ (NLL) in the Sudakov exponent.
To gauge the impact of the resummation on the DIS cross section, we can evaluate the charged-current (CC) structure function $F_2$ convoluting NLO and NLL-resummed [$\overline{\mbox{MS}}$ ]{}coefficient functions with the NLO PDF set CTEQ6M [@Pumplin:2002vw]. We consider $Q^2 = 31.62$ GeV$^2$, since it is one of the values of $Q^2$ at which the NuTeV collaboration collected data [@Naples:2003ne]. In Fig. \[fdel\] we plot $F_2(x)$ with and without resummation (Fig. 1a), as well as the normalized difference $\Delta = (F_2^{\mathrm{res}} - F_2^{\mathrm{NLO}})/F_2^{\mathrm{NLO}}$ (Fig. 1b). We note that the effect of the resummation is an enhancement of $F_2$ for $x > 0.6$. Such an enhancement is compensated by a decrease at smaller $x$: the resummation, in fact, does not change the first moment of $F_2$, since we include in the Sudakov exponent only terms $\sim \ln^k N$, which vanish for $N = 1$.
Our predictions for $F_2$ at different values of $Q^2$ can be compared with NuTeV data at large $x$. The results of the comparison are shown in Fig. \[fignut\]: although the resummation moves the prediction towards the data, we are still unable to reproduce the large-$x$ data.
Several effects are involved in the mismatch: at very large values of $x$, power corrections will certainly play a role. Moreover, we have used so far a parton set (CTEQ6M), extracted by a global fit which did not account for the NuTeV data. Rather, data from the CCFR experiment [@Yang:2000ju], which disagree at large $x$ with NuTeV [@Naples:2003ne], were used. The discrepancy has recently been described as understood [@tzanov]; however, it is not possible to draw any firm conclusion from our comparison.
We wish to reconsider the CC data in the context of an indipendent fit. We shall use NuTeV data on $F_2(x)$ and $x F_3(x)$ at $Q^2 = 31.62$ GeV$^2$ and 12.59 GeV$^2$, and extract NLO and NLL-resummed quark distributions from the fit. $F_2$ contains a gluon-initiated contribution $F_2^g$, which is not soft-enhanced and is very small at large $x$: we can therefore safely take $F_2^g$ from a global fit, e.g. CTEQ6M, and limit our fit to the quark-initiated term $F_2^q$. We choose a parametrization of the form $F_2^q (x) = F_2 (x) - F_2^g (x) = A
x^{- \alpha} (1 - x)^\beta (1 + b x)$; $ x F_3(x) = C x^{-\rho}
( 1 - x )^\sigma ( 1 + k x )$. The best-fit parameters and the $\chi^2$ per degree of freedom are quoted in [@noi]. In Fig. \[fits\], we present the NuTeV data on $F_2(x)$ and $x F_3(x)$ at $Q^2 = 12.59$ GeV$^2$, along with the best-fit curves. Similar plots at $Q^2 = 31.62$ GeV$^2$ are shown in Ref. [@noi].
In order to extract individual quark distributions, we need to consider also neutral current data. We use BCDMS [@Benvenuti:1989fm] and NMC [@Arneodo:1996qe] results, and employ the parametrization of the nonsinglet structure function $F_2^{\mathrm{ns}} = F_2^p - F_2^D$ provided by Ref. [@DelDebbio:2004qj]. The parametrization [@DelDebbio:2004qj] is based on neural networks trained on Monte-Carlo copies of the data set, which include all information on errors and correlations: this gives an unbiased representation of the probability distribution in the space of structure functions.
Writing $F_2$, $x F_3$ and $F_2^\mathrm{ns}$ in terms of their parton content, we can extract NLO and NLL-resummed quark distributions, according to whether we use NLO or NLL coefficient functions. We assume isospin symmetry of the sea, i.e. $s = \bar s$ and $\bar u = \bar d$, we neglect the charm density, and impose a relation $\bar s = \kappa \, \bar u$. We obtain a system of three equations, explicitly presented in [@noi], that can be solved in terms of $u$, $d$ and $s$. We begin by working in $N$-space, where the resummation has a simpler form and quark distributions are just the ratio of the appropriate structure function and coefficient function. We then revert to $x$-space using a simple parametrization $q(x) = D x^{-\gamma}(1 - x)^\delta$.
Figs. \[up\]–\[up1\] show the effect of the resummation on the up-quark distribution at $Q^2 = 12.59$ and 31.62 GeV$^2$, in $N$- and $x$-space respectively. The best-fit values of $D$, $\gamma$ and $\delta$, along with the $\chi^2/\mathrm{dof}$, can be found in [@noi]. The impact of the resummation is noticeable at large $N$ and $x$: there, soft resummation enhances the coefficient function and its moments, hence it suppresses the quark densities extracted from structure function data. In principle, also $d$ and $s$ densities are affected by the resummation; the errors on their moments, however, are too large for the effect to be statistically significant. In [@noi] it was also shown that the results for the up quark at 12.59 and 31.62 GeV$^2$ are consistent with NLO perturbative evolution.
In summary, we have presented a comparison of NLO and NLL-resummed quark densities extracted from large-$x$ DIS data. We found a suppression of valence quarks in the $10-20 \%$ range at $x > 0.5$, for moderate $Q^2$. We believe that it would be interesting and fruitful to extend this analysis and include large-$x$ resummation in the toolbox of global fits. Our results show in fact that this would be necessary to achieve precisions better than $10 \%$ in processes involving large-$x$ partons.
[99]{}
G. Sterman, *Nucl. Phys. B* [**281**]{} (1987) 310. S. Catani and L. Trentadue, *Nucl. Phys. B* [**327**]{} (1989) 323.
E. Laenen and S. O. Moch, *Phys. Rev. D* [**59**]{} (1999) 034027.
G. Corcella and A. D. Mitov, *Nucl. Phys. B* [**676**]{} (2004) 346.
J. Pumplin, D. R. Stump, J. Huston, H. L. Lai, P. Nadolsky and W. K. Tung, *JHEP* [**0207**]{} (2002) 012.
D. Naples [*et al.*]{} \[NuTeV Collaboration\], hep-ex/0307005.
U. K. Yang [*et al.*]{} \[CCFR/NuTeV Collaboration\], *Phys. Rev. Lett. * [**86**]{} (2001) 2742.
M. Tzanov [*et al.*]{} \[NuTeV Collaboration\], these proceedings.
G. Corcella and L. Magnea, hep-ph/0506278.
A. C. Benvenuti [*et al.*]{} \[BCDMS Collaboration\], *Phys. Lett. B* [**237**]{} (1990) 592.
M. Arneodo [*et al.*]{} \[New Muon Collaboration\], *Nucl. Phys. B* [**483**]{} (1997) 3.
L. Del Debbio, S. Forte, J. I. Latorre, A. Piccione and J. Rojo, *JHEP* [**0503**]{} (2005) 080.
[^1]: Talk given by G. Corcella at DIS 2005, XIII Workshop on Deep Inelastic Scattering, April 27–May 1, 2005, Madison, WI, U. S. A.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'Energy resolution, $\alpha/\beta$ ratio, pulse-shape discrimination for $\gamma$ rays and $\alpha$ particles, radioactive contamination were studied with neodymium doped yttrium-aluminum garnet (YAG:Nd). Applicability of YAG:Nd scintillators to search for $2\beta$ decay and $\alpha$ activity of natural neodymium isotopes are discussed.'
---
-0.5cm -0.5cm
0.2cm
F.A. Danevich[^1], V.V. Kobychev, S.S. Nagorny, V.I. Tretyak
0.2cm
[*Institute for Nuclear Research, MSP 03680 Kiev, Ukraine*]{}
0.5cm
0.4cm
PACS numbers: 29.40.Mc; 23.40.-s; 23.60.+e; 95.35.+d
0.4cm
Keywords: Scintillation detector, Double beta decay, Alpha decay, Dark matter, Neodymium doped yttrium-aluminum garnet
1.0cm
INTRODUCTION
============
The great interest to the double beta ($2\beta$) decay search is related, in particular, with the recent evidence of neutrino oscillations which strongly suggests that neutrinos have nonzero mass [@Ver02; @Futur; @Ell02; @DBD-tab]. While oscillation experiments are sensitive to the neutrinos mass difference, only the measured neutrinoless ($0\nu$) double beta decay rate could set the Majorana nature of the neutrinos and give the absolute scale of the effective neutrino mass.
$^{150}$Nd is one of the most promising candidate for $2\beta$ decay study due to its large transition energy ($Q_{2\beta}$). As a result, the calculated value of the phase space integral $G_{mm}^{0\nu}$ of the $0\nu2\beta$ decay of $^{150}$Nd is the largest one among 35 possible $2\beta^-$ decay candidates [@Doi85; @Suh98]. The theoretical predictions of the product $T_{1/2}^{0\nu}\cdot\langle m_\nu\rangle^2$ are in the interval $3.4\times 10^{22}-3.4\times 10^{24}$ yr (half-lives for the neutrino mass mechanism with $\langle m_{\nu} \rangle = 1$ eV) [@NME]. Moreover, the larger the $Q_{2\beta}$ energy, the simpler, from an experimental point of view, is to overcome background problems. Note that the background from natural radioactivity drops sharply above 2615 keV, which is the energy of the $\gamma$ from $^{208}$Tl decay ($^{232}$Th family). In addition, a contribution of cosmogenic activation, which is important problem of the next generation $2\beta$ decay experiments [@CARVEL], decreases at higher energies.
There exist no appropriate detector containing neodymium, which could serve as both source and detector simultaneously (so called “active” source experimental method [@Futur]). It should be noted that this experimental approach will assure high registration efficiency which is especially important if very expensive enriched $^{150}$Nd isotope would be used in an experiment.
The purpose of our work was investigation of scintillation properties and preliminary check of radioactive contamination of neodymium doped yttrium-aluminum garnet (YAG:Nd) as a possible detector for double beta decay experiment with $^{150}$Nd. The active source method allows also to investigate other rare processes in Nd nuclei, such as $\alpha$ decay of naturally occurring neodymium isotopes.
MEASUREMENTS AND RESULTS
========================
Energy resolution and $\alpha/\beta$ ratio
------------------------------------------
Main properties of YAG:Nd (chemical formula Y$_3$Al$_5$O$_{12}$:Nd) are presented in Table 1. These crystals are well known and their production is well developed due to their wide application in laser technique. The scintillation properties of cerium doped yttrium-aluminum garnet (YAG:Ce) have been studied in [@YAG:Ce], properties of YAG:Yb have been reported in [@Anto01], whereas YAG:Nd, to our knowledge, was never studied as a scintillator.
[|l|l|]{} Density (g/cm$^3$) & 4.56\
Melting point & 1970$^\circ$C\
Crystal structure & Cubic Garnet\
Hardness (Mohs) & 8.5\
Refractive index & 1.82\
Average decay time$^{\ast}$ & 4 $\mu$s\
Photoelectron yield relatively to NaI(Tl)$^{\ast}$ & 8%\
\
All the measurements were carried out for the $\oslash 17 \times
6$ mm YAG:Nd crystal with $\approx$2 mol% of Nd. The mass of the crystal is 7.16 g. Photoelectron yield was estimated with the Philips XP2412 bialkali photomultiplier (PMT) as 8% of NaI(Tl). The energy resolution FWHM=13.6% was measured for 662 keV $\gamma$ line of $^{137}$Cs with the YAG:Nd crystal wrapped by PTFE reflector tape and optically coupled to the PMT XP2412. A substantial improvement of the light output ($\approx$22%) and energy resolution was achieved by placing the crystal in liquid (silicone oil with index of refraction $\approx$1.5). The crystal was fixed in center of the teflon container $\oslash 70 \times 90$ mm and viewed by two PMTs XP2412. Fig. 1 demonstrates the energy spectra of $^{137}$Cs and $^{207}$Bi obtained in such a way with the YAG:Nd crystal. The energy resolution of 9.3% was obtained with the YAG:Nd crystal scintillator for 662 keV $\gamma$ line of $^{137}$Cs.
We estimate the energy resolution of the YAG:Nd scintillation detector at the energy $Q_{2\beta}$ of $^{150}$Nd using the results of measurements with $^{137}$Cs and $^{207}$Bi $\gamma$ sources. These data were fitted by function FWHM(keV)$~=a+\sqrt{b\times E_{\gamma}}$ with values $a=2$ keV and $b=5.2$ keV, where energy of $\gamma$ quanta $E_{\gamma}$ is in keV. The energy resolution FWHM$\approx$4% could be achieved at the $Q_{2\beta}$ energy of $^{150}$Nd.
The $\alpha/\beta$ ratio was measured with the help of the collimated $\alpha$ particles of a $^{241}$Am source. As it was checked by surface-barrier detector, the energy of $\alpha$ particles was reduced to about 5.25 MeV by $\approx 1$ mm of air, due to passing through the collimator. Fig. 2 shows the energy spectrum of the $\alpha$ particles measured by the YAG:Nd scintillator. The $\alpha/\beta$ ratio is 0.33. The YAG:Nd crystal was irradiated in three perpendicular directions with aim to check a possible dependence of the $\alpha $ signal on direction of irradiation. While earlier such a dependence has been found for CdWO$_4$ [@W-alpha] and ZnWO$_4$ [@ZWO] scintillators, we did not observe it for YAG:Nd crystal.
Pulse-shape discrimination for $\gamma$ quanta and $\alpha$ particles
---------------------------------------------------------------------
The shapes of scintillation light pulses in YAG:Nd crystals were studied for $5.25$ MeV $\alpha$ particles and $\approx 1.5-1.8$ MeV $\gamma$ quanta with the help of a 12 bit 20 MHz transient digitizer as described in [@W-alpha; @Faz98]. The pulse shape can be fitted by sum of exponential functions:
$f(t)=\sum A_{i}/(\tau_{i}-\tau_{0})\cdot (e^{-t/\tau
_{i}}-e^{-t/\tau _{0}}),\qquad t>0$,
where $A_{i}$ are intensities (in %), and $\tau_{i}$ are decay constants for different light emission components, $\tau_{0}$ is integration constant of electronics ($\approx$0.2 $\mu$s). The values of $A_{i}$ and $\tau_{i}$ obtained by fitting the average of $\approx$3 thousand individual $\alpha$ and $\gamma$ pulses in the time interval $0-60$ $\mu$s (see Fig. 3) are presented in Table 2. Observed difference in light pulse shapes allows to discriminate $\gamma $($\beta$) events from $\alpha$ particles. We applied for this purpose the optimal filter method proposed in [@Gatti] and developed in [@Faz98]. To obtain the numerical characteristic of YAG:Nd signal, so called shape indicator ($SI$), the following formula was applied for each pulse:
$SI=\sum f(t_k)\cdot P(t_k)/\sum f(t_k)$,
where the sum is over time channels $k,$ starting from the origin of pulse and up to 50 $\mu $s, $f(t_k)$ is the digitized amplitude (at the time $t_k$) of a given signal. The weight function $P(t)$ is defined as: $P(t)=\{{f}_\alpha (t)-{f}_\gamma (t)\}/\{f_\alpha
(t)+f_\gamma (t)\}$, where $f_\alpha (t)$ and $f_\gamma (t)$ are the reference pulse shapes for $\alpha$ particles and $\gamma$ quanta. Reasonable discrimination between $\alpha$ particles and $\gamma$ rays was achieved using this approach as one can see in Fig. 4 where the shape indicator distributions measured by the YAG:Nd scintillation crystal with $\alpha$ particles ($E_{\alpha}\approx
5.3$ MeV) and $\gamma$ quanta ($\approx 1.5-1.8$ MeV) are shown.
--------------------- ------------------ ------------------ ------------------ --
Type of irradiation
$\tau_1$ (A$_1$) $\tau_2$ (A$_2$) $\tau_3$ (A$_3$)
$\gamma$ rays 1.1 (10%) 4.1 (87%) 20 (3%)
$\alpha$ particles 1.0 (15%) 3.9 (83%) 18 (2%)
--------------------- ------------------ ------------------ ------------------ --
: Decay time of YAG:Nd scintillator for $\gamma$ quanta and $\alpha$ particles. The decay constants and intensities (in percentage of the total intensity) are denoted as $\tau_i$ and A$_i$, respectively.
### Radioactive contamination of the YAG:Nd crystal
To determine activity of $\alpha$ active nuclides from U/Th contamination in the YAG:Nd crystal, the pulse shape analysis was applied to data of 14.8 h low background measurement carried out at the earth surface laboratory in Kiev. The YAG:Nd crystal was viewed by the low-radioactive PMT (EMI D724KFLB) through plastic scintillator light-guide 10 cm in diameter and 4 cm long. The active light-guide reduces effect of $\gamma$ radiation from the PMT and provides suppression of cosmic rays induced background due to the pulse-shape discrimination of scintillation signals. Passive shield was consisted of steel (2 cm) and lead (10 cm). The energy resolution of the detector was 11% for 662 keV $\gamma$ rays.
To estimate contamination of the crystal by $\alpha$ active nuclides of U/Th families, the $\alpha$ events were selected from the data accumulated with the YAG:Nd crystal with the help of the pulse-shape discrimination technique. In the $\alpha$ spectrum there is no peculiarities which can be surely attributed to $\alpha$ active U/Th daughters. Analyzing these data we set the limit on total $\alpha$ activity of $^{235}$U, $^{238}$U, and $^{232}$Th daughters in the YAG:Nd crystal: $\leq 20$ mBq/kg.
DISCUSSION
==========
2$\beta$ decay of neodymium
---------------------------
Among neodymium isotopes there are three possible $2\beta$ candidates: $^{146}$Nd, $^{148}$Nd, and $^{150}$Nd. They are listed in Table 3 where the $Q_{2\beta}$ energies and the isotopic abundances are given. As it was mentioned in the Introduction, the most interesting of them is $^{150}$Nd due to the large value of the transition energy. Two neutrino $2\beta$ decay of $^{150}$Nd with the half-life $T_{1/2}=1.9^{+0.7}_{-0.4}\times10^{19}$ yr was observed in experiment [@Art95] by using a time projection chamber and samples of enriched $^{150}$Nd and natural neodymium. In [@Sil97] the double beta decay of $^{150}$Nd was studied with similar technique (only enriched $^{150}$Nd sample was used). The two-neutrino half-life $T_{1/2}=(6.8\pm0.8)\times10^{18}$ yr was measured. Recently, the observation of $^{150}$Nd $2\nu 2\beta$ decay to the second excited level of $^{150}$Sm ($0^+_1$ at 741 keV) with $T_{1/2}=1.4^{+0.5}_{-0.4}\times10^{20}$ yr was reported in [@Bar04], while only the limit $T_{1/2}>1.5\times10^{20}$ yr was determined previously at 90% C.L. [@Kli01]. The most stringent limit on the neutrinoless mode $T_{1/2}\geq1.2\times10^{21}$ yr was set in the experiment [@Sil97].
Transition Isotopic abundance (%) [@abundance] $Q_{2\beta}$ (keV) [@Aud03]
----------------------------------- ------------------------------------- -----------------------------
$^{146}$Nd$ \rightarrow ^{146}$Sm 17.2(0.3) 70.2(2.9)
$^{148}$Nd$ \rightarrow ^{148}$Sm 5.7(0.1) 1928.8(1.9)
$^{150}$Nd$ \rightarrow ^{150}$Sm 5.6(0.2) 3367.5(2.2)
: 2$\beta$ unstable neodymium isotopes
YAG:Nd scintillators provide a possibility to search for $0\nu2\beta$ decay of $^{150}$Nd by “active” source experimental method. The studied pulse-shape discrimination ability is an important advantage of YAG:Nd scintillators as a low counting rate detector. Background of scintillation detector caused by radioactive contamination by U/Th daughters can be effectively rejected due to pulse-shape analysis as it was demonstrated in experiments with CdWO$_4$ [@W-alpha; @Cd116] and CaWO$_4$ [@CARVEL; @CaWO] crystal scintillators. Obvious disadvantage of YAG:Nd detector is rather small concentration of Nd. In the studied crystal the mass concentration is only $\approx$0.03% of $^{150}$Nd. It should be noted, however, that YAG:Nd crystals can be produced with neodymium concentration up to 8 mol%, which corresponds to $\approx$0.1 mass% of $^{150}$Nd. In this connection we would like to refer the project CANDLES [@CANDLES] which intends to use a few tons of non-enriched CaF$_2$ crystals to search for $0\nu2\beta$ decay of $^{48}$Ca. The concentration of target $^{48}$Ca nuclei in CaF$_2$ detector is at the same level: $\approx$0.1 mass% of $^{48}$Ca.
The value of energy resolution FWHM=4% could allow $to$ $discover$ [@Zde04] the $0\nu2\beta$ decay of $^{150}$Nd with the half-life $T_{1/2}^{0\nu2\beta}\sim 10^{25}$ yr, which corresponds to the effective Majorana neutrino mass $m_{\nu}\sim 0.06-0.6$ eV (taking into account all existing calculations of nuclear matrix elements for $0\nu2\beta$ decay of $^{150}$Nd). As for the $sensitivity$ [@Zde04], an experiment involving $\approx$20 tons of non-enriched YAG:Nd crystals with $\approx$8 mol% of Nd could reach the half-life sensitivity $\approx3\times10^{26}$ yr (supposing zero background during ten years of measurements), which corresponds to the Majorana neutrino mass $0.01-0.1$ eV.
$\alpha$ decay of neodymium
---------------------------
Alpha decay is allowed energetically for five naturally occurring isotopes of neodymium. Their energies of $\alpha$ decay ($Q_{\alpha}$) and natural abundances are listed in Table 4. Calculation of half-life values based on models [@Buc91; @Poe83], and the value of the half-life of $^{145}$Nd calculated in [@Xu04] are also presented. Alpha decay was observed only for $^{144}$Nd. Average half-life $T_{1/2}=2.29\pm0.16\times10^{15}$ yr [@Son01] was derived on the basis of four measurements: $T_{1/2}=2.2\times10^{15}$ yr [@Por56], $T_{1/2}=1.9\times10^{15}$ yr [@Iso65], $T_{1/2}=2.4\pm0.3\times10^{15}$ yr [@Mac61] and $T_{1/2}=2.65\pm0.37\times10^{15}$ yr [@Alb87]. It is obvious that, due to low $Q_\alpha$ values and long expected $T_{1/2}$’s, there is no perspective to observe the $\alpha$ decay of $^{143}$Nd, $^{146}$Nd and $^{148}$Nd. As for $^{145}$Nd, only two $T_{1/2}$ limits are known: $T_{1/2}>6\times10^{16}$ yr [@Iso65] and $T_{1/2}>1\times10^{17}$ yr [@Kau66].
------------ -------------- ----------------- --------------------------------------- -------------------------------------- ------------------------ --------------------------------
Isotope Abund. $Q_\alpha$, MeV Exp. $T_{1/2}$, yr
[@abundance] [@Aud03] based on [@Buc91] based on [@Poe83] [@Xu04]
$^{143}$Nd 12.2% 0.520(7) $>1.1_{-1.0}^{+7.8}$$\times$$10^{80}$ $5.2_{-4.7}^{+51}$$\times$$10^{93}$ – –
$^{144}$Nd 23.8% 1.905(2) $1.9_{-0.1}^{+0.2}$$\times$$10^{15}$ $4.3_{-0.3}^{+0.3}$$\times$$10^{15}$ – $2.29\pm0.16$$\times$$10^{15}$
$^{145}$Nd 8.3% 1.578(2) $1.7_{-0.1}^{+0.2}$$\times$$10^{22}$ $3.9_{-0.4}^{+0.3}$$\times$$10^{23}$ $3.7$$\times$$10^{22}$ $>1$$\times$$10^{17}$
$^{146}$Nd 17.2% 1.182(2) $2.0_{-0.4}^{+0.4}$$\times$$10^{34}$ $3.9_{-0.7}^{+0.8}$$\times$$10^{34}$ – –
$^{148}$Nd 5.7% 0.599(3) $6.1_{-3.1}^{+6.7}$$\times$$10^{70}$ $1.1_{-0.6}^{+1.2}$$\times$$10^{71}$ – –
------------ -------------- ----------------- --------------------------------------- -------------------------------------- ------------------------ --------------------------------
: Theoretical calculations of half-lives for $\alpha$ decay of natural Nd isotopes. Uncertainties given for calculations with cluster model of Ref. [@Buc91] and semiempirical formula of Ref. [@Poe83] are related only with uncertainty in the $Q_\alpha$ values. For $^{143}$Nd, we take into account that change in a parity and spin additionally suppresses the $^{143}$Nd decay rate.
The indication on the $\alpha $ decay of natural $^{180}$W isotope with half-life $T_{1/2}=1.2_{-0.4}^{+0.8}$ (stat) $\pm 0.3$ (syst) $\times 10^{18}$ yr has been recently observed in the experiment [@W-alpha] with the help of the low background $^{116}
$CdWO$_4$ crystal scintillators. This result was confirmed in the measurements with CaWO$_4$ crystal as scintillator [@CaWO] and bolometer [@Coz04]. It should be also referred the excellent result on the detection of alpha decay of $^{209}$Bi with the half-life $T_{1/2}=(1.9\pm 0.2)\times 10^{19}$ y [@Mar03]. In both the experiments [@Coz04] and [@Mar03] the cryogenic technique, which uses simultaneous registration of heat and light signals, have been applied. This method provides perfect selection of $\alpha$ events on the background caused by $\gamma$ rays (electrons). It should be interesting to check YAG:Nd crystal as the cryogenic detector to search for alpha decay of $^{145}$Nd and to measure more accurately the $^{144}$Nd half-life.
It should be also mentioned a possibility to search for spin-dependent inelastic scattering of weakly interacting massive particles (WIMP) with excitation of low energy nuclear levels of $^{145}$Nd (the lowest one: $3/2^-$ 67 keV, E2 transition, and the second level: $5/2^-$ 72 keV, M1 transition) due to nonzero spin of this nucleus ($7/2^-$). Identification of such “mixed” (nuclear recoil plus $\gamma$ quanta) events could be possible due to the simultaneous registration of heat and light signals (such a techniques have been used in [@CRESST; @ROSEBUD]). A heat/light ratio for such events would differ from “pure” nuclear recoils or $\gamma$(electron) events.
CONCLUSIONS
===========
The scintillation properties of YAG:Nd crystals were studied. The energy resolution 9.3% (662 keV $^{137}$Cs $\gamma$ line) was obtained with the YAG:Nd crystal scintillator placed in liquid and viewed by two PMTs. Shapes of scintillation signals were investigated, and reasonable pulse-shape discrimination for $\gamma$ rays and $\alpha$ particles was achieved. The $\alpha/\beta$ ratio was measured with the YAG:Nd scintillator to be equal 0.33 at the energy of alpha particles 5.25 MeV. Radioactive contamination of the YAG:Nd crystal by $\alpha$ active nuclides from U/Th chains was estimated as $\leq20$ mBq/kg.
Three potentially $2\beta$ active neodymium isotopes can be studied with the help of YAG:Nd crystal. Due to good scintillation characteristics and pulse-shape discrimination ability the YAG:Nd scintillators seems to be encouraging material for the $2\beta$ experiment with $^{150}$Nd.
The YAG:Nd crystals could be used to search for $\alpha$ decay of natural neodymium isotopes, in particular, $^{145}$Nd. A strong signature of $\alpha$ decay can be assured via simultaneous registration of heat and light signals using cryogenic technique. However, ability of YAG:Nd crystal as cryogenic detector has to be elaborated in an experiment.
ACKNOWLEDGEMENT
===============
The authors would like to thank Dr. Yu.D. Glinka from the Institute of Surface Chemistry of the National Academy of Sciences of Ukraine for lending the YAG:Nd crystal used in the present study.
[99]{}
J.D. Vergados, Phys. Rep. 361 (2002) 1.
Yu.G. Zdesenko, Rev. Mod. Phys. 74 (2002) 663.
S.R. Elliot and P. Vogel, Ann. Rev. Nucl. Part. Sci. 52 (2002) 115.
V.I. Tretyak and Yu.G. Zdesenko, At. Data Nucl. Data Tables 61 (1995) 43; 80 (2002) 83.
M. Doi et al., Prog. Theor. Phys. Suppl. 83 (1985) 1.
J. Suhonen and O. Civitarese, Phys. Rep. 300 (1998) 123.
M. Doi et al., Prog. Th. Phys. 69 (1983) 602;\
A. Staudt et al., Europhys. Lett. 13 (1990) 31;\
T. Tomoda, Rep. Prog. Phys. 54 (1991) 53;\
M. Hirsh et al., Z. Phys. A 347 (1994) 151;\
J.G. Hirsh et al., Nucl. Phys. A 582 (1995) 124;\
F. Simkovic et al., Phys. Rev. C 60 (1999) 055502.
see, for instance: Yu.G. Zdesenko et al., “CARVEL: an experiment with $^{48}$CaWO$_4$ crystal scintillators for the double $\beta $ decay study of $^{48}$Ca”, submitted to Astropart. Phys.
M. Moszynski et al., Nucl. Instr. Meth. A 345 (1994) 461;\
T. Ludziejewski et al., Nucl. Instr. Meth. A 398 (1997) 287;\
T. Bhattacharjee et al., Nucl. Instr. Meth. A 484 (2002) 364.
P. Antonini et al., Nucl. Instr. Meth. A 460 (2001) 469.
F.A. Danevich et al., Phys. Rev. C 67 (2003) 014310.
F.A. Danevich, nucl-ex/0409014; submitted to Nucl. Instr. Meth. A.
T. Fazzini et al., Nucl. Instr. Meth. A 410 (1998) 213.
E. Gatti, F. De Martini, Nuclear Electronics 2, IAEA, Vienna, 1962, p. 265.
K.J.R. Rosman, P.D.P. Taylor, Pure and Appl. Chem. 70 (1998) 217.
G. Audi, A.H. Wapstra, C. Thibault, Nucl. Phys. A 729 (2003) 337.
V. Artemiev et al., Phys. Lett. B 345 (1995) 564.
A. De Silva et al., Phys. Rev. C 56 (1997) 2451.
A.S. Barabash et al., JETP Letters 79 (2004) 12.
A.A. Klimenko et al., Part. Nucl. Lett. 5 (2001) 80.
F.A. Danevich et al., Phys. Rev. C 62 (2000) 045501; 68 (2003) 035501.
Yu.G. Zdesenko et al., Nucl. Instr. Meth. A, in press.
I. Ogawa et al., Proc. 5-th Int. Workshop on Neutrino Oscillation and their Origin, NOON 2004, 10-15 Feb. 2004, Tokyo, Japan, to be published in Proc.
see discusion about discovery potential and sensitivity of $2\beta$ decay experiments in: Yu.G. Zdesenko et al., J. Phys. G.: Nucl. Part. Phys. 30 (2004) 971.
B. Buck, A.C. Merchant, S.M. Perez, J. Phys. G 17 (1991) 1223.
D.N. Poenaru, M. Ivascu, J. Physique 44 (1983) 791.
C. Xu, Z. Ren, Phys. Rev. C 69 (2004) 024614.
A.A. Sonzogni, Nucl. Data Sheets 93 (2001) 599.
W. Porschen, W. Riezler, Z. Naturforsch. 11a (1956) 143.
A. Isola, M. Nurmia, Z. Naturforsch. 20a (1965) 541.
R.D. Macfarlane, T.P. Kohman, Phys. Rev. 121 (1961) 1758.
B. Al-Bataina, J. Janecke, Radiochim. Acta 42 (1987) 159.
G. Kauw, Forsch. Laudes Nord.-Westfalen N 1640 (1966).
C. Cozzini et al., nucl-ex/0408006.
P. de Marcillac et al., Nature 422 (2003) 876.
M. Bravin et al., Astropart. Phys. 12 (1999) 107.
S. Cebrian et al., Phys. Lett. B 563 (2003) 48.
[^1]: Corresponding author. [*E-mail address:*]{} [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: |
We propose a new *CPT*-even and Lorentz-violating nonminimal coupling between fermions and Abelian gauge fields involving the *CPT*-even tensor $\left( K_{F}\right) _{\mu\nu\alpha\beta}$ of the standard model extension. We thus investigate its effects on the cross section of the electron-positron scattering by analyzing the process $e^{+}+e^{-}%
\rightarrow\mu^{+}+\mu^{-}$. Such a study was performed for the parity-odd and parity-even nonbirefringent components of the Lorentz-violating $\left(
K_{F}\right) _{\mu\nu\alpha\beta}$ tensor. Finally, by using experimental data available in the literature, we have imposed upper bounds as tight as $%
10^{-12}(\mbox{eV})^{-1}$ on the magnitude of the *CPT*-even and Lorentz-violating parameters while nonminimally coupled.
author:
- 'R. Casana$^{a}$, M. M. Ferreira Jr$^{a}$, R.V. Maluf$^{b}$, and F.E.P. dos Santos$^{a}$[^1]'
title: 'Effects of a *CPT*-even and Lorentz-violating nonminimal coupling on the electron-positron scattering'
---
Introduction
============
The standard model extension (SME) is a large theoretical framework that includes terms of Lorentz and *CPT* violation in the structure of the usual standard model [@Colladay]. This model was proposed after the verification about the possibility of having spontaneous violation of Lorentz symmetry in the context of string theories [@Samuel]. The Lorentz-violating (LV) terms are generated as vacuum expectation values of tensor quantities, keeping the coordinate invariance of the extended theory [@Lehnert]. This model has been scrutinized in many respects in the latest years, with studies embracing the fermion and gauge sectors, and gravitation extension [@Gravity]. The fermion sector [@Fermion] was much examined, mainly in connection with *CPT*-violating tests to impose some upper bounds on the magnitude of the LV terms [@CPT], dealing with other interesting aspects as well [@Fermion2]. The Abelian gauge sector of the SME is composed of a *CPT*-odd [@Jackiw] and a *CPT*-even sector, both intensively investigated in the latest years [@Adam; @Cherenkov1; @Photons1; @KM1; @Risse; @Cherenkov2; @Kostelec].
Besides the investigations undertaken into the structure of the SME, some other works were proposed to examine Lorentz-violating developments out of this broad framework. Some of them involve nonminimal coupling terms that modify the vertex interaction between fermions and photons. *CPT*-odd nonminimal couplings as $gv_{\mu}\tilde{F}^{\mu\nu}$and $g\gamma_{5}b_{\mu}%
\tilde{F}^{\mu\nu}$ were considered some time ago in the context of the Dirac equation, with interesting consequences in the nonrelativistic limit, involving topological phases [@NM1; @NM2], corrections on the hydrogen spectrum [@NM3]. Such nonminimal coupling has been reassessed in connection with its implications on the Aharonov-Bohm-Casher problem [NMabc]{}, the Bhabha cross section [@NMmaluf], and other respects [NMothers]{}. Recently, other types of nonminimal coupling, defined in the context of the Dirac equation, have been proposed for investigating the generation of topological and geometrical phases [@NMbakke].
Theoretical studies about cross section evaluation in the presence of Lorentz-violating terms were accomplished by some authors [@Colladay2], searching to elucidate the route for evaluating the cross section for a general scattering. Very recently, some authors performed a study on the Bhabha scattering [@NMmaluf], determining the effects induced by the nonminimal *CPT*-odd coupling on the Bhabha cross section. The results were compared with some available data concerning this scattering [Derrick]{} and used to impose the upper bound $\left\vert gv_{\mu}\right\vert
\leq10^{-12}\left( \mbox{eV}\right) ^{-1}.$
In this work, we reassess a well-known quantum electrodynamics process, the $%
e^{+}+e^{-}\,\rightarrow\mu^{+}+\,\mu^{-}$ scattering, in the presence of a new Lorentz-violating *CPT*-even nonminimal coupling involving the fermion and gauge sectors. First, we calculate the scattering amplitude, considering new Feynman diagrams due to the emergence of a new vertex in the theory. In order to evaluate the total cross-section, we first calculate the unpolarized squared amplitude, using the Casimir trick. We specialize our evaluations for the parity-odd and parity-even subsectors of the *CPT*-even gauge sector. At the end, following the approach of Refs. [NMmaluf]{} and [@Derrick], we compare the cross section results with the experimental data, finding an upper limit for the magnitude for the new nonminimal coupling as tight as $\left\vert \lambda\left( K_{F}\right)
\right\vert \leq10^{-12}\left( \mbox{eV}\right) ^{-1}.$
The theoretical model
=====================
We are interested in analyzing some aspects of a modified quantum electrodynamics, whose fermion sector is governed by the generalized Dirac equation, $$\left(i\gamma^{\mu}D_{\mu}-m\right)\Psi=0, \label{Dirac1}$$ in which the usual covariant derivative is supplemented by a nonminimal *CPT*-even coupling term, that is, $$D_{\mu}=\partial_{\mu}+ieA_{\mu}+\frac{\lambda}{2}\left( K_{F}\right)
_{\mu\nu\alpha\beta}\gamma^{\nu}F^{\alpha\beta}, \label{covader}$$ where $\left( K_{F}\right) _{\mu\nu\alpha\beta}$ is the tensor that embraces the 19 LV terms belonging to the *CPT*-even gauge sector of the SME. This tensor possesses the same symmetries of the Riemann tensor: $\left(
K_{F}\right) _{\alpha\nu\rho\varphi}=-\left( K_{F}\right) _{\nu\alpha
\rho\varphi},~\left( K_{F}\right) _{\alpha\nu\rho\varphi}=-\left(
K_{F}\right) _{\alpha\nu\varphi\rho},~\left( K_{F}\right) _{\alpha\nu
\rho\varphi}=\left( K_{F}\right) _{\rho\varphi\alpha\nu}$ and a double null trace, $\left( K_{F}\right) ^{\alpha\beta}{}_{\alpha\beta}=0,$ implying 19 components. Using the symmetries of the tensor $\left( K_{F}\right) _{\mu
\nu\alpha\beta}$ in the Dirac (\[Dirac1\]) equation, one obtains $$\left[ i\gamma^{\mu}\partial_{\mu}-e\gamma^{\mu}A_{\mu}+\frac{\lambda}{2}%
\left( K_{F}\right) _{\mu\nu\alpha\beta}\sigma^{\mu\nu}F^{\alpha\beta }-m%
\right] \Psi=0, \label{DiracM1}$$ with $$\sigma^{\mu\nu}=\frac{i}{2}(\gamma^{\mu}\gamma^{\nu}-\gamma^{\nu}\gamma^{\mu
})=\frac{i}{2}[\gamma^{\mu},\gamma^{\nu}], \label{OP1}$$ whose components, $\sigma^{0i}$ and $\sigma^{ij},$ are $$\sigma^{0i}=i\left(
\begin{array}{cc}
0 & \sigma^{i} \\
\sigma^{i} & 0%
\end{array}
\right) ,\text{ \ \ }\sigma^{ij}=-\left(
\begin{array}{cc}
\epsilon_{ijk}\sigma^{k} & 0 \\
0 & \epsilon_{ijk}\sigma^{k}%
\end{array}
\right) .$$ This new coupling, represented by $(\lambda K_{F})_{\mu\nu\alpha\beta}$, has mass dimension $[\lambda K_{F}]=-1$, which leads to a nonrenormalizable theory at power counting. This respect does not pose a problem for this investigation, once we are interested in analyzing the tree-level scattering process.
We now present the Lagrangian of the modified QED, $$\mathcal{L}_{\text{mod}QED}=\mathcal{L}_{QED}+\mathcal{L}_{I}^{new}
\label{L1}$$where $\mathcal{L}_{QED}$ is the usual Lagrangian density of QED in the Lorenz gauge, $$\mathrm{{\mathcal{L}}}_{QED}=\bar{\psi}(i{\rlap{\hbox{$\mskip 1 mu /$}}}%
\partial -e{\rlap{\hbox{$\mskip 1 mu /$}}}A-m)\psi -\frac{1}{4}F_{\mu \nu
}F^{\mu \nu }-\frac{1}{2\xi }(\partial _{\mu }A^{\mu })^{2},$$and $\mathrm{{\mathcal{L}}}_{I}^{new}~$represents the new interaction produced by the nonminimal coupling, to be regarded $${{\mathcal{L}}}_{I}^{new}=\frac{\lambda }{2}(K_{F})_{\mu \nu \alpha \beta }%
\bar{\psi}\sigma ^{\mu \nu }\psi F^{\alpha \beta }.$$In the next steps we will consider the Feynman gauge, $\xi =1$. The theory represented by Lagrangian (\[L1\]) has, besides the usual vertex, $\bullet
\rightarrow -ie\gamma ^{\mu }$, an additional LV vertex, represented as $$\times \rightarrow \lambda V_{\beta }=\lambda (K_{F})_{\mu \nu \alpha \beta
}\sigma ^{\mu \nu }q^{\alpha }, \label{V1}$$in the momentum space.
We are interested in analyzing how the electron-positron scattering, $\
e^{+}+e^{-}\rightarrow\mu^{+}+\mu^{-}$, is altered by this new vertex. This process may be depicted by the following tree-level Feynman diagrams:
{width="8cm"}
The tensor $K_{F}$ is composed of birefringent and nonbirefringent components. Without loss of generality, we restrain our investigation to the nonbirefringent sector [@Obs1], represented by nine coefficients and parametrized by a symmetric and traceless rank-2 tensor defined by the contraction $$\kappa ^{\mu \nu }=\left( K_{F}\right) _{\alpha }{}^{\mu \alpha \nu },
\label{k2}$$which fulfills $$\left( K_{F}\right) ^{\lambda \nu \delta \rho }=\frac{1}{2}\left[ g^{\lambda
\delta }\kappa ^{\nu \rho }-g^{\nu \delta }\kappa ^{\lambda \rho }+g^{\nu
\rho }\kappa ^{\lambda \delta }-g^{\lambda \rho }\kappa ^{\nu \delta }\right]
. \label{Alt}$$Hence, the interaction (\[V1\]) is rewritten as $${{\mathcal{L}}}_{I}^{new}=\lambda \kappa _{\nu \beta }\bar{\psi}\sigma ^{\mu
\nu }\psi F_{\mu }{}^{\beta }, \label{LI2}$$which implies the following vertex: $$\lambda V^{\mu }=\lambda q^{\beta }\left( \kappa ^{\nu \mu }\sigma _{\beta
\nu }-\kappa _{\nu \beta }\sigma ^{\mu \nu }\right) . \label{V2}$$
The components of the tensor can be classified by their parity properties: $%
\kappa_{00},\kappa_{ij}$ are parity even, while $\kappa_{0i}$ is parity odd.
The cross section evaluation
============================
In this section, we evaluate the differential and total cross section for the process, $$e^{+}+e^{-}\rightarrow \mu ^{+}+\mu ^{-},$$where the particles are labeled with momentum and spin variables as $%
e^{+}\left( p_{1};s_{1}\right) $, $e^{-}\left( p_{2};s_{2}\right) ,~\mu
^{+}\left( p_{1}^{\prime };s_{1}^{\prime }\right) $, and $\mu ^{-}\left(
p_{2}^{\prime };s_{2}^{\prime }\right) $. We work in the center of mass frame, in which it holds $p_{1}=\left( E,\mathbf{p}\right) $, $p_{2}=\left(
E,-\mathbf{p}\right) $, $p_{1}^{\prime }=\left( E,\mathbf{p}^{\prime
}\right) $, and $p_{2}^{\prime }=\left( E,-\mathbf{p}^{\prime }\right) $, with $p_{1},p_{2}$, and $p_{1}^{\prime },p_{2}^{\prime }$ being the momenta of the incoming and outgoing particles, respectively. Transfer momentum $%
\left( q=p_{1}+p_{2}\right) $ is $q^{\beta }=\left( \sqrt{s},0\right) ,$ where $\sqrt{s}$ is the energy in the center of mass. In this frame, it holds $\left\vert \mathbf{p}^{\prime }\right\vert ^{2}=\left\vert \mathbf{p}%
\right\vert ^{2}-m_{\mu }^{2}+m_{e}^{2},$ and $$\rlap{$/$}p_{2}=\gamma ^{0}\rlap{$/$}p_{1}\gamma ^{0}=\gamma ^{0}\rlap{$/$}%
p\gamma ^{0},\text{ \ }\rlap{$/$}p_{2}^{\prime }=\gamma ^{0}\rlap{$/$}%
p_{1}^{\prime }\gamma ^{0}=\gamma ^{0}\rlap{$/$}p^{\prime 0}, \label{pp}$$with ${m_{\mu }},m_{e}$ being the masses of the muon and the electron, respectively. The vertex components are $V^{0}=0,$ and $$V^{i}=\sqrt{s}\left( \kappa _{00}\sigma ^{0i}-\kappa _{ij}\sigma
^{0j}-\kappa _{0j}\sigma ^{ij}\right) . \label{V3}$$
Note that it holds $\kappa^{00}=\kappa^{ii}=\frac{3}{2}\kappa_{\text{tr}}, $ $\kappa^{ij}=-\left( \kappa_{e-}\right) ^{ij}+\frac{1}{2}\kappa_{\text{tr}%
}\delta^{ij}$, $\kappa^{0i}=-\kappa^{i},$ where $\kappa_{\text{tr}}$, and $%
\left( \kappa_{e-}\right) _{ij}$ correspond to the isotropic and anisotropic parity-even components of the *CPT*-even sector, respectively, while $%
\kappa^{i}$ represents the parity-odd components in accordance with Ref. [@KM1]. These vertex components can be read as $$V^{i}=V_{+I}^{i}+V_{+A}^{i}+V_{-}^{i}, \label{V5}$$ where $V_{+I}^{i}=\sqrt{s}\kappa_{00}\sigma^{0i}$ is the part associated with the parity-even isotropic coefficient, $V_{+A}^{i}=-\sqrt{s}%
\kappa_{ij}\sigma^{0j}$ is related to the anisotropic parity-even component, and $V_{-}^{i}=-\sqrt{s}\kappa_{0j}\sigma^{ij}=\sqrt{s}\kappa_{j}\sigma^{ij}$ is the contribution stemming from the parity-odd components.
In this scenario, the differential cross section (in natural units) is given by $$\frac{d\sigma }{d\Omega }=\frac{\left\vert \mathbf{p}^{\prime }\right\vert }{%
\left( 8\pi \right) ^{2}s\left\vert \mathbf{p}\right\vert }\left\vert {{%
\mathcal{M}}}\right\vert ^{2}. \label{CS1}$$The scattering amplitude is read off from the Feynman diagrams, $${{\mathcal{M}}}=\underset{a,b}{{\sum }}[\bar{v}^{s_{2}}(p_{2})\Gamma
_{\left( a\right) }^{\mu }u^{s_{1}}(p_{1})]\frac{1}{q^{2}}[\bar{u}%
^{s_{1}^{\prime }}(p_{1}^{\prime })\Gamma _{\left( b\right) \mu }v
^{s_{2}^{\prime }}(p_{2}^{\prime })], \label{amplitudes}$$where $a,b=0,1$ and $\Gamma _{\left( a\right) }^{\mu }$ defined by $$\Gamma _{(0)}^{\mu }=-ie\gamma ^{\mu },\text{ }\Gamma _{(1)}^{\mu }=\lambda
V^{\mu }, \label{Vertices}$$stands for the usual and new vertices. Here, $u^{s_{1}}(p_{1})$, $\bar{v}%
^{s_{2}}(p_{2})$ are the spinors for the electron and the positron, while $%
\bar{u}^{s_{1}^{\prime }}(p_{1}^{\prime })$, $v ^{s_{2}^{\prime }}\left(
p_{2}^{\prime }\right) $ represent the muon and antimuon spinors. For evaluating the unpolarized cross section , the relevant quantity is $\langle
|{{\mathcal{M}}}|^{2}\rangle $, defined as $\left\vert {{\mathcal{M}}}%
\right\vert ^{2}={\displaystyle\sum }{{\mathcal{M}}{\mathcal{M}}}^{\ast }$, where the sum is over the spin indices, $s_{1},s_{2},s_{1}^{\prime
},s_{2}^{\prime }$. This squared amplitude is carried out by means the Casimir’s trick, based on the use of spinor completeness relations and the trace properties of $\gamma $ matrices. Knowing that $${{\mathcal{M}}}^{\ast }=\sum [\bar{u}^{s_{1}}(p_{1})\bar{\Gamma}_{\left(
a\right) }^{\mu }v^{s_{2}}(p_{2})]\frac{1}{q^{2}}[\bar{v}^{s_{2}^{\prime
}}(p_{2}^{\prime })\bar{\Gamma}_{\left( b\right) \mu }u^{s_{1}^{\prime
}}(p_{1}^{\prime })], \label{amplitudeC}$$the squared amplitude is written as $$\begin{aligned}
\langle |{{\mathcal{M}}}|^{2}\rangle & =\frac{1}{4q^{4}}\sum \bar{v}%
^{s_{2}}(p_{2})\Gamma _{(a)}^{\mu }u^{s_{1}}(p_{1})\bar{u}^{s_{1}}(p_{1})%
\bar{\Gamma}_{(b)}^{\rho }v^{s_{2}}(p_{2}) \notag \\
& \times \bar{u}^{s_{1}^{\prime }}(p_{1}^{\prime })\Gamma _{(c)\mu
}v^{s_{2}^{\prime }}(p_{2}^{\prime })\bar{v}^{s_{2}^{\prime }}(p_{2}^{\prime
})\bar{\Gamma}_{(d)\rho }u^{s_{1}^{\prime }}(p_{1}^{\prime }).
\label{Msquared1}\end{aligned}$$where $\bar{\Gamma}_{(i)}^{\mu }=\gamma ^{0}\Gamma _{(i)}^{\mu \dag }\gamma
^{0},$ and the sum is over the spin indices and over $a,b,c,d$. Using the relation, $$\begin{aligned}
& \bar{v}^{s_{2}}(p_{2})\Gamma _{(a)}^{\mu }u^{s_{1}}(p_{1})\bar{u}%
^{s_{1}}(p_{1})\bar{\Gamma}_{(b)}^{\nu }v^{s_{2}}(p_{2}) \notag \\
& =\text{tr}(\Gamma _{(a)}^{\mu }u^{s_{1}}(p_{1})\bar{u}^{s_{1}}(p_{1})\bar{%
\Gamma}_{(b)}^{\nu }v^{s_{2}}(p_{2})\bar{v}^{s_{2}}(p_{2})),\end{aligned}$$the spin sum yields $$\langle |{{\mathcal{M}}}|^{2}\rangle =\frac{1}{4q^{4}}L_{T}^{\mu \nu }\left(
M_{T}\right) _{\mu \nu }, \label{amplitudespins}$$$$\begin{aligned}
L_{T}^{\mu \nu }& =L_{(00)}^{\mu \nu }+L_{(01)}^{\mu \nu }+L_{(10)}^{\mu \nu
}+L_{(11)}^{\mu \nu }, \\
M_{T}^{\mu \nu }& =M_{(00)}^{\mu \nu }+M_{(01)}^{\mu \nu }+M_{(10)}^{\mu \nu
}+M_{(11)}^{\mu \nu },\end{aligned}$$with $$\begin{aligned}
& \left. {L_{(ab)}^{\mu \nu }}{=}\text{{tr}}{[\Gamma _{(a)}^{\mu }(\rlap{$/$}%
p_{1}+m_{e})\bar{\Gamma}_{(b)}^{\nu }(\rlap{$/$}p_{2}-m_{e})],}\right.
\label{L1ab} \\
& {\;}\left. {M_{(ab)\mu \nu }=}\text{{tr}}{[\Gamma _{(a)\mu }(\rlap{$/$}%
p_{2}^{\prime }-m_{\mu })\bar{\Gamma}_{(b)\nu }(\rlap{$/$}p_{1}^{\prime
}+m_{\mu })].}\right. \label{M1ab}\end{aligned}$$Remember that the Latin indices inside parentheses $(a,b)$ can assume only two values, $0$ or $1$, corresponding to the usual and new nonminimal vertex, properly defined in Eqs. (\[V3\]) and (\[Vertices\]).
Next, in order to facilitate our evaluations and better discuss our results, we proceed to separate the contributions coming from the parity-odd and parity-even coefficients.
Parity-odd contribution
-----------------------
To calculate parity-odd contributions to the cross section, we restrict the vertex (\[V3\]) to $$V_{-}^{i}=\sqrt{s}\sigma ^{ij}\kappa _{j}, \label{V6}$$where $\kappa ^{i}=\left( K_{F}\right) ^{0jij}$. Using the trace technique, and using identity (\[pp\]), we show that$$L_{(01)}^{ij}=L_{(10)}^{ij}=M_{(01)}^{ij}=M_{(10)}^{ij}=0. \label{L011}$$The nonnull terms of the tensors (\[L1ab\],\[M1ab\]) are $$\begin{aligned}
& \left. {L_{(00)}^{ij}}{=e}^{2}\text{{tr}}{[\gamma ^{i}\rlap{$/$}%
p_{1}\gamma ^{j}\rlap{$/$}p_{2}-m_{e}^{2}\gamma ^{i}\gamma ^{j}],}\right. \\
& \left. {L_{(11)}^{ij}=\lambda }^{2}\text{{tr}}{[V^{i}\rlap{$/$}p_{1}V^{j}%
\rlap{$/$}p_{2}-m_{e}^{2}V^{i}V^{j}],}\right. \\
& \left. {M_{(00)}^{ij}}{={e}^{2}}\text{{tr}}{[\gamma ^{i}\rlap{$/$}%
p_{1}^{\prime }\gamma ^{j}\rlap{$/$}p_{2}^{\prime }-m_{\mu }^{2}\gamma
^{i}\gamma ^{j}],}\right. \\
& \left. {M_{(11)}^{ij}={\lambda }^{2}}\text{{tr}}{[V^{i}\rlap{$/$}%
p_{1}^{\prime }V^{j}\rlap{$/$}p_{2}^{\prime }-m_{\mu }^{2}V^{i}V^{j}]}%
,\right.\end{aligned}$$while $L_{(ab)}^{0\mu }=L_{(ab)}^{\mu 0}=M_{(ab)}^{0\mu }=M_{(ab)}^{\mu 0}=0$. These latter terms are explicitly carried out: $$\begin{aligned}
L_{(00)}^{ij}& ={e}^{2}\left( 2s\delta ^{ij}-8p^{i}p^{j}\right) , \notag \\%
[-0.3cm]
& \\
M_{(00)}^{ij}& ={e}^{2}\left( 2s\delta ^{ij}-8p^{\prime i}p^{\prime
j}\right) , \notag\end{aligned}$$$$\begin{aligned}
L_{(11)}^{ij}& =8{\lambda }^{2}s\varepsilon ^{ikm}\varepsilon
^{jln}p^{n}p^{m}\kappa ^{k}\kappa ^{l},\text{ } \\
M_{(11)}^{ij}& =8{\lambda }^{2}s\varepsilon ^{ikm}\varepsilon
^{jln}p^{\prime n}p^{\prime m}\kappa ^{k}\kappa ^{l}.\end{aligned}$$The squared amplitude is $$\begin{aligned}
\langle |{{\mathcal{M}}}|^{2}\rangle & =\frac{1}{4s^{2}}\left[ L_{\left(
00\right) }^{ij}M_{\left( 00\right) }^{ij}+L_{\left( 11\right)
}^{ij}M_{\left( 00\right) }^{ij}\right. \notag \\
& \left. +L_{\left( 00\right) }^{ij}M_{\left( 11\right) }^{ij}+L_{\left(
11\right) }^{ij}M_{\left( 11\right) }^{ij}\right] . \label{Msquared2}\end{aligned}$$
The differential cross section is obtained replacing these results in Eqs. ( \[Msquared2\]) and (\[CS1\]). The total cross section is obtained by integration, $$\sigma =\frac{\left\vert \mathbf{p}^{\prime }\right\vert }{\left( 8\pi
\right) ^{2}s\left\vert \mathbf{p}\right\vert }\int \left\langle \left\vert
\mathcal{M}\right\vert ^{2}\right\rangle d\Omega .$$Taking the background as fixed, we integrate only on the angular variables of the scattered particles, that is,$$\begin{aligned}
\int \left\langle \left\vert {{\mathcal{M}}}\right\vert ^{2}\right\rangle
d\Omega & =\frac{1}{4s^{2}}\left[ L_{00}^{ij}\int M_{00}^{ij}d\Omega
+L_{11}^{ij}\int M_{00}^{ij}d\Omega \right. \notag \\
& +\left. L_{00}^{ij}\int M_{11}^{ij}d\Omega +L_{11}^{ij}\int
M_{11}^{ij}d\Omega \right] .\end{aligned}$$These integrals provide$$\begin{aligned}
\int M_{\left( 00\right) }^{ij}d\Omega & =\frac{16{{e}^{2}}}{3}\left(
s+2m_{\mu }^{2}\right) \pi \delta ^{ij}, \\
\int M_{\left( 11\right) }^{ij}d\Omega & =\frac{8{{\lambda }^{2}}}{3}\pi
s\left( s-4m_{\mu }^{2}\right) \left( \delta ^{ij}\boldsymbol{\kappa }%
^{2}-\kappa ^{i}\kappa ^{j}\right) ,\end{aligned}$$where the integral $\int p^{\prime i}p^{\prime j}d\Omega =\frac{1}{3}\left(
s-4m_{\mu }^{2}\right) \pi \delta ^{ij}$ was used. In the ultrarelativistic limit, we take $m_{e}=m_{\mu }=0$. The resulting cross section (at second order) is$$\sigma =\sigma _{QED}\left[ 1+\frac{1}{4e^{2}}\lambda ^{2}\left( 3s%
\boldsymbol{\kappa }^{2}-4\left( \mathbf{p}\cdot \boldsymbol{\kappa }\right)
^{2}\right) \right] .$$
The results can be presented in two ways, concerning the beam orientation in relation to the background vector, $\kappa^{i}.$ For the case where the beam is perpendicular to the background, $\boldsymbol{\kappa}\cdot\mathbf{p}=0$, we achieve $$\sigma=\sigma_{QED}\left( 1+\frac{3s}{4e^{2}}\lambda^{2}|\boldsymbol{\kappa }%
|^{2}\right) , \label{CSortog}$$ while for the case where the beam is parallel to the background, $%
\boldsymbol{\kappa }\cdot\mathbf{p}=\left\vert \boldsymbol{\kappa}%
\right\vert \sqrt{s}/2, $ the total cross section is $$\sigma=\sigma_{QED}\left( 1+\frac{s}{2e^{2}}\lambda^{2}|\boldsymbol{\kappa }%
|^{2}\right) . \label{CSlong}$$
Experimental data from Ref. [@Derrick] for the $e^{+}+e^{-}\,\rightarrow
\mu^{+}+\,\mu^{-}$ scattering yields $$\frac{\sigma-\sigma_{QED}}{\sigma_{QED}}=\pm\frac{2s}{\Lambda_{\pm}^{2}},
\label{Experiment}$$ where $\sqrt{s}=29\; \mbox{GeV}$ and $\Lambda_{+}=170\; \mbox{GeV}$ with $%
95\%$ confidence level. Comparing (\[CSortog\]) and (\[CSlong\]) with (\[Experiment\]), we obtain the following upper bound: $$|\lambda\boldsymbol{\kappa}|<3\times10^{-12}\,(\mbox{eV})^{-1}.
\label{Bound1}$$
Parity-even contribution
------------------------
We begin considering the parity-even and isotropic contribution, whose associated vertex is $V_{+I}^{i}=\sqrt{s}\kappa _{00}\sigma ^{0i}.$ In this case, the elements of the tensors (\[L1ab\],\[M1ab\]) are $$\begin{aligned}
L_{(00)}^{\mu \nu }& =e^{2}\text{tr}[\gamma ^{\mu }\rlap{$/$}p_{1}\gamma
^{\nu }\rlap{$/$}p_{2}-m_{e}^{2}\gamma ^{\mu }\gamma ^{\nu }], \\
L_{(01)}^{\mu \nu }& =ie\lambda m_{e}\text{tr}[\gamma ^{\mu }\rlap{$/$}%
p_{1}V_{+I}^{\nu }-\gamma ^{\mu }V_{+I}^{\nu }\rlap{$/$}p_{2}], \\
L_{(10)}^{\mu \nu }& =-ie\lambda m_{e}\text{tr}[V_{+I}^{\mu }\rlap{$/$}%
p_{1}\gamma ^{\nu }-V_{+I}^{\mu }\gamma ^{\nu }\rlap{$/$}p_{2}], \\
L_{(11)}^{\mu \nu }& =\lambda ^{2}\text{tr}[V_{+I}^{\mu }\rlap{$/$}%
p_{1}V_{+I}^{\nu }\rlap{$/$}p_{2}-m_{e}^{2}V_{+I}^{\mu }V_{+I}^{\nu }],\end{aligned}$$The components of tensor $M_{(ab)}^{\mu \nu }$ are written in the same way, changing $p_{1},p_{2},m_{e}$ by $p_{1}^{\prime },p_{2}^{\prime },m_{\mu }.$ In this case, $$L_{\left( ab\right) }^{0\mu }=L_{\left( ab\right) }^{\mu 0}=M_{\left(
ab\right) }^{0\mu }=M_{\left( ab\right) }^{\mu 0}=0,$$remaining as nonnull only the components $L_{\left( ab\right) }^{ij}$, $%
M_{\left( ab\right) }^{ij}$, given as $$\begin{aligned}
L_{(01)}^{ij}& =L_{(01)}^{ij}=4e\lambda \kappa _{00}sm_{e}\delta ^{ij}, \\
M_{(01)}^{ij}& =M_{(01)}^{ij}=4e\lambda \kappa _{00}sm_{\mu }\delta ^{ij}, \\
L_{(11)}^{ij}& =8s\lambda ^{2}\left( \kappa _{00}\right) ^{2}\left(
m_{e}^{2}\delta ^{ij}+p^{i}p^{j}\right) , \\
M_{(11)}^{ij}& =8s\lambda ^{2}\left( \kappa _{00}\right) ^{2}\left( m_{\mu
}^{2}\delta ^{ij}+p^{\prime i}p^{\prime j}\right) .\end{aligned}$$
The squared amplitude now is $$\begin{aligned}
\langle| {{\mathcal{M}}}|^{2}\rangle & =\frac{1}{4s^{2}}\left(
L_{(00)}^{ij}+2L_{(01)}^{ij}+L_{(11)}^{ij}\right) \ \notag \\
& \times\left( M_{(00)}^{ij}+2M_{(01)}^{ij}+M_{(11)}^{ij}\right) .\end{aligned}$$
Proceeding with the integration evaluations, and taking the ultrarelativistic limit $\left( m_{e}=m_{\mu }=0\right) ,$ the total cross section (at second order) is$$\sigma =\sigma _{QED}\left( 1+\frac{s}{e^{2}}\left\vert \lambda \kappa
_{00}\right\vert ^{2}\right) .$$By using the same conditions as in Eq. (\[Experiment\]), we achieve $$|\lambda \kappa _{00}|<2.5\times 10^{-12}(\mbox{eV})^{-1}.$$
We continue regarding the anisotropic parity-even contribution, whose vertex is $V_{-A}^{i}=-\sqrt{s}\kappa ^{ij}\sigma ^{0j}.$ In this case, for turning feasible the evaluations, we consider the ultrarelativistic limit $\left( {%
m_{e}}={m_{\mu }}=0\right) $ well in the beginning. The operators (\[L1ab\],\[M1ab\]) are rewritten as $$\begin{aligned}
L_{(00)}^{\mu \nu }& \approx e^{2}\text{tr}[\gamma ^{\mu }\rlap{$/$}%
p_{1}\gamma ^{\nu }\rlap{$/$}p_{2}], \\
L_{(11)}^{\mu \nu }& \approx \lambda ^{2}\text{tr}[V_{-A}^{\mu }\rlap{$/$}%
p_{1}V_{-A}^{\nu }\rlap{$/$}p_{2}],\end{aligned}$$with components of the tensor $M_{(ab)}^{\mu \nu }$ read similarly by changing $p_{1},p_{2},m$ by $p_{1}^{\prime },p_{2}^{\prime },M.$ Some evaluations lead to $$\begin{aligned}
L_{\left( ab\right) }^{0\mu }& =L_{\left( ab\right) }^{\mu 0}=M_{\left(
ab\right) }^{0\mu }=M_{\left( ab\right) }^{\mu 0}=0, \\
L_{(11)}^{ij}& =8s\lambda ^{2}\kappa ^{ik}\kappa ^{jl}p^{l}p^{k}, \\
M_{(11)}^{ij}& =8s\lambda ^{2}\kappa ^{ik}\kappa ^{jl}p^{\prime l}p^{\prime
k},\end{aligned}$$implying $$\begin{aligned}
& \left. \langle |{{\mathcal{M}}}|^{2}\rangle =\frac{1}{4s^{2}}\left[
L_{00}^{ij}M_{00}^{ij}\right. \right. \notag \\
& \left. +8s\left( \kappa \right) ^{ik}\left( \kappa \right) ^{jl}\left(
p^{l}p^{k}M_{00}^{ij}+L_{00}^{ij}p^{\prime l}p^{\prime k}\right) \right] .\end{aligned}$$
Doing the corresponding integrations in the solid angle, we achieve $$\int \!\!\left\langle \left\vert {{\mathcal{M}}}\right\vert
^{2}\right\rangle d\Omega =\frac{16\pi e^{4}}{3}\left[ 1+\frac{\lambda ^{2}}{%
4e^{2}}\left( s\left( \kappa ^{2}\right) ^{ii}+4\left( \kappa
^{ij}p^{j}\right) ^{2}\right) \right] ,$$where $\left( \kappa ^{2}\right) ^{ii}=\kappa ^{ij}\kappa ^{ji}$. Choosing a beam direction so that $\kappa ^{ij}p_{j}=0,$ we attain $$\int \left\langle \left\vert {{\mathcal{M}}}\right\vert ^{2}\right\rangle
d\Omega =\frac{16e^{4}\pi }{3}\left( 1+\frac{\lambda ^{2}s}{4e^{2}}\left(
\kappa ^{2}\right) ^{ii}\right) . \label{4.140)}$$$\ $This evaluation leads to $$\sigma =\sigma _{QED}\left( 1+\frac{\lambda ^{2}s}{4e^{2}}\left( \kappa
^{2}\right) ^{ii}\right) , \label{4.142)}$$implying the following upper bound: $$|\lambda \kappa ^{ij}|<5\times 10^{-12}(\mbox{eV})^{-1}. \label{4.143)}$$
We notice that the upper bound on the parity-even parameters have the same order of magnitude as the one on the parity-odd coefficients.
Conclusions
===========
In this work, we have studied the influence of a Lorentz-violating *CPT*-even nonminimal coupling in the context of the Dirac equation, focusing specifically on the $e^{+}+e^{-}\rightarrow \mu ^{+}+\mu ^{-}$ scattering process. This new coupling implied the insertion of a new vertex, increasing the number of Feynman diagrams representing the level tree process. We have carried out the contributions of the nonminimally *CPT*-even LV terms on the unpolarized cross section, using the Casimir’s trick. This evaluation was performed with details for the parity-odd and parity-even coefficients in the ultrarelativistic limit $\left( {m_{e}}={m_{\mu }}=0\right) $. Comparing the attained results with scattering data in the literature [Derrick]{}, we have succeeded in imposing upper bounds at the level of $%
10^{-12}\,(\mbox{eV})^{-1}$ on the parity-odd and parity-even nonbirefringent coefficients of the quantity $\lambda \left( K_{F}\right)
_{\mu \nu \alpha \beta },$ representing a good route to constrain the strength of this new nonminimal coupling in a relativistic environment. It is important to mention that these bounds should not be directly compared with the upper bounds imposed on the coefficients of the dimensionless *CPT*-even tensor $\left( K_{F}\right) _{\mu \nu \alpha \beta }$ in Refs. [@KM1; @Risse]. The bounds here achieved restrain the dimensional quantity $\lambda \left( K_{F}\right) _{\mu \nu \alpha \beta }$, representing a constraint on the way the CPT-even is coupled to the fermion sector.
Although we have restricted our study to the nonbirefringent sector of the *CPT*-even tensor $\left( K_{F}\right) _{\mu \nu \alpha \beta },$ we could have considered the ten birefringent components of the tensor $\left(
K_{F}\right) _{\mu \nu \alpha \beta }$ as well. The point is that these coefficients contribute to the modified cross section also in second order, implying the same upper bound attained on the nonbirefringent components. This reasoning allows to extend the bounds here achieved to all the components of the tensor $\left( K_{F}\right) ,$ that is $\left\vert \lambda
\left( K_{F}\right) \right\vert \leq 10^{-12}\left( \mbox{eV}\right) ^{-1},$ circumventing some cumbersome and unnecessary evaluations.
An interesting investigation concerns the possible connections between this dimension-5 nonminimal coupling and the higher-dimensional operators belonging to the photon sector presented in Ref. [@Kostelec]. The proposed nonminimal coupling is a dimension-5 operator which is not contained in the framework of Refs. [@Kostelec], once this term refers to the interaction between fermions and photons. The connection begins to appear when one performs the radiative corrections generated by this nonminimal coupling. Indeed, the evaluation of the one-loop vacuum polarization diagram of the photon leads to operators with dimension-4 and -6. The dimension-4 operator is exactly the *CPT*-even term $%
(K_{F})_{\mu \nu \rho \sigma }F^{\mu \nu }F^{\rho \sigma }$. The operators of dimension-six are second order in $K_{F}$ and could be encompassed in Ref. [@Kostelec]. The fact that the dimension-4 operator can be generated by radiative corrections allows one to use the existing bounds on the *CPT*-even $\left( K_{F}\right) _{\mu \nu \alpha \beta }$ to attain even better bounds on the magnitude of the quantity $\lambda \left(
K_{F}\right) _{\mu \nu \alpha \beta }.$ The detailed analysis of this issue is under development now [@Frede2].
The authors are grateful to CNPq, CAPES and FAPEMA (Brazilian research agencies) for invaluable financial support.
[99]{} D. Colladay and V. A. Kostelecky, Phys. Rev. D **55**, 6760 (1997); **58**, 116002 (1998); S.R. Coleman and S.L. Glashow, Phys. Rev. D **59**, 116008 (1999).
V. A. Kostelecky and S. Samuel, Phys. Rev. Lett. 63, 224 (1989); Phys. Rev. Lett. 66, 1811 (1991); Phys. Rev. D 39, 683 (1989); Phys. Rev. D 40, 1886 (1989), V. A. Kostelecky and R. Potting, Nucl. Phys. B 359, 545 (1991); Phys. Lett. B 381, 89 (1996); V. A. Kostelecky and R. Potting, Phys. Rev. D 51, 3923 (1995).
V. A. Kostelecky and R. Lehnert, Phys. Rev. D **63**, 065008 (2001).
V. A. Kostelecky, Phys. Rev. D **69**, 105009 (2004); V. A. Kostelecky, Neil Russell, and J. D. Tasson, Phys. Rev. Lett. **100**, 111102 (2008); V. A. Kostelecky and J. D. Tasson, Phys. Rev. Lett. **102**, 010402 (2009); Q. G. Bailey, V.A. Kostelecky, Phys.Rev. D **74**, 045001 (2006); Q. G. Bailey, Phys.Rev. D **80**, 044004 (2009); V.A. Kostelecky and R. Potting, Phys.Rev. D **79**, 065018 (2009); Q. G. Bailey, Phys. Rev. D **82**, 065012 (2010); V.B. Bezerra, C.N. Ferreira, J.A. Helayel-Neto, Phys.Rev. D **71**, 044018 (2005); J.L. Boldo, J.A. Helayel-Neto, L.M. de Moraes, C.A.G. Sasaki, V.J. V. Otoya, Phys. Lett. B **689**, 112 (2010);V. A. Kostelecky and J. D. Tasson, Phys. Rev. D **83**, 016013 (2011).
B. Altschul, Phys. Rev. D **70**, 056005 (2004); G. M. Shore, Nucl. Phys. B **717**, 86 (2005); D. Colladay and V. A. Kostelecky, Phys. Lett. B **511**, 209 (2001); O. G. Kharlanov and V. Ch. Zhukovsky, J. Math. Phys. **48**, 092302 (2007); R. Lehnert, Phys. Rev. D **68**, 085003 (2003); V.A. Kostelecky and C. D. Lane, J. Math. Phys. **40**, 6245 (1999); R. Lehnert, J. Math. Phys. **45**, 3399 (2004); W. F. Chen and G. Kunstatter, Phys. Rev. D **62**, 105029 (2000); B. Goncalves, Y. N. Obukhov, I. L. Shapiro, Phys.Rev.D **80**, 125034 (2009); V. A. Kostelecky and R. Lehnert, Phys. Rev. D** 63,** 065008 (2001)**;** S. Chen, B. Wang, and R. Su, Class. Quant.Grav. **23**, 7581 (2006).
R. Bluhm, V.A. Kostelecky, and N. Russell, Phys. Rev. Lett. **79**, 1432 (1997); R. Bluhm, V.A. Kostelecky, and N. Russell, Phys. Rev. D **57**, 3932 (1998); R. Bluhm, V.A. Kostelecky, C. D. Lane, and N. Russell, Phys. Rev. Lett. **88**, 090801 (2002); R. Bluhm and V.A. Kostelecky, Phys. Rev. Lett.** 84**, 1381 (2000); R. Bluhm, V.A. Kostelecky, and C. D. Lane, Phys. Rev. Lett. **84**, 1098 (2000); R. Bluhm, V.A. Kostelecky, and N. Russell, Phys. Rev. Lett. **82**, 2254 (1999); V.A. Kostelecky and C.D. Lane, Phys. Rev. D **60**, 116010 (1999).
M. Gomes, J. R. Nascimento, A. Yu. Petrov, A. J. da Silva, Phys. Rev. D **81**, 045018 (2010); T. Mariz, J. R. Nascimento, A. Yu. Petrov, Phys. Rev. D **85**, 125003 (2012); G. Gazzola, H. G. Fargnoli, A. P. Baeta Scarpelli, M. Sampaio, and M. C. Nemes, J. Phys. G **39**, 035002 (2012); A. P. Baeta Scarpelli, Marcos Sampaio, M. C. Nemes, B. Hiller, Eur. Phys. J. C **56**, 571 (2008); F.A. Brito, L.S. Grigorio, M.S. Guimaraes, E. Passos, C. Wotzasek, Phys.Rev. D **78**, 125023 (2008); F.A.Brito, E. Passos, P.V. Santos, Europhys. Lett. **95**, 51001 (2011); C. F. Farias, A. C. Lehum, J. R. Nascimento, A. Yu. Petrov, Phys. Rev. D **86**, 065035 (2012);A. P. Baeta Scarpelli, J. Phys. G **39**, 125001 (2012).
S.M. Carroll, G.B. Field and R. Jackiw, Phys. Rev.**D **41**, 1231 (1990).
C. Adam and F. R. Klinkhamer, Nucl. Phys.* *B **607**, 247 (2001); C. Adam and F. R. Klinkhamer, Nucl. Phys.* *B **657**, 214 (2003); A.A. Andrianov and R. Soldati, Phys. Rev. D **51**, 5961 (1995); Phys. Lett. B **435**, 449 (1998); A.A. Andrianov, R. Soldati and L. Sorbo, Phys. Rev. D **59**, 025002 (1998); A. A. Andrianov, D. Espriu, P. Giacconi, R. Soldati, J. High Energy Phys. **0909**, 057 (2009); J. Alfaro, A.A. Andrianov, M. Cambiaso, P. Giacconi, R. Soldati, Int.J.Mod.Phys.A **25**, 3271 (2010); V. Ch. Zhukovsky, A. E. Lobanov, E. M. Murchikova, Phys.Rev. D**73** 065016, (2006).
R. Lehnert and R. Potting, Phys. Rev. Lett. **93**, 110402 (2004); R. Lehnert and R. Potting, Phys. Rev. D **70**, 125010 (2004); B. Altschul, Phys. Rev. D **75**, 105003 (2007); C. Kaufhold and F.R. Klinkhamer, Nucl. Phys. B **734, 1** (2006).
A. P. Baeta Scarpelli, H. Belich, J. L. Boldo, J.A. Helayel-Neto, Phys. Rev. D 67, 085021 (2003); H. Belich *et al.*, Nucl. Phys. B Suppl. 127, 105 (2004); M. B. Cantcheff, Eur. Phys. J. C 46, 247 (2006); J.A. de Sales, T. Costa-Soares, V. J. Vasquez Otoya, Physica A **391**, 5422 (2012); O. M. Del Cima, D. H. T. Franco, and A. H. Gomes, J. M. Fonseca, O. Piguet, Phys. Rev. D **8**5, 065023 (2012).
V. A. Kostelecky and M. Mewes, Phys. Rev. Lett. **87**, 251304 (2001); V. A. Kostelecky and M. Mewes, Phys. Rev. D** 66**, 056005 (2002); V. A. Kostelecky and M. Mewes, Phys. Rev. Lett. **97**, 140401 (2006).
F.R. Klinkhamer and M. Risse, Phys. Rev. D **77**, 016002 (2008); F.R. Klinkhamer and M. Risse, Phys. Rev. D **77**, 117901 (2008); F. R. Klinkhamer and M. Schreck, Phys. Rev. D **78**, 085026 (2008); B. Altschul, Phys. Rev. Lett. **98**, 041603 (2007); M. Schreck, Phys. Rev. D **86**, 065038 (2012).
B. Altschul, Nucl. Phys. B** 796**, 262 (2008); B. Altschul, Phys. Rev. Lett. **98**, 041603 (2007); C. Kaufhold and F.R. Klinkhamer, Phys. Rev. D **76**, 025024 (2007)**.**
V. A. Kostelecky and M. Mewes, Phys. Rev.* *D **80**, 015020 (2009); M. Cambiaso, R. Lehnert, R. Potting, Phys.Rev. D **85** 085023 (2012); M. Mewes, Phys. Rev. D** 85**, 116012 (2012).
H. Belich, T. Costa-Soares, M.M. Ferreira Jr., J. A. Helayël-Neto, Eur. Phys. J. C **41,** 421 (2005).
H. Belich, T. Costa-Soares, M.M. Ferreira Jr., J. A. Helayël-Neto, M.T. D. Orlando, Phys. Lett. **B*** ***639,** 678 (2006); H. Belich, L.P. Colatto, T. Costa-Soares, J.A. Helayël-Neto, M.T.D. Orlando, Eur. Phys. J. C **62**, 425 (2009).
H. Belich, T. Costa-Soares, M.M. Ferreira Jr., J. A. Helayël-Neto, and F. M. O. Moucherek, Phys. Rev. D **74**, 065009 (2006).
H. Belich, M.M. Ferreira Jr., E.O. Silva, and M.T. D. Orlando, Phys. Rev. D **83**, 125025 (2011).
B. Charneski, M. Gomes, R. V. Maluf, A. J. da Silva, Phys. Rev. D **86**, 045003 (2012).
L. R. Ribeiro, E. Passos, C. Furtado, J. Phys. G. **39**, 105004 (2012).
K. Bakke, H. Belich, J. Phys. G **39**, 085001 (2012); K. Bakke, H. Belich, E. O. Silva, J. Math. Phys. **52**, 063505 (2011); J. Phys. G **39**, 055004 (2012); Annalen der Physik (Leipzig) **523**, 910 (2011); K. Bakke and H. Belich, Eur. Phys. J. Plus **127**, 102 (2012).
D. Colladay and V. A. Kostelecky, Phys. Lett. B **511**, 209 (2001); B. Altschul, Phys. Rev. D **70**, 056005 (2004); G. Rubtsov, P. Satunin, and S.Sibiryakov, Phys. Rev. D **86**, 085012 (2012).
M. Derrick *et al*. Phys. Rev. D **31**, 2352 (1985); D. Bender *et al.* Phys. Rev. D **30**, 515 (1984); M. E. Levi *et al*., Phys. Rev. Lett.** 51**, 1941 (1983).
An additional reason for limiting our evaluations on the nonbirefringent sector is provided in the conclusions.
R. Casana, M. M. Ferreira Jr, R.V. Maluf, F.E.P. dos Santos, work under development.
[^1]: e-mails: [email protected], [email protected],
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
author:
- 'R.A. Ong$^{1}$, I.H. Bond$^{2}$, P.J. Boyle$^{3}$, S.M. Bradbury$^{2}$, J.H. Buckley$^{4}$, I. de la Calle Perez$^{2}$, D.A. Carter-Lewis$^{5}$, O. Celik$^{1}$, S. Criswell$^{6}$, W. Cui$^{7}$, M. Daniel$^{5}$, A. Falcone$^{7}$, D.J. Fegan$^{8}$, S.J. Fegan$^{6}$, J.P. Finley$^{7}$, L.F. Fortson$^{3}$, J.A. Gaidos$^{7}$, S. Gammell$^{8}$, K. Gibbs$^{6}$, D. Hanna$^{9}$, J. Hall$^{10}$, A.M. Hillas$^{2}$, J. Holder$^{2}$, D. Horan$^{6}$, M. Jordan$^{4}$, M. Kertzman$^{11}$, D. Kieda$^{10}$, J. Kildea$^{9}$, J. Knapp$^{2}$, K. Kosack$^{4}$, H. Krawczynski$^{4}$, F. Krennrich$^{5}$, S. LeBohec$^{5}$, E. Linton$^{3}$, J. Lloyd-Evans$^{2}$, P. Moriarty$^{12}$, D. Müller$^{3}$, T. Nagai$^{10}$, S. Nolan$^{7}$, R. Pallassini$^{2}$, F. Pizlo$^{7}$, B. Power-Mooney$^{8}$, J. Quinn$^{8}$, K. Ragan$^{9}$, P. Rebillot$^{4}$, J. Reynolds$^{13}$, H.J. Rose$^{2}$, M. Schroedter$^{6}$, G.H. Sembroski$^{7}$, S.P. Swordy$^{3}$, V.V. Vassiliev$^{10}$, S.P. Wakely$^{3}$, G. Walker$^{10}$, T.C. Weekes$^{6}$, J. Zweerink$^{1}$'
- |
$^{1}$Dept. of Physics & Astron., Univ. of California, Los Angeles, CA 90095, USA\
$^{2}$Dept. of Physics, University of Leeds, Leeds, LS2 9JT, Yorkshire, UK\
$^{3}$Enrico Fermi Inst., University of Chicago, Chicago, IL 60637, USA\
$^{4}$Dept. of Physics, Washington University, St. Louis, MO 63130, USA\
$^{5}$Physics & Astronomy Dept., Iowa State University, Ames, IA 50011, USA\
$^{6}$F. Lawrence Whipple Obs., Harvard-Smithsonian CfA, Amado, AZ 85645, USA\
$^{7}$Dept. of Physics, Purdue University, West Lafayette, IN 47907, USA\
$^{8}$Dept. of Physics, National University of Ireland, Dublin 4, Ireland\
$^{9}$Dept. of Physics, McGill University, Montreal, Quebec H3A 2T8, Canada\
$^{10}$High Energy Astrophysics Inst., Univ. of Utah, Salt Lake City, UT 84112, USA\
$^{11}$Physics Dept., De Pauw University, Greencastle, IN 46135, USA\
$^{12}$School of Science, Galway-Mayo Institute of Technology, Galway, Ireland\
$^{13}$Dept. of Applied Physics, Cork Institute of Technology, Cork, Ireland\
---
The VERITAS Project
===================
Abstract {#abstract .unnumbered}
--------
VERITAS is a new, major ground-based $\gamma$-ray observatory designed to significantly advance our understanding of extreme astrophysical processes in the universe. The observatory comprises seven large-aperture (12$\,$m diameter) Cherenkov telescopes, each equipped with an imaging camera. The first phase of VERITAS (consisting of four telescopes) is currently under construction. Here we outline the key features of VERITAS and provide an update on its status \[1\].
Introduction
------------
Ground-based $\gamma$-ray astronomy came of age in the last decade. The discovery of numerous sources by telescopes using the imaging atmospheric Cherenkov technique proved that very high energy (VHE) astronomy can be done from the ground. Roughly one dozen VHE sources have now been detected with varying levels of significance \[2\]. The established sources include pulsar-powered nebulae, active galaxies of the BL Lac type, and shell type supernova remnants.
These exciting discoveries strongly motivate the construction of new instruments to substantially increase the VHE source catalog and to enable much more detailed studies of individual sources. VERITAS combines the successful features of the Whipple Observatory (large-aperture reflector, imaging camera) with those of HEGRA (array of reflectors). Relative to current operating imaging telescopes, VERITAS will have substantially [*better flux sensitivity*]{} (factor of five to ten improvement, depending on energy), [*reduced energy threshold*]{} (peak energy near 100 GeV), [*improved energy resolution*]{} (resolution of 10-15% over a broad energy range), and [*improved angular resolution*]{} (4.3 arc-min at 1 TeV).
VERITAS Design
--------------
Serious consideration of VERITAS started in 1996, and its design was developed and finalized over the next several years. The design uses modular construction and proven technology to a large degree, but it also encompasses new technical innovations where appropriate. VERITAS will be an array of seven telescopes, each employing a $12\,$m diameter optical reflector. Each telescope has a camera of 499 photomultiplier tube elements, covering a field of view of 3.5$^\circ$ diameter. The photomultiplier tubes are read out through high-bandwidth electronics by a Flash ADC (FADC) system sampling at 500$\,$MSps. The FADC sampling, a flexible trigger system, and extensive electronic and optical calibration systems are among the important new capabilities of VERITAS. Details on the design and expected performance of VERITAS have been published earlier \[3,4\]. This report provides an update on the key developments that have happened recently.
[rcl]{} Characteristic & E& Value\
Peak Energy $^a$ & & 110 GeV\
Flux sensitivity & 100GeV & 3.4$\times$10$^{-11}$cm$^{-2}$s$^{-1}$\
& 1TeV & 6.5$\times$10$^{-13}$cm$^{-2}$s$^{-1}$\
& 10TeV & 2.1$\times$10$^{-13}$cm$^{-2}$s$^{-1}$\
Angular resolution & 100GeV & 7.5 arc min\
& 1TeV & 4.3 arc min\
& 10TeV & 1.6 arc min\
Collection area & 100GeV & 3.3$\times$10$^8$cm$^2$\
& 1TeV & 2.2$\times$10$^9$cm$^2$\
Crab Nebula & $>$100GeV & 40/minute\
$\gamma$-ray rates & $>$300GeV & 15/minute\
& $>$1TeV & 4/minute\
Energy resolution$^b$ & & $<$15%\
\
\
VERITAS will be sited on Mt. Hopkins in southern Arizona, USA, near the Basecamp of the Whipple Observatory (1350$\,$m altitude). The site will be well protected from light pollution (within 20% of the darkest sites available in the U.S.). The construction will be carried out in two phases. During the first phase (VERITAS-4), four telescopes will be constructed and deployed as shown in Figure 1. In the second phase, three additional telescopes will be added to complete the hexagonal array.
VERITAS-4 Performance
---------------------
We have carried out detailed simulation studies to estimate the expected performance of the four telescope array, VERITAS-4 – some key results are shown in Table 1. In quoting sensitivity, we make the conservative requirement of at least a five standard deviation $\gamma$-ray excess in each energy bin, of width one-quarter decade.
Technical Progress
------------------
A great deal of technical progress has been made on VERITAS, and we are well on our way towards the construction of a prototype telescope encompassing all important design elements. Here we describe the salient features of the VERITAS design and recent progress made on the construction of the prototype that will become operational in mid-2003.
### Telescope, Optical Support Structure, and Mirrors
The VERITAS telescopes consist of a tubular-steel, space-frame optical support structure (OSS) mounted on a commercial positioner. In the OSS design, as shown in Figure 1, quadrapod arms penetrate the mirror surface to hold up the camera. We expect excellent optical performance from the OSS – the blur will be well less than $0.01^\circ$ over the usable elevation range of the telescope, and the de-centering will be less than $0.002^\circ$. The positioner for the prototype telescope will be delivered in April 2003.
We use the Davies-Cotton optical design in which the telescope reflecting surface is spherical and the mirror facets are identical in shape. Each $12\,$m diameter (f/1.0) mirror is made from 350 hexagonal facets and has a total mirror area of 110$\,$m$^2$. The facets are made of float glass that has been slumped and polished by the manufacturer (DOTI Technologies). They are aluminized and anodized in a dedicated facility near the Whipple Observatory Basecamp. All the facets for the prototype telescope have been received; mirror coating is in progress. Laser measurements indicate that the optical quality of the facets is better than originally specified: 1) the tolerance of the radius of curvature is better than $0.4$% (as opposed to 1.0% specified), and 2) the average blur is less than 0.5$\,$cm (as opposed to 1.0$\,$cm specified).
{height="16pc"}
### Camera, Cables, & High Voltage
The cameras for VERITAS comprise 499 photomultiplier tubes (PMTs), arranged in a close-packed hexagonal pattern with an angular separation between PMTs of $0.15^\circ$. Light cones increase the photon collection efficiency and protect the PMT’s from stray light pollution. A high-speed, custom-made amplifier is used to boost the signal before transmission through $40\,$m of high-bandwidth 75$\Omega$ coaxial cable. The high voltage level for each PMT can be individually programmed, and custom-designed electronics monitor individual PMT anode currents.
All of the key camera components were tested in November 2001 when thirty PMT channels were installed in the focus box of the Whipple 10m telescope. Performance characteristics (e.g. linearity, rise time, noise levels, etc.) of the full electronics chain of VERITAS were verified. The assembly of the camera for the prototype telescope is currently in progress at the University of Chicago. The integrated camera will be shipped to Mt. Hopkins in spring 2003. Figure 2 shows the camera box during assembly and an FADC trace for a Cherenkov pulse.
{height="12pc"}
### Flash ADCs (FADCs) and Data Acquisition
The custom-built FADCs digitize the Cherenkov pulse waveform at a rate of 500$\,$MSps to provide the maximum possible information about the shape and time structure of the pulse. Each PMT signal is sampled by a separate FADC with an effective dynamic range of 11 bits and a memory depth of 64$\,\mu$s. The FADCs are packaged in 9U VME boards with 10 channels/board. The data acquisition is based around standard VME architecture, comprising a fast VME backplane and crate Single Board Computers (SBC’s) connected to a local event-building workstation via the fast SCI protocol. The FADC design has been finalized, and the boards for the prototype telescope are being built. The VME readout for the FADC system is also near completion; crate-to-crate transfer speeds of 100$\,$MB/s have been achieved.
### Trigger Electronics and Calibration
VERITAS employs a three level trigger system to select Cherenkov events at the lowest possible energy threshold. Level 1 consists of constant fraction discriminators (CFDs) to determine when a PMT pulse exceeds a given threshold. The CFDs are mounted directly on the main FADC board to minimize noise and signal delay. Level 2 comprises a hardware pattern trigger designed to select compact Cherenkov events, as opposed to night sky background overlaps. Level 3 uses the Level 2 telescope triggers to determine when the array has triggered. The designs for all three trigger levels are well advanced and complete systems for the prototype telescope are now being integrated with the other electronics.
The calibration system is designed to calibrate and monitor the performance of each telescope and the combined array. There are three major calibration components: charge injection, optical injection, and atmospheric monitoring. The charge injection system distributes a calibrated amount of charge to the front-end electronics. The initial optical system uses a dye laser to simultaneously flash all PMTs in the camera.
### Software
The online software can be divided into a number of components: 1) FADC and VME data acquisition, 2) Telescope acquisition (event building), 3) Array acquisition/online analysis (Quicklook), and 4) Array Control and Database. The software uses object-oriented (C++) programs running on Intel-based Linux computers. Reliance is made on widely available software packages (e.g. CORBA, SQL, Qt, etc.) for database and graphic user interface tasks and for inter-process communication and control. The majority of the online code for the prototype telescope has been written, and now the major task is integration. Initial software systems for offline analysis and simulation are also in place; these will be augmented and refined in the future.
Schedule & Summary
------------------
VERITAS is a new state-of-the-art ground-based $\gamma$-ray observatory for VHE astronomy. An initial phase (VERITAS-4) consists of four atmospheric Cherenov telescopes. A prototype telescope will begin operating in summer 2003, and funding permitting, first light for the VERITAS-4 array is expected in late 2005, well before the launch of the Gamma-ray Large Area Space Telescope (GLAST).
References
----------
1. Talk presented at the international symposium: [*The Universe Viewed in Gamma-Rays*]{}, 25-28 Sep 2002, Kashiwa, Japan. Symposium proceedings published by Universal Academy Press, Inc, edited by R. Enomoto.
2. Ong, R.A. 2003, these proceedings.
3. Quinn, J. [*et al.*]{} 2001, Proc. 27th. Int. Cosmic Ray Conf. (Hamburg) 7, 2781.
4. Weekes, T.C. [*et al.*]{} 2002, Astroparticle Physics 17, 221.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'Monogeny classes and epigeny classes have proved to be useful in the study of direct sums of uniserial modules and other classes of modules. In this paper, we show that they also turn out to be useful in the study of direct products.'
address:
- 'Adel Alahmadi, Department of Mathematics, Faculty of Science, King Abdulaziz University, Jeddah 21589, Saudi Arabia'
- 'Alberto Facchini, Dipartimento di Matematica Università di Padova, 35121 Padova, Italy'
author:
- Adel Alahmadi
- Alberto Facchini
title: Direct Products of Modules Whose Endomorphism Rings have at Most Two Maximal Ideals
---
[^1]
Introduction
============
Two right $R$-modules $M$ and $N$ are said to [*belong to the same monogeny class*]{} (written $[M]_m=[N]_m$) if there exist a monomorphism $M\to N$ and a monomorphism $N\to M$. Dually, $M$ and $N$ are said to [*belong to the same epigeny class*]{} ($[M]_e=[N]_e$) if there exist an epimorphism $M\to N$ and an epimorphism $N\to M$. Recall that a module is [*uniserial*]{} if its lattice of submodules is linearly ordered under inclusion. In [@TAMS Theorem 1.9], it was proved that if $U_1$, $\dots,$ $U_n$, $V_1$, $\dots,$ $V_t$ are non-zero uniserial right $R$-modules, then $U_1\oplus\dots\oplus U_n\cong V_1\oplus\dots\oplus V_t$ if and only if $n=t$ and there are two permutations $\sigma,\tau$ of $\{1,2,\dots,n\}$ such that $[U_i]_m=[V_{\sigma(i)}]_m$ and $[U_i]_e=[V_{\tau(i)}]_e$ for every $i=1,2,\dots,
n$. This result, which made possible the solution of a problem [@Warf p. 189] posed by Warfield in 1975, was then generalized in various directions. On the one hand, it was extended to the case of arbitrary, non-necessarily finite, families $\{\,U_i\mid i\in I\,\}$, $\{\,V_j\mid j\in J\,\}$ of uniserial modules [@Pri Theorem 2.6] (see Theorem \[3.2\] below). On the other hand, it was shown that similar theorems hold not only for uniserial modules, but also for cyclically presented modules over a local ring $R$, for kernels of morphisms between indecomposable injective modules, for couniformly presented modules, and more generally, for several classes of modules with at most two maximal right ideals (see [@Bull Section 5] and [@AlbPav5]).
In this paper, we prove that a similar result holds not only for direct sums, but also for direct products of arbitrary families $\{\,U_i\mid i\in I\,\}$, $\{\,V_j\mid j\in J\,\}$ of uniserial modules. We show (Theorem \[suff’\]) that if there exist two bijections $\sigma,\tau\colon
I\to J$ such that $[U_i]_m=[V_{\sigma(i)}]_m$ and $[U_i]_e=[V_{\tau(i)}]_e$ for every $i\in I$, then $\prod_{i\in I}U_i\cong \prod_{j \in J}V_j$. In fact, the theorem we prove is much more general, and involves completely prime ideals in categories of modules whose endomorphism rings have at most two maximal right ideals (Theorem \[suff\]). This allows us to apply our theorem not only to uniserial modules, but also to several other classes of modules, like the class of cyclically presented modules over a local ring and the class of kernels of morphisms between indecomposable injective modules (Section \[App\]).
We then show with some examples that in general it is not possible to reverse our result (find the converse of it). It is possible to reverse it only in the particular case of slender modules (Theorems \[nec\] and \[buupbp\]). For this class of modules, it is possible to argue as in the recent paper [@Franetic].
The rings we deal with are associative rings with identity $1\ne 0$, and modules are unitary modules.
The main result
===============
In order to present our result in the most general setting, that of modules whose endomorphism rings have at most two maximal right ideals, we adopt the point of view of [@AlbPav5 Section 6]. Thus, let $R$ be an associative ring with identity and ${\operatorname{Mod-\!}}R$ the category of all right $R$-modules. Let ${{\mathcal C}}$ be a full subcategory of ${\operatorname{Mod-\!}}R$ whose class of objects $\operatorname{Ob}({{\mathcal C}})$ consists of indecomposable right $R$-modules. Recall that a [*completely prime ideal*]{} ${{\mathcal P}}$ of ${{\mathcal C}}$ consists of a subgroup ${{\mathcal P}}(A,B)$ of the additive abelian group ${\operatorname{Hom}}_R(A,B)$ for every pair of objects $A,B\in{}\operatorname{Ob}({{\mathcal C}})$ such that: (1) for every $A,B,C\in{}\operatorname{Ob}({{\mathcal C}})$, every $f\colon A\to B$ and every $g\colon B\to C$, there holds $gf\in {{\mathcal P}}(A,C)$ if and only if either $f\in {{\mathcal P}}(A,B)$ or $g\in {{\mathcal P}}(B,C)$; and (2) ${{\mathcal P}}(A,A)$ is a proper subgroup of ${\operatorname{Hom}}_R(A,A)$ for every object $A\in{}\operatorname{Ob}({{\mathcal C}})$.
If $A,B$ are objects of ${{\mathcal C}}$, we say that $A$ and $B$ [*belong to the same ${{\mathcal P}}$ class*]{}, and write $[A]_{{{\mathcal P}}}=[B]_{{{\mathcal P}}}$, if there exist $f\colon A\to B$ and $g\colon B\to A$ such that $f\notin{{\mathcal P}}(A,B)$ and $g\notin{{\mathcal P}}(B,a)$, that is, if ${{\mathcal P}}(A,B)\ne{\operatorname{Hom}}(A,B)$ and ${{\mathcal P}}(B,A)\ne{\operatorname{Hom}}(B,A)$. The full subcategory ${{\mathcal C}}$ of ${\operatorname{Mod-\!}}R$ is said to [*satisfy Condition (DSP)*]{} (direct summand property) if whenever $A,B,C,D$ are right $R$-modules with $A\oplus B\cong C\oplus D$ and $A,B,C\in{}\operatorname{Ob}({{\mathcal C}})$, then also $D\in{}\operatorname{Ob}({{\mathcal C}})$.
We begin by recalling the Weak Krull-Schmidt Theorem, the proof of which can be found in [@AlbPav5 Theorem 6.2], followed by a preparatory lemma.
\[WKST\][(Weak Krull-Schmidt Theorem)]{} Let ${{\mathcal C}}$ be a full subcategory of ${\operatorname{Mod-\!}}R$ in which all objects are indecomposable right $R$-modules and let ${{\mathcal P}},{{\mathcal Q}}$ be two completely prime ideals of ${{\mathcal C}}$ with the property that, for every $A\in{}\operatorname{Ob}({{\mathcal C}})$, $f\colon A\to A$ is an automorphism if and only if $f\notin {{\mathcal P}}(A,A)\cup {{\mathcal Q}}(A,A)$. Let $A_1,\dots,A_n$, $B_1,\dots,B_t$ be objects of ${{\mathcal C}}$. Then the modules $A_1 \oplus \dots \oplus A_n$ and $B_1
\oplus \dots \oplus B_t$ are isomorphic if and only if $n=t$ and there are two permutations $\sigma, \tau$ of $\{ 1, 2,
\dots, n \}$ with $[A_i]_{{{\mathcal P}}} = [B_{\sigma(i)}]_{{{\mathcal P}}}$ and $[A_i]_{{{\mathcal Q}}} = [B_{\tau(i)}]_{{{\mathcal Q}}}$ for every $i=1,\dots,n$.
\[2.2\] Let ${{\mathcal C}}$ be a full subcategory of ${\operatorname{Mod-\!}}R$ in which all objects are indecomposable right $R$-modules and let ${{\mathcal P}},{{\mathcal Q}}$ be a pair of completely prime ideals of ${{\mathcal C}}$ with the property that, for every $A\in{}\operatorname{Ob}({{\mathcal C}})$, $f\colon A\to A$ is an automorphism if and only if $f\notin {{\mathcal P}}(A,A)\cup {{\mathcal Q}}(A,A)$. Assume that ${{\mathcal C}}$ satisfies Condition (DSP). Let $A,B,C\in{}\operatorname{Ob}({{\mathcal C}})$ with $[C]_{{{\mathcal P}}}=[A]_{{{\mathcal P}}}$ and $[C]_{{{\mathcal Q}}}=[B]_{{{\mathcal Q}}}$. Then there exists $D\in{}\operatorname{Ob}({{\mathcal C}})$ with $A\oplus B\cong C\oplus D$. Moreover, $[D]_{{{\mathcal P}}}=[B]_{{{\mathcal P}}}$ and $[D]_{{{\mathcal Q}}}=[A]_{{{\mathcal Q}}}$.
Let $A,B,C$ be objects of ${{\mathcal C}}$ such that $[C]_{{{\mathcal P}}}=[A]_{{{\mathcal P}}}$ and $[C]_{{{\mathcal Q}}}=[B]_{{{\mathcal Q}}}$. Then there exist morphisms $f\colon C\to A$, $g\colon A\to C$, $h\colon C\to B$ and $\ell\colon B\to C$ such that $f\notin{{\mathcal P}}(C,A)$, $g\notin{{\mathcal P}}(A,C)$, $h\notin{{\mathcal Q}}(C,B)$ and $\ell\notin{{\mathcal Q}}(B,C)$. Thus $gf\notin{{\mathcal P}}(C,C)$ and $\ell h\notin{{\mathcal Q}}(C,C)$.
We have four cases according to whether $gf\notin{{\mathcal Q}}(C,C)$ or $gf\in{{\mathcal Q}}(C,C)$ and $\ell h\notin{{\mathcal P}}(C,C)$ or $\ell h\in{{\mathcal P}}(C,C)$. If $gf\notin{{\mathcal Q}}(C,C)$, then $gf$ is an automorphism of $C$. Since the composite mapping of $h\colon C\to B$ and $(\ell h)^{-1}\ell\colon B\to C$ is the identity mapping of $C$, it follows that $C$ is isomorphic to a direct summand of $B$. But $B$ and $C$ are indecomposable, so that $C\cong B$. In particular, $[B]_{{{\mathcal P}}}=[A]_{{{\mathcal P}}}$. It follows that $D:=A$ has the required properties. Similarly if $\ell h\notin{{\mathcal P}}(C,C)$.
Hence it remains to consider the case $gf\in{{\mathcal Q}}(C,C)$ and $\ell h\in{{\mathcal P}}(C,C)$. In this case, we have that $gf+\ell h\notin{{\mathcal P}}(C,C)\cup{{\mathcal Q}}(C,C)$, so $gf+\ell h$ is an automorphism of $C$. Now the composite mapping of $\binom{f}{h}\colon C\to A\oplus B$ and $(gf+\ell h)^{-1}(g\ \ell)\colon A\oplus B\to C$ is the identity mapping of $C$. Thus $C$ is isomorphic to a direct summand of $A\oplus B$. By Condition [*(DSP)*]{}, there exists an object $D$ of ${{\mathcal C}}$ with $A\oplus B\cong C\oplus D$. Finally, $[D]_{{{\mathcal P}}}=[B]_{{{\mathcal P}}}$ and $[D]_{{{\mathcal Q}}}=[A]_{{{\mathcal Q}}}$ by Theorem \[WKST\].
We are ready for the proof of the main result of this paper.
\[suff\] Let ${{\mathcal C}}$ be a full subcategory of ${\operatorname{Mod-\!}}R$ in which all objects are indecomposable right $R$-modules and let ${{\mathcal P}},{{\mathcal Q}}$ be two prime ideals of ${{\mathcal C}}$ with the property that, for every $A\in{}\operatorname{Ob}({{\mathcal C}})$, $f\colon A\to A$ is an automorphism if and only if $f\notin {{\mathcal P}}(A,A)\cup {{\mathcal Q}}(A,A)$. Assume that ${{\mathcal C}}$ satisfies Condition (DSP). Let $\{\,A_i\mid i\in I\,\}$ and $\{\,B_j\mid j\in J\,\}$ be two families of objects of ${{\mathcal C}}$. Assume that there exist two bijections $\sigma,\tau\colon
I\to J$ such that $[A_i]_{{{\mathcal P}}} = [B_{\sigma(i)}]_{{{\mathcal P}}}$ and $[A_i]_{{{\mathcal Q}}} = [B_{\tau(i)}]_{{{\mathcal Q}}}$ for every $i\in I$. Then the $R$-modules $\prod_{i\in I}A_i$ and $\prod_{j \in J}B_j$ are isomorphic.
The proof is very similar to the proof of [@DF Theorem 3.1]. Let $\sigma,\tau\colon
I\to J$ be two bijections such that $[A_i]_{{{\mathcal P}}}=[B_{\sigma(i)}]_{{{\mathcal P}}}$ and $[A_i]_{{{\mathcal Q}}}=[B_{\tau(i)}]_{{{\mathcal Q}}}$ for every $i\in I$. We want to prove that $\prod_{i\in I}A_i\cong \prod_{j \in J}B_j$. Re-indexing the family $\{\,B_j\mid j\in J\,\}$ in the set $I$ via the bijection $\sigma$, we can suppose, without loss of generality, that $I=J$ and $\sigma\colon I\to I$ is the identity. Thus we have that $\tau$ is an element of the symmetric group $S_I$ of all permutations of $I$, and that $[A_i]_{{{\mathcal P}}}=[B_i]_{{{\mathcal P}}}$ and $[A_i]_{{{\mathcal Q}}}=[B_{\tau(i)}]_{{{\mathcal Q}}}$ for every $i\in I$. We must show that $\prod_{i\in I}A_i\cong \prod_{i\in I}B_i$.
The symmetric group $S_I$ acts on the set $I$. Let $C$ be the cyclic subgroup of $S_I$ generated by $\tau$, so that $C$ acts on $I$. For every $i\in I$, let $[i]=\{\,\tau^z(i)\mid z\in {\mathbb{Z}}\,\}$ be the $C$-orbit of $i$. As $C$ is countable, the $C$-orbits are either finite or countable. We claim that $\prod_{k\in[i]}A_k\cong \prod_{k\in[i]}B_k$. If the orbit $[i]$ is finite, in which case direct sum and direct product coincide, the claim follows immediately from Theorem \[WKST\], because the modules $A_k$ and $B_k$ belong to the same ${{\mathcal P}}$ class for every $k$, and the modules in the finite sets $\{\,A_k\mid k\in[i]\,\}$ and $\{\,B_k\mid k\in[i]\,\}$ have their ${{\mathcal Q}}$ classes permuted by the restriction of $\tau$ to the orbit $[i]$. Hence we can assume that the orbit $[i]$ is infinite. For simplicity of notation, set $i_z:=\tau^z(i)$, $A_z:=A_{i_z}$ and $B_z:=B_{i_z}$ for every $z\in{\mathbb{Z}}$. In this notation, we have that $\tau(i_z)=\tau^{z+1}(i)=i_{z+1}$ for every $z$, so that $[A_z]_{{{\mathcal P}}}=[B_z]_{{{\mathcal P}}}$ and $[A_z]_{{{\mathcal Q}}}=[B_{{z+1}}]_{{{\mathcal Q}}}$ for every $z\in {\mathbb{Z}}$, and we must show that $\prod_{z\in{\mathbb{Z}}}A_z\cong \prod_{z\in{\mathbb{Z}}}B_z$.
For every integer $n\ge 0$, we will construct by induction on $n$ a 4-tuple $$(X_n,Y_n,X'_n,Y'_n)$$ of modules isomorphic to objects of ${{\mathcal C}}$ with the following five properties for every $n\ge 0$:
1. $X_n\oplus Y_n=A_{n}\oplus A_{-n-1}$;
2. $X'_n\oplus Y'_n=B_{n+1}\oplus B_{-n-1}$;
3. $[X_n]_{{{\mathcal P}}}=[B_n]_{{{\mathcal P}}}$ and $[X_n]_{{{\mathcal Q}}}=[B_{-n}]_{{{\mathcal Q}}}$;
4. $X_{n+1}\cong X'_n$;
5. $Y_n\cong Y'_n$.
As $[B_0]_{{{\mathcal P}}}=[A_0]_{{{\mathcal P}}}$ and $[B_0]_{{{\mathcal Q}}}=[A_{-1}]_{{{\mathcal Q}}}$, there is a direct-sum decomposition $X_0\oplus Y_0$ of $A_0\oplus A_{-1}$ such that $X_0\cong B_0$, $[Y_0]_{{{\mathcal P}}}=[A_{-1}]_{{{\mathcal P}}}$ and $[Y_0]_{{{\mathcal Q}}}=[A_0]_{{{\mathcal Q}}}$ (Lemma \[2.2\]). Hence $[Y_0]_{{{\mathcal P}}}=[B_{-1}]_{{{\mathcal P}}}$ and $[Y_0]_{{{\mathcal Q}}}=[B_1]_{{{\mathcal Q}}}$. Thus we can apply Lemma \[2.2\] to the module $Y_0$ and the direct sum $B_1\oplus B_{-1}$, getting that there is a direct-sum decomposition $B_1\oplus B_{-1}= X'_0\oplus Y'_0$ of $B_1\oplus B_{-1}$ with $Y'_0\cong Y_0$ and $X'_0\in{}\operatorname{Ob}({{\mathcal C}})$. This proves that the 4-tuple $(X_0,Y_0,X'_0,Y'_0)$ has the required properties.
Now suppose $n\ge 1$ and that the 4-tuple $(X_t,Y_t,X'_t,Y'_t)$ with the required properties has been defined for every $t$ with $0\le t<n$. Then $X_{n-1}\oplus Y_{n-1}=A_{n-1}\oplus A_{-n}$ by (a), and $[X_{n-1}]_{{{\mathcal P}}}=[B_{n-1}]_{{{\mathcal P}}}=[A_{n-1}]_{{{\mathcal P}}}$ by (c). Thus, by Theorem \[WKST\], we know that $[Y_{n-1}]_{{{\mathcal P}}}=[A_{-n}]_{{{\mathcal P}}}=[B_{-n}]_{{{\mathcal P}}}$. Similarly, from $X_{n-1}\oplus Y_{n-1}=A_{n-1}\oplus A_{-n}$ and $[X_{n-1}]_{{{\mathcal Q}}}=[B_{-n+1}]_{{{\mathcal Q}}}=[A_{-n}]_{{{\mathcal Q}}}$ (property (c)), we have that $[Y_{n-1}]_{{{\mathcal Q}}}=[A_{n-1}]_{{{\mathcal Q}}}=[B_n]_{{{\mathcal Q}}}$. Applying (e), it follows that $[Y'_{n-1}]_{{{\mathcal P}}}=[B_{-n}]_{{{\mathcal P}}}$ and $[Y'_{n-1}]_{{{\mathcal Q}}}=[B_n]_{{{\mathcal Q}}}$. These last two equalities and $X'_{n-1}\oplus Y'_{n-1}=B_{n}\oplus B_{-n}$ imply that $$[X'_{n-1}]_{{{\mathcal P}}}=[B_{n}]_{{{\mathcal P}}}\quad\mbox{\rm and}\quad [X'_{n-1}]_{{{\mathcal Q}}}=[B_{-n}]_{{{\mathcal Q}}}.\label{(3)}$$ Thus $[X'_{n-1}]_{{{\mathcal P}}}=[A_{n}]_{{{\mathcal P}}}$ and $[X'_{n-1}]_{{{\mathcal Q}}}=[A_{-n-1}]_{{{\mathcal Q}}}$. By Lemma \[2.2\], there exist $X_n$ and $Y_n$ such that $X_n\oplus Y_n=A_n\oplus A_{-n-1}$ and $X'_{n-1}\cong X_n$. Thus properties (a) and (d) hold. Property (c) also follows easily.
From equalities (\[(3)\]), we have that $[X_n]_{{{\mathcal P}}}=[X'_{n-1}]_{{{\mathcal P}}}=[B_n]_{{{\mathcal P}}}=[A_n]_{{{\mathcal P}}}$ and $[X_n]_{{{\mathcal Q}}}=[X'_{n-1}]_{{{\mathcal Q}}}=[B_{-n}]_{{{\mathcal Q}}}=[A_{-n-1}]_{{{\mathcal Q}}}$. These equalities and $X_n\oplus Y_n=A_n\oplus A_{-n-1}$ imply that $[Y_n]_{{{\mathcal P}}}=[A_{-n-1}]_{{{\mathcal P}}}$ and $[Y_n]_{{{\mathcal Q}}}=[A_n]_{{{\mathcal Q}}}$. Hence $[Y_n]_{{{\mathcal P}}}=[B_{-n-1}]_{{{\mathcal P}}}$ and $[Y_n]_{{{\mathcal Q}}}=[B_{n+1}]_{{{\mathcal Q}}}$. These two equalities give a direct-sum decomposition $X'_n\oplus Y'_n=B_{n+1}\oplus B_{-n-1}$ with $Y_n\cong Y'_n$ (Lemma \[2.2\] and Condition [*(DSP)*]{}). Thus properties (b) and (e) also hold, and the construction by induction is completed.
Notice that from (c) it follows that $X_0\cong B_0$. Then $$\begin{array}{l}\prod_{z\in{\mathbb{Z}}}A_z\cong \prod_{n\ge0}A_{n}\oplus A_{-n-1}\cong \prod_{n\ge0}X_n\oplus Y_n\cong \\ \qquad\cong X_0\oplus \prod_{n\ge0}X_{n+1}\oplus Y_n\cong B_0\oplus \prod_{n\ge0}X'_n\oplus Y'_n
\cong \\ \qquad\cong B_0\oplus \prod_{n\ge0}B_{n+1}\oplus B_{-n-1}\cong \prod_{z\in{\mathbb{Z}}}B_z.\end{array}$$ This concludes the proof of the claim. Now the orbits $[i]$ form a partition of $I$, that is, $I$ is the disjoint union of the orbits, and $\prod_{k\in[i]}A_k\cong \prod_{k\in[i]}B_k$ for every $i$ by the claim. Taking the direct product we conclude that $\prod_{i\in I}A_i\cong\prod_{i \in I}B_i$, as desired.
In [@AdelAlbKor], the authors have considered the condition “the canonical functor ${{\mathcal C}}\to {{\mathcal C}}/{{\mathcal P}}\times{{\mathcal C}}/{{\mathcal Q}}$ is local”. Here an additive functor $F\colon{{\mathcal A}}\to{{\mathcal B}}$ between preadditive categories ${{\mathcal A}}$ and ${{\mathcal B}}$ is said to be a [*local*]{} functor if, for every morphism $f\colon A\to B$ in the category ${{\mathcal A}}$, $F(f)$ isomorphism in ${{\mathcal B}}$ implies $f$ isomorphism in ${{\mathcal A}}$. Let us prove that if ${{\mathcal C}}$ is a preadditive category, ${{\mathcal P}}$ and ${{\mathcal Q}}$ are two completely prime ideals of ${{\mathcal C}}$, and for every $A\in{}\operatorname{Ob}({{\mathcal C}})$, $f\colon A\to A$ is an automorphism if and only if $f\notin {{\mathcal P}}(A,A)\cup {{\mathcal Q}}(A,A)$, then the canonical functor ${{\mathcal C}}\to {{\mathcal C}}/{{\mathcal P}}\times{{\mathcal C}}/{{\mathcal Q}}$ is local.
In order to see this, let ${{\mathcal C}}$ be a preadditive category, ${{\mathcal P}}$ and ${{\mathcal Q}}$ be two completely prime ideals of ${{\mathcal C}}$, and suppose that for every $A\in{}\operatorname{Ob}({{\mathcal C}})$, $f\colon A\to A$ is an automorphism if and only if $f\notin {{\mathcal P}}(A,A)\cup {{\mathcal Q}}(A,A)$. We will apply [@AdelAlbKor Theorem 2.4] to the ideal ${{\mathcal I}}:={{\mathcal P}}\cap{{\mathcal Q}}$ of ${{\mathcal C}}$. For any object $A\in{}\operatorname{Ob}({{\mathcal C}})$, ${{\mathcal P}}(A,A)\cup{{\mathcal Q}}(A,A)$ is the set of all non-invertible elements of the ring ${\mbox{\rm End}}_R(A)$. It follows that every right (or left) ideal of ${\mbox{\rm End}}_R(A)$ is either contained in ${{\mathcal P}}(A,A)$ or ${{\mathcal Q}}(A,A)$, that is, the maximal right (or left) ideals of ${\mbox{\rm End}}_R(A)$ are at most ${{\mathcal P}}(A,A)$ and ${{\mathcal Q}}(A,A)$. In any case, ${{\mathcal I}}(A,A)={{\mathcal P}}(A,A)\cap{{\mathcal Q}}(A,A)$ is contained in the Jacobson radical of ${\mbox{\rm End}}_R(A)$. By [@AdelAlbKor Theorem 2.4], if ${{\mathcal J}}$ is the Jacobson radical of ${{\mathcal C}}$, one has that ${{\mathcal I}}={{\mathcal P}}\cap{{\mathcal Q}}\subseteq {{\mathcal J}}$. It follows that the canonical functor ${{\mathcal C}}\to {{\mathcal C}}/{{\mathcal P}}\times{{\mathcal C}}/{{\mathcal Q}}$ is local [@AdelAlbKor Theorem 5.3].
The implication we have just proved in the previous paragraph cannot be reversed. In order to see it, consider the full subcategory ${{\mathcal C}}$ of ${\operatorname{Mod-\!}}{\mathbb{Z}}$ with the unique object ${\mathbb{Z}}$. Let ${{\mathcal P}}={{\mathcal Q}}=0$ be the zero ideal of ${{\mathcal C}}$, which is a completely prime ideal of ${{\mathcal C}}$. The canonical functor ${{\mathcal C}}\to {{\mathcal C}}/{{\mathcal P}}\times{{\mathcal C}}/{{\mathcal Q}}$ is trivially local. Multiplication by $n\ge 2$ is an endomorphism of ${\mathbb{Z}}$ that is not an automorphism and does not belong to $0={{\mathcal P}}({\mathbb{Z}},{\mathbb{Z}})\cup {{\mathcal Q}}({\mathbb{Z}},{\mathbb{Z}})$.
Applications {#App}
============
Now we are going to apply Theorem \[suff\] to a number of examples.
Biuniform modules. {#bm}
------------------
Let $R$ be a ring and ${{\mathcal B}}$ be the full subcategory of ${\operatorname{Mod-\!}}R$ whose objects are all biuniform right $R$-modules; that is, the modules that are both uniform and couniform (=hollow; biuniform modules are those of Goldie dimension $1$ and dual Goldie dimension $1$). If $A$ and $B$ are biuniform $R$-modules, let ${{\mathcal P}}(A,B)$ be the group of all non-injective morphisms $A\to B$ and ${{\mathcal Q}}(A,B)$ be the group of all non-surjective morphisms $A\to B$. Then ${{\mathcal P}}$ and ${{\mathcal Q}}$ are completely prime ideals of ${{\mathcal B}}$ [@book Lemma 6.26], the category ${{\mathcal B}}$ clearly satisfies Condition [*(DSP)*]{}, and the pair ${{\mathcal P}},{{\mathcal Q}}$ satisfies the hypotheses of Theorems \[WKST\] and \[suff\]. Thus, from Theorem \[suff\], we immediately get that:
\[suff’\] Let $\{\,U_i\mid i\in I\,\}$ and $\{\,V_j\mid j\in J\,\}$ be two families of biuniform modules over an arbitrary ring $R$. Assume that there exist two bijections $\sigma,\tau\colon
I\to J$ such that $[U_i]_m=[V_{\sigma(i)}]_m$ and $[U_i]_e=[V_{\tau(i)}]_e$ for every $i\in I$. Then $\prod_{i\in I}U_i\cong \prod_{j \in J}V_j$.
Since non-zero uniserial modules are biuniform, Theorem \[suff’\] holds in particular for families $\{\,U_i\mid i\in I\,\}$ and $\{\,V_j\mid j\in J\,\}$ of non-zero uniserial modules.
Uniserial modules, quasismall modules.
--------------------------------------
Quasismall modules have a decisive role in the study of direct sums of uniserial modules. Recall that a module $N_R$ is [*quasismall*]{} if for every set $\{\,M_i\mid i\in I\,\}$ of $R$-modules such that $N_R$ is isomorphic to a direct summand of $\oplus_{i\in
I}M_i$, there exists a finite subset $F$ of $ I$ such that $N_R$ is isomorphic to a direct summand of $\oplus_{i\in
F}M_i$. For instance, every finitely generated module is quasismall, every module with local endomorphism ring is quasismall, and every uniserial module is either quasismall or countably generated. There exist uniserial modules that are not quasismall [@Puni].
Pavel Prihoda proved in [@Pri] (the necessity of the condition had already been proved in [@DF]) that:
\[3.2\] Let $\{\,U_i\mid i\in I\,\}$ and $\{\,V_j\mid j\in J\,\}$ be two families of uniserial modules over an arbitrary ring $R$. Let $I'$ be the sets of all indices $i\in I$ with $U_i$ quasismall, and similarly for $J'$. Then $\bigoplus_{i\in I}U_i\cong \bigoplus_{j \in J}V_j$ if and only if there exist a bijection $\sigma\colon
I\to J$ such that $[U_i]_m=[V_{\sigma(i)}]_m$ and a bijection $\tau\colon
I'\to J'$ such that $[U_i]_e=[V_{\tau(i)}]_e$ for every $i\in I'$.
Therefore it is natural to ask whether Theorem \[suff’\] remains true for uniserial modules if we weaken its hypotheses to the condition studied by Prihoda. That is, assume that $\{\,U_i\mid i\in I\,\}$ and $\{\,V_j\mid j\in J\,\}$ are two families of uniserial modules over an arbitrary ring $R$. Let $I'$ be the sets of all indices $i\in I$ with $U_i$ quasismall, and similarly for $J'$. Suppose that there exist a bijection $\sigma\colon
I\to J$ such that $[U_i]_m=[V_{\sigma(i)}]_m$ for every $i\in I$ and a bijection $\tau\colon
I'\to J'$ such that $[U_i]_e=[V_{\tau(i)}]_e$ for every $i\in I'$. Is it true that $\prod_{i\in I}U_i\cong \prod_{j \in J}V_j$?
Equivalently, if $\{\,U_i\mid i\in I\,\}$ and $\{\,V_j\mid j\in J\,\}$ are two families of uniserial modules over an arbitrary ring $R$ and $\bigoplus_{i\in I}U_i\cong \bigoplus_{j \in J}V_j$, is it true that $\prod_{i\in I}U_i\cong \prod_{j \in J}V_j$? We don’t know what the answer to this question is, but, in a very special case, it is possible to find a result dual to Prihoda’s Theorem \[3.2\]. For this purpose, we now recall the main results of [@FG Section 6]. Recall that if $_SA$ and $_SB$ are left modules over a ring $S$, $_SA$ is said to be [*cogenerated by $_SB$*]{} if $_SA$ is isomorphic to a submodule of a direct product of copies of $_SB$.
Let $R$ be any ring. Fix a set $\{\,E_\lambda\mid\lambda\in\Lambda\,\}$ of representatives up to isomorphism of all injective right $R$-modules that are injective envelopes of some non-zero uniserial $R$-module. Set $E_R:=E(\oplus_{\lambda\in\Lambda}E_\lambda)$ and $S:={\mbox{\rm End}}(E_R)$. Then $S/J(S)$ is a von Neumann regular ring and idempotents can be lifted modulo $J(S)$, so that $S$ is an exchange ring [@W72]. Thus idempotents can be lifted modulo every left (respectively right) ideal [@Nicho]. Moreover, ${}_SE_R$ turns out to be an $S$-$R$-bimodule and $${\operatorname{Hom}}(-,{}_SE_R)\colon {\operatorname{Mod-\!}}R\to S{\mbox{\rm -Mod}}$$ is a contravariant exact functor. Let ${{\mathcal C}}_R$ denote the full subcategory of ${\operatorname{Mod-\!}}R$ whose objects are all uniserial right $R$-modules. Let ${}_S{{\mathcal C}}'$ be the full subcategory of $S {\mbox{\rm -Mod}}$ whose objects are all uniserial left $S$-modules that have a projective cover and are cogenerated by $_SE$. It is possible to prove that if a non-zero uniserial module $U$ has a projective cover $P$, then $P$ is a couniform module [@FG Lemma 2.2], so that, in particular, $P$, hence $U$, are cyclic modules. Thus all the $S$-modules in ${}_S{{\mathcal C}}'$ are quasismall. The following result is proved in [@FG Proposition 6.1 and Corollary 6.2].
\[vhl\] [(a)]{} The restriction $H$ of the functor ${\operatorname{Hom}}(-,{}_SE_R)\colon {\operatorname{Mod-\!}}R\to S{\mbox{\rm -Mod}}$ is a duality between the categories ${{\mathcal C}}_R$ and ${}_S{{\mathcal C}}'$.
[(b)]{} Two uniserial right $R$-modules $U_R$ and $U'_R$ belong to the same monogeny class if and only if the uniserial left $S$-modules $H (U_R)$ and $H(U'_R)$ belong to the same epigeny class.
[(c)]{} Two uniserial right $R$-modules $U_R$ and $U'_R$ belong to the same epigeny class if and only if the uniserial left $S$-modules $H(U_R)$ and $H(U'_R)$ belong to the same monogeny class.
Call [*dually quasismall*]{} any object of ${}_S{{\mathcal C}}'$ isomorphic to $H(U_R)$ for some quasismall uniserial $R$-module $U_R$. Then we have that:
\[Pridual\] Let $\{\,U_i\mid i\in I\,\}$, $\{\,V_j\mid j\in J\,\}$ be two families of objects of ${}_S{{\mathcal C}}'$. Let $I'$ be the sets of all the indices $i\in I$ with $U_i$ dually quasismall, and $J'$ be the sets of all the indices $j\in J$ with $V_j$ dually quasismall. Suppose that there exist a bijection $\sigma\colon
I'\to J'$ such that $[U_i]_m=[V_{\sigma(i)}]_m$ for every $i\in I'$ and a bijection $\tau\colon
I\to J$ such that $[U_i]_e=[V_{\tau(i)}]_e$ for every $i\in I$. Then $\prod_{i\in I}U_i\cong \prod_{j \in J}V_j$.
For every $i\in I$, let $X_i$ be a uniserial right $R$-module with $H(X_i)\cong U_i$ and, for every $j\in J$, let $Y_j$ be a uniserial right $R$-module with $H(Y_j)\cong V_j$. The module $X_i$ is quasismall if and only if $i\in I'$, and $Y_j$ is quasismall if and only if $j\in J'$. By Proposition \[vhl\], $[X_i]_e=[Y_{\sigma(i)}]_e$ for every $i\in I'$ and $[X_i]_m=[Y_{\tau(i)}]_m$ for every $i\in I$. From Theorem \[3.2\], we know that $\bigoplus_{i\in I}X_i\cong \bigoplus_{j \in J}Y_j$. Applying the functor ${\operatorname{Hom}}(-,{}_SE_R)$, we obtain the desired conclusion $\prod_{i\in I}U_i\cong \prod_{j \in J}V_j$.
Cyclically presented modules
----------------------------
For any ring $R$, we denote by $U(R)$ the group of all invertible elements of $R$ and by $J(R)$ the Jacobson radical of $R$. Let $R$ be a local ring and ${{\mathcal C}}$ be the full subcategory of ${\operatorname{Mod-\!}}R$ whose objects all the modules $R/rR$ with $r\in J(R)\setminus\{0\}$. Since $R$ is local, all the modules in ${{\mathcal C}}$ are couniform, and therefore all objects of ${{\mathcal C}}$ are indecomposable modules. If $R/rR,R/sR\in\operatorname{Ob}({{\mathcal C}})$, every morphism $R/rR\to R/sR$ is induced by left multiplication by some element $t\in R$, and ${\operatorname{Hom}}(R/rR,R/sR)\cong \{\, t\in R\mid tr\in sR\,\}/sR$. Let ${{\mathcal P}}(R/rR,R/sR)$ be the group $\{\, t\in R\mid tr\in sJ(R)\,\}/sR$. Then ${{\mathcal P}}$ turns out to be a completely prime ideal of ${{\mathcal C}}$. If $R/rR$ and $R/sR$ are objects of ${{\mathcal C}}$ in the same ${{\mathcal P}}$ class, the modules $R/rR$ and $R/sR$ are said to [*have the same lower part*]{}, denoted by $[R/rR]_l=[R/sR]_l$ [@Amini]. It is easily seen that $[R/rR]_l=[R/sR]_l$ if and only if there exist $u,v\in U(R)$ and $x,y\in R$ with $ru= xs$ and $sv= yr$. As in the previous Example \[bm\], let ${{\mathcal Q}}(R/rR,R/sR)$ be the group of all non-surjective morphisms $R/rR\to R/sR$. In this case, ${{\mathcal Q}}(R/rR,R/sR)\cong \{\, t\in J(R)\mid tr\in sR\,\}/sR$.
The pair ${{\mathcal P}},{{\mathcal Q}}$ of completely prime ideals satisfies the hypotheses of Theorems \[WKST\] and \[suff\], and the category ${{\mathcal C}}$ satisfies Condition [*(DSP)*]{}, so that, from Theorem \[suff\], we immediately obtain:
\[suff”\] Let $R$ be a local ring and $\{\,U_i\mid i\in I\,\}$ and $\{\,V_j\mid j\in J\,\}$ be two families of right $R$-modules in ${{\mathcal C}}$. Suppose that there exist two bijections $\sigma,\tau\colon
I\to J$ such that $[U_i]_l=[V_{\sigma(i)}]_l$ and and $[U_i]_e=[V_{\tau(i)}]_e$ for every $i\in I$. Then $\prod_{i\in I}U_i\cong \prod_{j \in J}V_j$.
More generally, lower part and epigeny class can be defined for couniformly presented right modules over a (non-necessarily local) ring. Here an $R$-module $M$ is said to be [*couniformly presented*]{} if there exists an exact sequence $0\to M_1\to P\to M\to 0$ with $P$ projective and $P$ and $M_1$ of dual Goldie dimension $1$; cf. [@FG]. Also in this case, we have an analogue of Theorem \[suff”\] for couniformly presented $R$-modules.
Kernels of morphisms between indecomposable injective modules.
--------------------------------------------------------------
Let $A$ and $B$ be two modules. We say that $A$ and $B$ [*have the same upper part*]{}, and write $[A]_u=[B]_u$, if there exist a homomorphism $\varphi\colon
E(A)\rightarrow E(B)$ and a homomorphism $\psi\colon E(B)\rightarrow E(A)$ such that $\varphi^{-1}(B)=A$ and $\psi^{-1}(A)=B$. If $E_1,E_2,E'_1,E'_2$ are injective indecomposable right modules over an arbitrary ring $R$ and $\varphi\colon E_1\to E_2$, $\varphi'\colon E'_1\to E'_2$ are arbitrary morphisms, then $\ker\varphi \cong \ker\varphi '$ if and only if $[\ker\varphi ]_m=[\ker\varphi ']_m$ and $[\ker\varphi ]_u=[\ker\varphi ']_u$ [@Tufan Lemma 2.4].
Let $R$ be an arbitrary ring and ${{\mathcal K}}$ be the full subcategory of ${\operatorname{Mod-\!}}R$ whose objects are all kernels of morphisms $f\colon E_1\to E_2$, where $E_1$ and $E_2$ range in the class of all uniform injective modules. The canonical functor $P\colon {\operatorname{Mod-\!}}R\to{\operatorname{Spec}}({\operatorname{Mod-\!}}R)$, where ${\operatorname{Spec}}({\operatorname{Mod-\!}}R)$ denotes the spectral category of ${\operatorname{Mod-\!}}R$ [@gabrieloberst], is a left exact, covariant, additive functor, which has an $n$-th right derived functor $P^{(n)}$ for every $n\ge 0$ [@Chennai Proposition 2.2]. The restriction of $P^{(1)}$ to ${{\mathcal K}}$ is a functor ${{\mathcal K}}\to {\operatorname{Spec}}({\operatorname{Mod-\!}}R)$. If ${{\mathcal Q}}(A,B)$ consists of all morphisms $f\colon A\to B$ in ${{\mathcal K}}$ with $P^{(1)}(f)=0$, then ${{\mathcal Q}}$ is a completely prime ideal of ${{\mathcal K}}$. If ${{\mathcal P}}$ is the ideal of all non-injective homomorphisms, as in §\[bm\] for biuniform modules, then the pair ${{\mathcal P}},{{\mathcal Q}}$ satisfies the hypotheses of Theorems \[WKST\] and \[suff\]. In particular, the class ${{\mathcal K}}$ satisfies Condition [*(DSP)*]{}.
Notice that in [@Ece] it was proved that if $\{\,A_i\mid i\in I\,\}$ and $ \{\,B_j\mid j\in J\,\}$ are two families of modules over a ring $R$, all the $B_j$’s are kernels of non-injective morphisms between indecomposable injective modules and there exist bijections $\sigma,\tau\colon I\to J$ such that $[A_i]_m=[B_{\sigma(i)}]_m$ and $[A_i]_u=[B_{\tau(i)}]_u$ for every $i\in I$, then $\oplus_{i\in I}A_i\cong
\oplus_{j \in J}B_j.$ From Theorem \[suff\] we see that, under the same hypotheses, $\prod_{i\in I}A_i\cong
\prod_{j \in J}B_j.$
As a final example for this section, we can consider the following category ${{\mathcal C}}$. Let $R$ be a ring and let $S_1,S_2$ be two fixed non-isomorphic simple right $R$-modules. Let ${{\mathcal C}}$ be the full subcategory of ${\operatorname{Mod-\!}}R$ whose objects are all artinian right $R$-modules $A_R$ with ${\mbox{\rm soc}}(A_R)\cong S_1\oplus S_2$. Set ${{\mathcal P}}_i(A,B):=\{\,f\in{\operatorname{Hom}}_R(A,B)\mid f({\mbox{\rm soc}}_{S_i}(A))=0\,\}$ [@AlbPav5 Example 6.3(7)]. The pair of completely prime ideals ${{\mathcal P}}_1,{{\mathcal P}}_2$ satisfies the hypotheses of Theorem \[suff\].
Reversing the main result
=========================
In this section, we consider the problem of reversng the implications in Theorems \[suff\] and \[3.2\], that is, whether a direct product of uniserial modules determines the monogeny classes and the epigeny classes of the factors. We give four examples that prove that the answer is negative in general.
The following example shows on the one hand that it is impossible to reverse the implication in Theorem \[3.2\] and, on the other hand, that it is impossible to prove a result for direct products analogous to the result proved by Prihoda for direct sums of uniserial modules. More precisely, the example shows that there are two families $\{\,U_i\mid i\in I\,\}$ and $\{\,V_j\mid j\in J\,\}$ of non-zero uniserial quasismall modules over a ring $R$ with no bijection $\sigma\colon
I\to J$ such that $[U_i]_m=[V_{\sigma(i)}]_m$ for every $i\in I$ and no a bijection $\tau\colon
I\to J$ such that $[U_i]_e=[V_{\tau(i)}]_e$ for every $i\in I$, but with $\prod_{i\in I}U_i\cong \prod_{j \in J}V_j$. (Equivalently, $\bigoplus_{i\in I}U_i\not\cong \bigoplus_{j \in J}V_j$, but $\prod_{i\in I}U_i\cong \prod_{j \in J}V_j$.)
In this example, $R$ is the localization of the ring ${\mathbb{Z}}$ of integers at a maximal ideal $(p)$, $I={\mathbb N}$ (the set of non-negative integers), $J={\mathbb N}^*={\mathbb N}\setminus\{0\}$ (the set of positive integers), $U_0$ is the field of fractions ${\mathbb{Q}}$ of $R$ and $U_n=V_n={\mathbb{Z}}(p^\infty)={\mathbb{Q}}/R$ (the Prüfer group) for every $n\ge 1$. Both the $R$-modules ${\mathbb{Q}}$ and ${\mathbb{Z}}(p^\infty)$ are uniserial and with a local endomorphism ring. Hence they are quasismall. Since $[{\mathbb{Q}}]_m\ne [{\mathbb{Z}}(p^\infty)]_m$ and $[Q]_e\ne [{\mathbb{Z}}(p^\infty)]_e$, there are no bijections $\sigma,\tau\colon {\mathbb N}\to{\mathbb N}^*$ preserving the monogeny classes and the epigeny classes, respectively. In order to show that the $R$-modules ${\mathbb{Q}}\oplus({\mathbb{Z}}(p^\infty))^{{\mathbb N}^*}$ and $({\mathbb{Z}}(p^\infty))^{{\mathbb N}^*}$ are isomorphic, it suffices to prove that these two divisible groups are isomorphic. Recall that two divisible abelian groups are isomorphic if and only if they have the same torsion-free rank and the same $p$-rank for every prime $p$. Thus it is enough to show that the torsion-free rank of the divisible abelian group $({\mathbb{Z}}(p^\infty))^{{\mathbb N}^*}$ is infinite, that is, that the group $({\mathbb{Z}}(p^\infty))^{{\mathbb N}^*}$ contains a free abelian subgroup of infinite rank. Now $$({\mathbb{Z}}(p^\infty))^{{\mathbb N}^*}=({\mathbb{Q}}/R)^{{\mathbb N}^*}=\{\,(q_n+R)_{n\ge 1}\mid q_n\in{\mathbb{Q}}\,\}.$$ Consider the infinitely many elements $$\left(\frac{1}{p^{nt}}+R\right)_{n\ge 1}$$ of $({\mathbb{Z}}(p^\infty))^{{\mathbb N}^*}$, with $t\in{\mathbb N}^*$. It is easy to see that these countably many elements form a free set of generators of a free abelian subgroup of $({\mathbb{Z}}(p^\infty))^{{\mathbb N}^*}$. Thus the divisible abelian group $({\mathbb{Z}}(p^\infty))^{{\mathbb N}^*}$ has infinite torsion-free rank; hence the $R$-modules ${\mathbb{Q}}\oplus({\mathbb{Z}}(p^\infty))^{{\mathbb N}^*}$ and $({\mathbb{Z}}(p^\infty))^{{\mathbb N}^*}$ are isomorphic.
Here is another example that proves that it is impossible to reverse the implication in Theorem \[suff\]. Let $R$ be a ring and ${{\mathcal C}}$ be the full subcategory of ${\operatorname{Mod-\!}}R$ whose objects are all injective indecomposable $R$-modules. If $A$ and $B$ are objects of ${{\mathcal C}}$, let ${{\mathcal P}}(A,B)$ be the group of all non-injective morphisms $A\to B$, so that ${{\mathcal P}}$ is a completely prime ideal of ${{\mathcal C}}$, the category ${{\mathcal C}}$ satisfies Condition [*(DSP)*]{}, and the ideals ${{\mathcal P}}={{\mathcal Q}}$ satisfy the hypotheses of Theorems \[WKST\] and \[suff\]. Now take, for instance, $R={\mathbb{Z}}$ and consider the family consisting of all the Prüfer groups ${\mathbb{Z}}(p^\infty)$, where $p$ ranges in the set of prime numbers, and the group ${\mathbb{Q}}$. The groups in this family are pair-wise non-isomorphic and have distinct ${{\mathcal P}}$ classes. In order to show that Theorem \[suff\] cannot be reversed, it suffices to prove that ${\mathbb{Q}}\oplus\prod_p{\mathbb{Z}}(p^\infty)\cong \prod_p{\mathbb{Z}}(p^\infty)$, because there does not exist a bijection $\sigma$ preserving the ${{\mathcal P}}$ classes. And, as in the previous example, to prove that ${\mathbb{Q}}\oplus\prod_p{\mathbb{Z}}(p^\infty)\cong \prod_p{\mathbb{Z}}(p^\infty)$, it is enough to show that the divisible abelian group $\prod_p{\mathbb{Z}}(p^\infty)$ has infinite torsion-free rank. Now the torsion subgroup of $\prod_p{\mathbb{Z}}(p^\infty)$ is the countable divisible group $\bigoplus_p{\mathbb{Z}}(p^\infty)$, so that $\prod_p{\mathbb{Z}}(p^\infty)/\bigoplus_p{\mathbb{Z}}(p^\infty)$ is a torsion-free divisible abelian group of cardinality $2^{\aleph_0}$. Thus $\prod_p{\mathbb{Z}}(p^\infty)/\bigoplus_p{\mathbb{Z}}(p^\infty)$ is a divisible group of torsion-free rank $2^{\aleph_0}$. Hence the divisible group $\prod_p{\mathbb{Z}}(p^\infty)$ has torsion-free rank $2^{\aleph_0}$. Therefore ${\mathbb{Q}}\oplus\prod_p{\mathbb{Z}}(p^\infty)\cong \prod_p{\mathbb{Z}}(p^\infty)$, though the ${{\mathcal P}}$ class of ${\mathbb{Q}}$ does not appear in the set of the ${{\mathcal P}}$ classes of the ${\mathbb{Z}}(p^\infty)$’s. Notice that $\prod_p{\mathbb{Z}}(p^\infty)\cong{\mathbb{T}}:={\mathbb{R}}/{\mathbb{Z}}$.
This argument can be extended to any non-semilocal commutative Dedekind domain $R$ with cardinality $|R|=\alpha$, with maximal spectrum (set of maximal ideals) of cardinality $\beta$ for which $\alpha<\alpha^\beta$. Let $R$ be such a ring. As finite domains are fields, hence semilocal rings, it follows that $\alpha\ge\aleph_0$. Let $Q$ be the field of fractions of $R$, which also must have cardinality $\alpha$. Since all ideals in a Dedekind domain can be generated with two elements, we get that $\aleph_0\le\beta\le\alpha$. In a Dedekind domain, divisible modules coincide with injective modules. For every maximal ideal $P$ of $R$, we have that $|E(R/P)|=|Q/R_P|=\alpha$, where $R_P$ denotes the localization of $R$ at $P$. Thus $|\prod_P Q/R_P|=\alpha^\beta$. Let us prove that the torsion submodule of $\prod_P Q/R_P$ is $\bigoplus_PQ/R_P$. Let $$(q^{(P)}+R_P)_P\in \prod_P Q/R_P$$ be a torsion element. Then there exists a non-zero $r\in R$ such that $rq^{(P)}\in R_P$ for every maximal ideal $P$. Thus $Rrq^{(P)}\subseteq R_P$ for every $P$. The non-zero ideal $Rr$ of $R$ is contained in only finitely many maximal ideals of $R$, so that $R_Pr=R_P$ for almost all maximal ideals $P$. Thus $Rrq^{(P)}\subseteq R_P$ implies that $R_Prq^{(P)}\subseteq R_P$, so $R_Pq^{(P)}\subseteq R_P$, that is, $q^{(P)}\in R_P$ for almost all $P$. This proves that the torsion submodule of $\prod_P Q/R_P$ is $\bigoplus_PQ/R_P$, which has cardinality $\alpha<\alpha^\beta$. Therefore it is possible to argue as in the previous paragraph.
[This example is taken from [@Franetic Example 2.1]. Let $p$ be a prime number and $\widehat{{\mathbb{Z}}_p}$ be the ring of $p$-adic integers, so that ${\mathbb{Z}}/p^n{\mathbb{Z}}$ is a module over $\widehat{{\mathbb{Z}}_p}$ for every integer $n\ge 1$. Let $$\varphi\colon\prod_{n\ge1}{\mathbb{Z}}/p^n{\mathbb{Z}}\to \prod_{n\ge1}{\mathbb{Z}}/p^n{\mathbb{Z}}$$ be the $\widehat{{\mathbb{Z}}_p}$-module morphism defined by $\varphi(a_n+n{\mathbb{Z}})_{n\ge1}=(a_{n+1}-a_n+n{\mathbb{Z}})_{n\ge1}$. This morphism $\varphi$ is onto and its kernel is isomorphic to $\widehat{{\mathbb{Z}}_p}$. Thus there is an exact sequence $\xymatrix{
0 \ar[r] & \widehat{{\mathbb{Z}}_p}\ar[r] & \prod_{n\ge1}{\mathbb{Z}}/p^n{\mathbb{Z}}\ar[r]^{\varphi} & \prod_{n\ge1}{\mathbb{Z}}/p^n{\mathbb{Z}}\ar[r] & 0,
}
$ which is a pure-exact sequence, and $\widehat{{\mathbb{Z}}_p}$ is pure-injective, so that the pure-exact sequence splits. Thus $ \widehat{{\mathbb{Z}}_p}\oplus\prod_{n\ge1}{\mathbb{Z}}/p^n{\mathbb{Z}}\cong \prod_{n\ge1}{\mathbb{Z}}/p^n{\mathbb{Z}}$. In these direct products, all factors $\widehat{{\mathbb{Z}}_p}$ and ${\mathbb{Z}}/p^n{\mathbb{Z}}$ ($n\ge1$) are pair-wise non-isomorphic uniserial $\widehat{{\mathbb{Z}}_p}$-modules, have distinct monogeny classes and distinct epigeny classes. Hence there cannot be bijections $\sigma$ and $\tau$ preserving the monogeny and the epigeny classes in the two direct-product decompositions. Notice that all factors have a local endomorphism ring; hence they are quasismall uniserial modules. ]{}
[Here is a further example that shows that the condition in Theorem \[suff’\] is sufficient but not necessary for the isomorphism $\prod_{i\in I}U_i\cong \prod_{j \in J}V_j$ to hold. Let $U_0$ be a uniserial non-quasismall $R$-module (it is known that such modules exist [@Puni]). Then $U_0$ is countably generated [@DF Lemma 4.2]; hence it is a union of an ascending chain $U_n$, $n\ge 1$ of cyclic submodules. Then $\oplus_{n\ge0}U_n\cong\oplus_{n\ge1}U_n$ [@DF Theorem 4.9]. Since $U_n$ is cyclic for $n\ge 1$ but not for $n=0$, it follows that $[U_0]_e\ne [U_n]_e$ for every $n\ge 1$. Hence there does not exist a bijection between the epigeny classes of $\{\,U_n\mid n\ge 0\,\}$ and the epigeny classes of $\{\,U_n\mid n\ge 1\,\}$. Anyway, there does exist a bijection between the monogeny classes of $\{\,U_n\mid n\ge 0\,\}$ and the monogeny classes of $\{\,U_n\mid n\ge 1\,\}$ (Theorem \[3.2\]). Applying the duality $H$ of Proposition \[vhl\], we obtain two isomorphic direct-product decompositions $\prod _{n\ge0}H(U_n)\cong\prod _{n\ge1}H(U_n)$, with all the modules $H(U_n)$, $n\ge0$, cyclic uniserial left $S$-modules, for which there exists a bijection between the epigeny classes of $\{\,H(U_n)\mid n\ge 0\,\}$ and the epigeny classes of $\{\,H(U_n)\mid n\ge 1\,\}$, but there does not exist a bijection between the monogeny classes of $\{\,H(U_n)\mid n\ge 0\,\}$ and the monogeny classes of $\{\,H(U_n)\mid n\ge 1\,\}$. Hence the condition in Theorem \[suff’\] is not necessary. Notice that by Theorem \[Pridual\], there is a bijection between the monogeny classes of the dually quasismall modules, that is, the modules $H(U_n)$ with $n\ge1$.]{}
Slender modules.
================
Now we adopt the point of view of [@Franetic], restricting our attention to slender modules. Let $R$ be a ring and $R^\omega=\prod_{n<\omega}e_nR$ be the right $R$-module that is the direct product of countably many copies of the right $R$-module $R_R$, where $e_n$ is the element of $R^\omega$ with support $\{n\}$ and equal to $1$ in $n$. A right $R$-module $M_R$ is [*slender*]{} if, for every homomorphism $f\colon R^\omega\to M$ there exists $n_0<\omega$ such that $f(e_n)=0$ for all $n\ge n_0$. The most important property of slender modules we need in the sequel is the following [@EM Theorem 1.2]: A module $M_R$ is slender if and only if for every countable family $\{\,P_n\mid n\ge 0\,\}$ of right $R$-modules and any homomorphism $f\colon \prod_{n\ge 0}P_n\to M_R$ there exists $m\ge0$ such that $f(\prod_{n\ge m}P_n)=0$. Here $\prod_{n\ge m}P_n$ is the subgroup of $\prod_{n\ge 0}P_n$ consisting of all elements with support contained in $\{m,m+1,m+2,\dots\}$. In the following, the cardinality of any set $I$ is denoted by $|I|$. If $M_R$ is slender and $\{\,P_i\mid i\in I\,\}$ is a family of right $R$-modules with $|I|$ non-measurable, then ${\operatorname{Hom}}(\prod_{i\in I}P_i,M_R)\cong\bigoplus_{i\in I}{\operatorname{Hom}}(P_i,M_R)$.
Every submodule of a slender module is a slender module [@EM Lemma 1.6(i)], so that:
If $U_R$ is a slender module, then every module in the same monogeny class as $U_R$ is slender.
As far as Condition [*(DSP)*]{} is concerned, it is easily seen that:
If $U_R, V_R$ are slender modules, then every direct summand of $U_R\oplus V_R$ is slender.
\[nec\] Let ${{\mathcal C}}$ be a full subcategory of ${\operatorname{Mod-\!}}R$ in which all objects are indecomposable slender right $R$-modules and let ${{\mathcal P}},{{\mathcal Q}}$ be a pair of completely prime ideals of ${{\mathcal C}}$ with the property that, for every $A\in{}\operatorname{Ob}({{\mathcal C}})$, $f\colon A\to A$ is an automorphism if and only if $f\notin {{\mathcal P}}(A,A)\cup {{\mathcal Q}}(A,A)$. Assume that ${{\mathcal C}}$ satisfies Condition (DSP). Let $\{\,A_i\mid i\in I\,\}$ and $\{\,B_j\mid j\in J\,\}$ be two families of objects of ${{\mathcal C}}$ with $|I|$ and $|J|$ non-measurable. Assume that:
[(a)]{} In both families, there are at most countably many modules in each ${{\mathcal P}}$ class.
[(b)]{} In both families, there are at most countably many modules in each ${{\mathcal Q}}$ class.
[(c)]{} The $R$-modules $\prod_{i\in I}A_i$ and $\prod_{j \in J}B_j$ are isomorphic.
Then there exist two bijections $\sigma,\tau\colon
I\to J$ such that $[A_i]_{{{\mathcal P}}} = [B_{\sigma(i)}]_{{{\mathcal P}}}$ and $[A_i]_{{{\mathcal Q}}} = [B_{\tau(i)}]_{{{\mathcal Q}}}$ for every $i\in I$.
[*Step 1. Assume that a slender $R$-module $B$ is isomorphic to a direct summand of the direct product $\prod_{i\in I}A_i$, with $|I|$ non-measurable. Then there is a finite subset $F$ of $I$ such that $B$ is isomorphic to a direct summand of $\bigoplus_{i\in F}A_i$*]{}.
This is [@Franetic Lemma 1.1].
[*Step 2. For every $j\in J$ there exist $i,k\in I$ such that $[A_i]_{{{\mathcal P}}}=[B_j]_{{{\mathcal P}}}$ and $[A_k]_{{{\mathcal Q}}}=[B_j]_{{{\mathcal Q}}}$.*]{}
Fix $j\in J$. Since $\prod_{i\in I}A_i\cong\prod_{j \in J}B_j$, we know that $B_j$ is isomorphic to a direct summand of $\prod_{i\in I}A_i$. By Step 1, there exists a finite subset $F=\{i_1,\dots,i_t\}$ of $I$ such that $B_j$ is isomorphic to a direct summand of $\bigoplus_{i\in F}A_{i}=\bigoplus_{\ell=1}^tA_{i_\ell}$. Thus there are morphisms $\varphi\colon B_j\to \bigoplus_{\ell=1}^tA_{i_\ell}$ and $\psi\colon \bigoplus_{\ell=1}^tA_{i_\ell}\to B_j$ with $\psi\varphi=1_{B_j}$. In matrix notation, we have that $$\varphi=\left(\begin{array}{c}\varphi_1\\ \vdots \\ \varphi_t \end{array}\right)\qquad\mbox{\rm and}\qquad\psi=\left(\begin{array}{ccc}\psi_1, & \dots & ,\psi_t\end{array}\right)$$ for suitable morphisms $\varphi_\ell\colon B_j\to A_{i_\ell}$ and $\psi_\ell\colon A_{i_\ell}\to B_j$. Thus $\sum_{\ell=1}^t\psi_\ell\varphi_\ell=1_{B_j}$. In particular, $\sum_{\ell=1}^t\psi_\ell\varphi_\ell\notin{{\mathcal P}}(B_j,B_j)$. Since ${{\mathcal P}}(B_j,B_j)$ is an ideal of the ring ${\mbox{\rm End}}_R(B_j)$, it follows that there exists an index $\ell=1,\dots,t$ such that $\psi_\ell\varphi_\ell\notin{{\mathcal P}}(B_j,B_j)$. Thus $\psi_\ell\notin{{\mathcal P}}(A_{i_\ell},B_j)$ and $\varphi_\ell\notin{{\mathcal P}}(B_j,A_{i_\ell})$. Hence $[A_{i_\ell}]_{{{\mathcal P}}}=[B_j]_{{{\mathcal P}}}$, and the index $i:=i_\ell\in I$ has the required property. By symmetry, the same holds for ${{\mathcal Q}}$ classes.
[*Step 3. Proof of the statement of the Theorem.*]{}
For every slender right $R$-module $D$, set $I_D:=\{\,i\in I\mid [A_i]_{{{\mathcal P}}}=[D]_{{{\mathcal P}}}\,\}$ and $J_D:=\{\,j\in J\mid [B_j]_{{{\mathcal P}}}=[D]_{{{\mathcal P}}}\,\}$. It suffices to show that $|I_D|=|J_D|$ for every such slender right module $D$. By contradiction, suppose that $|I_D|\ne|J_D|$ for some slender module $D$. By symmetry, we can suppose without loss of generality that $|I_D|<|J_D|$. By (a), the cardinal $|I_D|$ must be finite. Also, $J_D$ must be non-empty, so that there exists $j\in J_D$. By Step 2, there exist $i,k\in I$ such that $[A_i]_{{{\mathcal P}}}=[B_j]_{{{\mathcal P}}}$ and $[A_k]_{{{\mathcal Q}}}=[B_j]_{{{\mathcal Q}}}$. If $i=k$, then $A_i\cong B_j$, and since $A_i\cong B_j$ has a semilocal endomorphism ring, it can be canceled from direct sums [@book Corollary 4.6], so that $\prod_{i'\in I}A_{i'}\cong\prod_{j" \in J}B_{j'}$ implies $\prod_{i'\in I\setminus\{i\}}A_{i'}\cong\prod_{j" \in J\setminus\{j\}}B_{j'}$. If $i\ne k$, then $B_j\oplus X_j\cong A_i\oplus A_k$ for some object $X_j$ of ${{\mathcal C}}$ (Lemma \[2.2\]). Then $\prod_{i'\in I}A_{i'}\cong \prod_{j' \in J}B_{j'}$ can be rewritten as $B_j\oplus X_j\oplus\prod_{i'\in I\setminus\{i,k\}}A_{i'}\cong \prod_{j' \in J}B_{j'}$, and canceling as before we get that $X_j\oplus\prod_{i'\in I\setminus\{i,k\}}A_{i'}\cong \prod_{j' \in J\setminus\{j\}}B_{j'}$. In both cases $i=k$ and $i\ne k$, we have obtained two direct-product decompositions in which the families $I_D$ and $J_D$ of ${{\mathcal P}}$ classes have one element less. We now proceed recursively, after $|I_D|$ steps we obtain two direct-product decompositions in which the family $I_D$ is empty and the family $J_D$ is not empty. This contradicts Step 2, and the contradiction proves that the bijection $\sigma$ with the required property exists. It is a similar argument for $\tau$.
\[nec’\] Let ${{\mathcal C}}$ be a full subcategory of ${\operatorname{Mod-\!}}R$ in which all objects are indecomposable slender right $R$-modules and let ${{\mathcal P}},{{\mathcal Q}}$ be a pair of completely prime ideals of ${{\mathcal C}}$ with the property that, for every $A\in{}\operatorname{Ob}({{\mathcal C}})$, $f\colon A\to A$ is an automorphism if and only if $f\notin {{\mathcal P}}(A,A)\cup {{\mathcal Q}}(A,A)$. Assume that ${{\mathcal C}}$ satisfies Condition (DSP). Let $\{\,A_i\mid i\in I\,\}$ and $\{\,B_j\mid j\in J\,\}$ be two countable families of objects of ${{\mathcal C}}$. Assume that $\prod_{i\in I}A_i\cong\prod_{j \in J}B_j$.Then there exist two bijections $\sigma,\tau\colon
I\to J$ such that $[A_i]_{{{\mathcal P}}} = [B_{\sigma(i)}]_{{{\mathcal P}}}$ and $[A_i]_{{{\mathcal Q}}} = [B_{\tau(i)}]_{{{\mathcal Q}}}$ for every $i\in I$.
If ${{\mathcal D}}$ is a preadditive category, $A$ is an object of ${{\mathcal D}}$ and $I$ is an ideal of the ring ${\mbox{\rm End}}_{{{\mathcal D}}}(A)$, let ${{\mathcal A}}_I$ be the ideal of the category ${{\mathcal D}}$ defined in the following way. A morphism $f \colon X \to Y$ in ${{\mathcal D}}$ belongs to ${{\mathcal A}}_I(X,Y)$ if and only if $\beta f \alpha \in I$ for every pair of morphisms $\alpha \colon
A \to X$ and $\beta \colon Y \to A$ in the category ${{\mathcal D}}$. The ideal ${{\mathcal A}}_I$ is called the [*ideal of ${{\mathcal D}}$ associated to*]{} $I$ [@AlbPav3]. The ideal ${{\mathcal A}}_I$ is the greatest of the ideals ${{\mathcal Q}}$ of ${{\mathcal D}}$ with ${{\mathcal Q}}(A,A)\subseteq I$. It is easily seen that ${{\mathcal A}}_I(A,A) = I$.
\[buupbp\] Let ${{\mathcal C}}$ be a full subcategory of ${\operatorname{Mod-\!}}R$ in which all objects are slender right $R$-modules and let ${{\mathcal P}}$ be a completely prime ideal of ${{\mathcal C}}$. Let $\{\,A_i\mid i\in I\,\}$ and $\{\,B_j\mid j\in J\,\}$ be two families of objects of ${{\mathcal C}}$ with $|I|$ and $|J|$ non-measurable. Assume that:
[(a)]{} For every object $A$ of ${{\mathcal C}}$, ${{\mathcal P}}(A,A)$ is a maximal right ideal of ${\mbox{\rm End}}_R(A)$.
[(b)]{} There are at most countably many modules in each ${{\mathcal P}}$ class in both families $\{\,A_i\mid i\in I\,\}$ and $\{\,B_j\mid j\in J\,\}$.
[(c)]{} The $R$-modules $\prod_{i\in I}A_i$ and $\prod_{j \in J}B_j$ are isomorphic.
Then there is a bijection $\sigma_{{{\mathcal P}}}\colon
I\to J$ such that $[A_i]_{{{\mathcal P}}} = [B_{\sigma_{{{\mathcal P}}}(i)}]_{{{\mathcal P}}}$ for every $i\in I$.
Fix an object $D$ in ${{\mathcal C}}$. Let ${{\mathcal A}}_{{{\mathcal P}}(D,D)}$ be the ideal of ${\operatorname{Mod-\!}}R$ associated to the maximal right ideal, hence maximal two-sided ideal, ${{\mathcal P}}(D,D)$ of ${\mbox{\rm End}}_R(D)$. Let $P\colon{\operatorname{Mod-\!}}R\to{\operatorname{Mod-\!}}R/{{\mathcal A}}_{{{\mathcal P}}(D,D)}$ be the canonical functor. The ideal ${{\mathcal A}}_{{{\mathcal P}}(D,D)}$ restricts to an ideal of the category ${{\mathcal C}}$, and we will also denote this restriction by ${{\mathcal A}}_{{{\mathcal P}}(D,D)}$. Similarly, the restriction of $P$ will be still denoted by $P$, so that $P\colon{{\mathcal C}}\to{{\mathcal C}}/{{\mathcal A}}_{{{\mathcal P}}(D,D)}$.
[*Step 1. $D$ is a non-zero object in the factor category ${{\mathcal C}}/{{\mathcal A}}_{{{\mathcal P}}(D,D)}$.*]{}
The endomorphism ring of the object $D$ in the factor category ${{\mathcal C}}/{{\mathcal A}}_{{{\mathcal P}}(D,D)}$ is ${\mbox{\rm End}}_R(D)/{{\mathcal A}}_{{{\mathcal P}}(D,D)}(D,D)={\mbox{\rm End}}_R(D)/{{\mathcal P}}(D,D)$, which is a division ring.
[*Step 2. For every object $A$ of ${{\mathcal C}}$ with $[A]_{{{\mathcal P}}}=[D]_{{{\mathcal P}}}$, the objects $A$ and $D$ are isomorphic objects in the factor category ${{\mathcal C}}/{{\mathcal A}}_{{{\mathcal P}}(D,D)}$.*]{}
Let $A$ be any object of ${{\mathcal C}}$. Suppose $[A]_{{{\mathcal P}}}=[D]_{{{\mathcal P}}}$. Since ${{\mathcal A}}_{{{\mathcal P}}(D,D)}(D,D)={{\mathcal P}}(D,D)$, it follows that ${{\mathcal A}}_{{{\mathcal P}}(D,D)}\supseteq{{\mathcal P}}$. In particular, ${{\mathcal A}}_{{{\mathcal P}}(D,D)}(A,A)\supseteq{{\mathcal P}}(A,A)$, which is a maximal ideal. Thus either ${{\mathcal A}}_{{{\mathcal P}}(D,D)}(A,A)={\mbox{\rm End}}_R(A)$ or ${{\mathcal A}}_{{{\mathcal P}}(D,D)}(A,A)={{\mathcal P}}(A,A).$
In the first case, $1_A\in {{\mathcal A}}_{{{\mathcal P}}(D,D)}(A,A)$, so that for every $\alpha\colon D\to A$ and every $\beta\colon A\to~D$ one has that $\beta\alpha\in {{\mathcal P}}(D,D)$. But $[A]_{{{\mathcal P}}}=[D]_{{{\mathcal P}}}$; hence there exist $\alpha\colon D\to A$ and $\beta\colon A\to D$ with $\alpha\notin {{\mathcal P}}(D,A)$ and $\beta\notin{{\mathcal P}}(A,D)$. This contradicts the fact that ${{\mathcal P}}$ is a completely prime ideal.
It shows that ${{\mathcal A}}_{{{\mathcal P}}(D,D)}(A,A)={{\mathcal P}}(A,A)$, so that the endomorphism ring of $A$ in the category ${{\mathcal C}}/{{\mathcal A}}_{{{\mathcal P}}(D,D)}$ is ${\mbox{\rm End}}_R(A)/{{\mathcal P}}(A,A)$, which is a division ring. From $[A]_{{{\mathcal P}}}=[D]_{{{\mathcal P}}}$, it follows that ${{\mathcal P}}(A,D)\ne{\operatorname{Hom}}(A,D)$ and ${{\mathcal P}}(D,A)\ne{\operatorname{Hom}}(D,A)$. Thus there exist $R$-module morphisms $f\colon A\to D$ and $g\colon D\to A$ with $f\notin {{\mathcal P}}(A,D)$ and $g\notin {{\mathcal P}}(D,A)$. But ${{\mathcal P}}$ is completely prime, so that $fg\notin {{\mathcal P}}(D,D)$ and $gf\notin{{\mathcal P}}(A,A)$. Hence $fg+ {{\mathcal P}}(D,D)$ is an automorphism of $D$ and $gf+{{\mathcal P}}(A,A)$ is an automorphism of $A$ in the factor category ${{\mathcal {{\mathcal }}}} C/{{\mathcal A}}_{{{\mathcal P}}(D,D)}$. Thus the morphism $f+ {{\mathcal A}}_{{{\mathcal P}}(D,D)}(A,D)$ is both right invertible and left invertible in the category ${{\mathcal C}}/{{\mathcal A}}_{{{\mathcal P}}(D,D)}$. Hence it is an isomorphism in ${{\mathcal C}}/{{\mathcal A}}_{{{\mathcal P}}(D,D)}$. It follows that $A$ and $D$ are isomorphic objects of the category ${{\mathcal C}}/{{\mathcal A}}_{{{\mathcal P}}(D,D)}$.
[*Step 3. For every object $A$ of ${{\mathcal C}}$ with $[A]_{{{\mathcal P}}}\ne[D]_{{{\mathcal P}}}$, the object $A$ is a zero object in the factor category ${{\mathcal C}}/{{\mathcal A}}_{{{\mathcal P}}(D,D)}$.*]{}
From $[A]_{{{\mathcal P}}}\ne[D]_{{{\mathcal P}}}$, it follows that ${{\mathcal P}}(A,D)={\operatorname{Hom}}(A,D)$ or ${{\mathcal P}}(D,A)={}$ ${\operatorname{Hom}}(D,A)$. In both cases, for every $\alpha\colon D\to A$ and every $\beta\colon A\to D$, one has that $\beta1_A\alpha\in{{\mathcal P}}(D,D)$. Thus $1_A\in {{\mathcal A}}_{{{\mathcal P}}(D,D)}$. That is, $A=0$ in ${{\mathcal C}}/{{\mathcal A}}_{{{\mathcal P}}(D,D)}$.
[*Step 4. Suppose $F:=\{\,i\in I\mid [A]_{{{\mathcal P}}}=[D]_{{{\mathcal P}}}\,\}$ is a finite set. Then $P(\prod_{i\in I}A_i)$ is the coproduct in ${\operatorname{Mod-\!}}R/{{\mathcal A}}_{{{\mathcal P}}(D,D)}$ of $|F|$ objects whose endomorphism rings are isomorphic to the division ring ${\mbox{\rm End}}_R(D)/{{\mathcal P}}(D,D)$.*]{}
Since the functor $P$ is additive, one has that $$P(\prod_{i\in I}A_i)\cong P(\oplus_{i\in F}A_i)\coprod P(\prod_{i\in I\setminus F}A_i).$$ By Step 2, $P(\oplus_{i\in F}A_i)$ is the coproduct of $|F|$ objects isomorphic to $D$ in $${\operatorname{Mod-\!}}R/{{\mathcal A}}_{{{\mathcal P}}(D,D)}.$$ Moreover the endomorphism ring of $D$ in ${\operatorname{Mod-\!}}R/{{\mathcal A}}_{{{\mathcal P}}(D,D)}$ is isomorphic to the division ring ${\mbox{\rm End}}_R(D)/{{\mathcal P}}(D,D)$. Let us prove that $$P(\prod_{i\in I\setminus F}A_i)=0$$ in ${\operatorname{Mod-\!}}R/{{\mathcal A}}_{{{\mathcal P}}(D,D)}$; that is, $1_{\prod_{i\in I\setminus F}A_i}\in {{\mathcal A}}_{{{\mathcal P}}(D,D)}(\prod_{i\in I\setminus F}A_i,\prod_{i\in I\setminus F}A_i)$. For this, it suffices to show that, for every $\alpha\colon D\to \prod_{i\in I\setminus F}A_i$ and every $\beta\colon \prod_{i\in I\setminus F}A_i\to D$, one has that $\beta\alpha\in {{{\mathcal P}}(D,D)}$. Now $D$ is slender and $|I|$ is non-measurable, so ${\operatorname{Hom}}(\prod_{i\in I\setminus F}A_i, D)\cong \bigoplus_{i\in I\setminus F}{\operatorname{Hom}}(A_i,D)$. Thus if $\pi_j\colon \prod_{i\in I\setminus F}A_i\to A_j$ denotes the canonical projection, then there exists a finite subset $G$ of $I$ disjoint from $F$ and morphisms $\beta_j\colon A_j\to D$ for every $j\in G$ such that $\beta=\sum_{j\in G}\beta_j\pi_j$. As we have seen in Step 3, $A_i$ is a zero object in ${{\mathcal C}}/{{\mathcal A}}_{{{\mathcal P}}(D,D)}$ for every $i\in I\setminus F$. Thus $\beta_j\pi_j\alpha$ is the zero morphism, that is, $\beta\alpha$ is the zero morphism, in ${\operatorname{Mod-\!}}R/{{\mathcal A}}_{{{\mathcal P}}(D,D)}$. We can conclude that $\beta\alpha\in {{{\mathcal P}}(D,D)}$, as desired.
[*Step 5. Suppose $\{\,i\in I\mid [A_i]_{{{\mathcal P}}}=[D]_{{{\mathcal P}}}\,\}$ is an infinite set. Then, for every integer $n\ge1$, $P(\prod_{i\in I}A_i)$ is the coproduct of $n$ objects whose endomorphism rings are isomorphic to the division ring ${\mbox{\rm End}}_R(D)/{{\mathcal P}}(D,D)$ and one more object of ${\operatorname{Mod-\!}}R/{{\mathcal A}}_{{{\mathcal P}}(D,D)}$.*]{}
The proof is the same as the first part of the proof of Step 4.
[*Step 6. Suppose that $\prod_{i\in I}A_i\cong\prod_{j \in J}B_j$. Let $D$ be an object of ${{\mathcal C}}$. Then the set $\{\,i\in I\mid [A]_{{{\mathcal P}}}=[D]_{{{\mathcal P}}}\,\}$ is finite if and only if $\{\,j\in J\mid [B_j]_{{{\mathcal P}}}=[D]_{{{\mathcal P}}}\,\}$ is finite. Moreover, in this case, they have the same number of elements.*]{}
Let ${{\mathcal C}}'$ be any additive category in which idempotents split that contains ${{\mathcal C}}$, contains $P(\prod_{i\in I\setminus F}A_i)$ for every finite subset $F$ of $I$, and contains $P(\prod_{j \in J\setminus G}B_j)$ for every finite subset $G$ of $J$ [@Indiana p. 676]. Now apply the Krull-Schmidt-Azumaya Theorem for additive categories [@Bass p. 20] to the category ${{\mathcal C}}'$ and to the object $P(\prod_{i\in I}A_i)\cong P(\prod_{j \in J}B_j)$. Since the endomorphism rings of the non-zero objects $P(A_i)\cong P(B_j)\cong P(D)$ in ${{\mathcal C}}'$ are division rings, hence local rings, $P(D)^n$ cannot have a direct summand isomorphic to $P(D)^m$ for every $m>n$ by the Krull-Schmidt-Azumaya Theorem. Thus $\{\,i\in I\mid [A]_{{{\mathcal P}}}=[D]_{{{\mathcal P}}}\,\}$ is finite if and only if $\{\,j\in J\mid [B_j]_{{{\mathcal P}}}=[D]_{{{\mathcal P}}}\,\}$ is finite. Moreover, $P(D)^n\cong P(D)^m$ implies $n=m$. Thus the sets $\{\,i\in I\mid [A]_{{{\mathcal P}}}=[D]_{{{\mathcal P}}}\,\}$ and $\{\,j\in J\mid [B_j]_{{{\mathcal P}}}=[D]_{{{\mathcal P}}}\,\}$ are equipotent when they are finite.
In order to conclude the proof of the theorem, it is now sufficient to remark that $\{\,i\in I\mid [A]_{{{\mathcal P}}}=[D]_{{{\mathcal P}}}\,\}$ and $\{\,j\in J\mid [B_j]_{{{\mathcal P}}}=[D]_{{{\mathcal P}}}\,\}$ are always equipotent by hypothesis (b). Now glue the bijections between the sets $\{\,i\in I\mid [A]_{{{\mathcal P}}}=[D]_{{{\mathcal P}}}\,\}$ and $\{\,j\in J\mid [B_j]_{{{\mathcal P}}}=[D]_{{{\mathcal P}}}\,\}$ to obtain a bijection $\sigma_{{{\mathcal P}}}$ that preserves the ${{\mathcal P}}$ classes.
As an application of the previous theorem, we get the following Corollary, which is Theorem 2.8 in [@Franetic].
[[@Franetic Theorem 2.8]]{} Let $R$ be a ring and $\{\,A_i\mid i\in I\,\}$ be a family of slender right $R$-modules with local endomorphism rings. Let $\{\,B_j\mid j\in J\,\}$ be a family of indecomposable slender right $R$-modules. Assume that:
[(a)]{} $|I|$ and $|J|$ are non-measurable cardinals.
[(b)]{} There are at most countably many mutually isomorphic modules in each of the two families $\{\,A_i\mid i\in I\,\}$ and $\{\,B_j\mid j\in J\,\}$.
[(c)]{} The $R$-modules $\prod_{i\in I}A_i$ and $\prod_{j \in J}B_j$ are isomorphic.
Then there exists a bijection $\sigma\colon
I\to J$ such that $A_i\cong B_{\sigma(i)}$ for every $i\in I$.
First of all, let us prove that all the modules $B_j$ also have local endomorphism rings. From (c), each $B_j$ is isomorphic to a direct summand of $\prod_{i\in I}A_i$. As in Step 1 of the proof of Theorem \[nec\] [@Franetic Lemma 1.1], for each $j\in J$ there exists a finite subset $F_j$ of $I$ such that $B_j$ is a isomorphic to direct summand of $\prod_{i\in F_j}A_i$. Thus $B_j\oplus C_j\cong \prod_{i\in F_j}A_i$, where each $C_j$ is a direct sum of finitely many indecomposable objects, and these indecomposable objects plus $B_j$ are isomorphic, up to a permutation, to the modules $A_i$ with $ i\in F_j$ (Krull-Schmidt-Azumaya Theorem [@Bass p. 20]). Thus the modules $B_j$ have local endomorphism rings, and now the role of the two families $\{\,A_i\mid i\in I\,\}$ and $\{\,B_j\mid j\in J\,\}$ has become symmetric.
For every pair $A,B$ of objects of ${{\mathcal C}}$, let ${{\mathcal P}}(A,B)$ be the set of all the homomorphisms $A\to B$ that are not isomorphisms. It is easily checked that ${{\mathcal P}}$ is a completely prime ideal of ${{\mathcal C}}$. Moreover, two objects $A,B$ of ${{\mathcal C}}$ are isomorphic modules if and only if $[A]_{{{\mathcal P}}}=[B]_{{{\mathcal P}}}$. The corollary now follows immediately from Theorem \[buupbp\].
We conclude the paper with two elementary examples of applications of Theorem \[buupbp\]. As a first example, let $R$ be the ring ${\mathbb{Z}}$ of integers (it could be any other countable principal ideal domain that is not a field). Let ${{\mathcal C}}$ be the full subcategory of ${\operatorname{Mod-\!}}{\mathbb{Z}}$ whose objects are all torsion-free ${\mathbb{Z}}$-modules $G$ of torsion-free rank $1$ such that $pG\ne G$ for every prime $p$. There are $2^{\aleph_0}$ pair-wise non-isomorphic such ${\mathbb{Z}}$-modules (recall that torsion-free abelian groups $G$ of torsion-free rank $1$ are completely determined up to isomorphism by their type ${{\mathbf t}}(G)$ [@Fuchs2 Section 85]). All the modules in ${{\mathcal C}}$ are slender modules [@EM Corollary III.2.3]. For each prime $p$ in ${\mathbb{Z}}$ and pair $G,H$ of objects of ${{\mathcal C}}$, set ${{\mathcal P}}_p(G,H)=p{\operatorname{Hom}}(G,H)$. The group ${\operatorname{Hom}}(G,H)$ is torsion-free of rank $1$ and type ${{\mathbf t}}(H): {{\mathbf t}}(G)$ if ${{\mathbf t}}(G)\leq{{\mathbf t}}(H)$ [@Fuchs2 Proposition 85.4], so ${{\mathcal P}}_p(G,H)<{\operatorname{Hom}}(G,H)$ in this case. Otherwise, if ${{\mathbf t}}(G)\nleq{{\mathbf t}}(H)$, then $0={{\mathcal P}}_p(G,H)={\operatorname{Hom}}(G,H)$. It is then very easy to prove that ${{\mathcal P}}_p$ is a completely prime ideal of ${{\mathcal C}}$. For every object $A$ of ${{\mathcal C}}$, ${{\mathcal P}}_p(A,A)$ is the maximal ideal of ${\mbox{\rm End}}_R(A)\cong{\mathbb{Z}}$ generated by $p$. Thus Theorem \[buupbp\] applies to this situation. Notice that, for every prime $p$ and objects $G,H$ of ${{\mathcal C}}$, one has that $[G]_{{{\mathcal P}}_p} = [H]_{{{\mathcal P}}_p}$ if and only if $G\cong H$, because if $[G]_{{{\mathcal P}}_p} = [H]_{{{\mathcal P}}_p}$, then ${{\mathcal P}}_p(G,H)<{\operatorname{Hom}}(G,H)$ and ${{\mathcal P}}_p(H,G)<{\operatorname{Hom}}(H,G)$, so that ${{\mathbf t}}(G)\leq{{\mathbf t}}(H)$ and ${{\mathbf t}}(H)\leq{{\mathbf t}}(G)$; that is, ${{\mathbf t}}(G)={{\mathbf t}}(H)$, and $G\cong H$.
Here is a second example. Recall that a [*rigid system*]{} of abelian groups is a set $\{\,A_i\mid i\in I\,\}$ of non-zero torsion-free abelian groups for which ${\operatorname{Hom}}(A_i,A_j)$ is isomorphic to a subgroup of ${\mathbb{Q}}$ if $i=j$, and is $0$ if $i\ne j$. It is known that there exist rigid systems of abelian groups of finite rank which are homogeneous of type $(0,0,0,\dots)$ [@Fuchs2 Theorem 88.4]. Such groups are torsion-free, reduced and countable, hence slender [@EM Corollary 2.3]. Let ${{\mathcal C}}$ be the full subcategory of ${\operatorname{Mod-\!}}{\mathbb{Z}}$ with class of objects a rigid system of groups of finite rank homogeneous of type $(0,0,0,\dots)$. Let $p$ be any prime number and let ${{\mathcal P}}$ be the completely prime ideal of ${{\mathcal C}}$ defined, for every $A,B\in\operatorname{Ob}({{\mathcal C}})$, by ${{\mathcal P}}(A,B)=p{\operatorname{Hom}}(A,B)$. Then the hypotheses of Theorem \[buupbp\] hold. In this case, $[A]_{{{\mathcal P}}} = [B]_{{{\mathcal P}}}$ if and only if $A=B$, for every $A,B\in \operatorname{Ob}({{\mathcal C}})$. Notice that, in these last two examples, the endomorphism rings of the indecomposable direct factors are not local rings.
[99]{}
A. Alahmadi and A. Facchini, [*Some remarks on categories of modules modulo morphisms with essential kernel or superfluous image,*]{} J. Korean Math. Soc. [**50**]{} (2013), 557–578.
B. Amini, A. Amini and A. Facchini, [*Equivalence of diagonal matrices over local rings,*]{} J. Algebra [**320**]{} (2008), 1288–1310.
H. Bass, “Algebraic K-Theory”, W. A. Benjamin Inc., New York, 1968.
N. V. Dung and A. Facchini, [*Weak Krull-Schmidt for infinite direct sums of uniserial modules,*]{} J. Algebra [**193**]{} (1997), 102–121.
Ş. Ecevit, A. Facchini and M. T. Koşan, [*Direct sums of infinitely many kernels,*]{} J. Australian Math. Soc. [**89**]{} (2010), 199–214.
P. C. Eklof and A. H. Mekler, [*Almost free modules. Set-theoretic methods,*]{} Revised edition. North-Holland Mathematical Library [**65**]{}, North-Holland Publishing Co., Amsterdam, 2002.
A. Facchini, [*Krull-Schmidt fails for serial modules*]{}, Trans. Amer. Math. Soc. [**348**]{} (1996), 4561–4575.
A. Facchini, “Module Theory. Endomorphism rings and direct sum decompositions in some classes of modules”, Progress in Mathematics [**167**]{}, Birkhäuser Verlag, Basel, 1998. Reprinted in Modern Birkhäuser Classics, Birkhäuser Verlag, Basel, 2010.
A. Facchini, [*Representations of additive categories and direct-sum decompositions of objects,*]{} Indiana Univ. Math. J. [**56**]{} (2007), 659–680.
A. Facchini, [*Injective modules, spectral categories, and applications*]{}, in“Noncommutative Rings, Group Rings, Diagram Algebras and Their Applications”, S. K. Jain and S. Parvathi Eds., Contemporary Math. 456, Amer. Math. Soc., 2008, pp. 1–17.
A. Facchini, [*Direct-sum decompositions of modules with semilocal endomorphism rings,*]{} Bull. Math. Sci. [**2**]{} (2012), 225–279.
A. Facchini, Ş. Ecevit and M. T. Koşan, [*Kernels of morphisms between indecomposable injective modules,*]{} Glasgow Math. J. [**52A**]{} (2010), 69–82.
A. Facchini and N. Girardi, [*Couniformly presented modules and dualities*]{}, in “Advances in Ring Theory”, Dinh Van Huynh and Sergio R. López Permouth Eds., Birkhäuser Verlag, Basel, 2010, pp. 149–164.
A. Facchini and P. Příhoda, [*Factor categories and infinite direct sums*]{}, Int. Electron. J. Algebra [**5**]{} (2009), 135–168.
A. Facchini and P. Příhoda, [*The Krull-Schmidt Theorem in the case two*]{}, Algebr. Represent. Theory [**14**]{} (2011), 545–570.
D. Franetic, [*A Krull-Schmidt theorem for infinite products of modules*]{}, to appear in J. Algebra, 2014.
L. Fuchs, “Infinite abelian groups”, vol. II, Academic Press, 1973.
P. Gabriel and U. Oberst, [*Spektralkategorien und reguläre Ringe im von-Neumannschen Sinn*]{}, Math. Z. [**92**]{} (1966), 389–395.
W. K. Nicholson, [*Lifting idempotents and exchange rings,*]{} Trans. Amer. Math. Soc. 229 (1977), 269–278.
P. Příhoda, [*A version of the weak Krull-Schmidt theorem for infinite direct sums of uniserial modules,*]{} Comm. Algebra [**34**]{} (2006), no. 4, 1479–1487.
G. Puninski, [*Some model theory over an exceptional uniserial ring and decompositions of serial modules,*]{} J. London Math. Soc. (2) [**64**]{} (2001), 311–326.
R. B. Warfield, Jr., [*Exchange rings and decompositions of modules,*]{} Math. Ann. [**199**]{} (1972), 31–36.
R. B. Warfield, Jr., [*Serial rings and finitely presented modules*]{}, J. Algebra [**37**]{} (1975), 187–222.
[^1]: Partially supported by Università di Padova (Progetto ex 60% “Anelli e categorie di moduli”) and Fondazione Cassa di Risparmio di Padova e Rovigo (Progetto di Eccellenza “Algebraic structures and their applications”.)
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
author:
- 'Florian Brown-Altvater'
- Gabriel Antonius
- Tonatiuh Rangel
- Matteo Giantomassi
- Claudia Draxl
- Xavier Gonze
- 'Steven G. Louie'
- 'Jeffrey B. Neaton'
bibliography:
- 'main.bib'
title: 'Band gap renormalization, carrier mobilities, and the electron-phonon self-energy in crystalline naphthalene'
---
Lattice parameters
==================
------------------------------------ ------- ------- ------- --------- ----------
$a$ $b$ $c$ $\beta$ $\Omega$
PBE-D3 ([$\Omega_\mathrm{DFT}$]{}) 8.077 5.900 8.620 124.35 339.14
Exp. 8.080 5.933 8.632 124.65 340.41
Exp. ([$\Omega_\mathrm{295K}$]{}) 8.256 5.983 8.677 122.73 360.56
------------------------------------ ------- ------- ------- --------- ----------
: \[SI:tab:lattices\] Lattice parameters used in this work. Experimental crystal structures are available at the Cambridge Structural Database [@Thomas2010; @*csd_url]. The identifiers for the structures measured at and are NAPHTA31 and NAPHTA36, respectively, and published in association with [@Capelli2006]. (Lengths in , angles in degrees, volumes $\Omega$ in .)
Phonon frequencies
==================
![\[SI:fig:phonons\_h8\] Phonon band structure of intermolecular bands of perdeuterated naphthalene calculated with PBE-D3 at [$\Omega_\mathrm{DFT}$]{}. Experimental neutron scattering frequencies are taken from Ref. [@Natkaniec1980]. ](q464_phonons_d8.pdf)
![\[SI:fig:phonons\_d8\] Phonon band structure of naphthalene calculated with PBE-D3 at [$\Omega_\mathrm{DFT}$]{}. Experimental Raman and IR frequencies are taken from Ref. [@Suzuki1968]. The discrepancy of the highest phonon frequencies above coming from C-H stretch modes likely is caused by anharmonic effects, which are ignored by the perturbative approach used here. Because these phonon modes contribute only very little to the electron-phonon coupling, the discrepancy of about of these frequencies will not affect our results in any significant way. ](q464_phonons_h8.pdf)
![\[SI:fig:phonons\_dos\] Comparison of phonon band structures calculated with lattice parameters relaxed with PBE-D3 ([$\Omega_\mathrm{DFT}$]{}) and with fixed experimental room temperature lattice parameters ([$\Omega_\mathrm{295K}$]{}). Both calculations are done with PBE-D3. The lattice parameters mainly affect the soft intermolecular modes below . ](q464_phonons_0K_vs_RT.pdf)
Convergence of self-energy
==========================
![\[SI:fig:conv\_nq\] Convergence w.r.t. ${\mathbf{q}}$-grid spacing of the electron-phonon self-energy of the highest valence and lowest conduction band at $\Gamma$. The four ${{\mathbf{q}}}$ grids used from right to left are ([$2\times4\times2$]{}), ([$4\times6\times4$]{}), ([$6\times8\times6$]{}), and ([$12\times14\times12$]{}). Interpolated grids use square markers. Notable is the much smoother convergence of the conduction band, which is the absolute minimum at $\Gamma$. The valence band at $\Gamma$ is below the valence band maximum. ](convergence_Nq.pdf)
![\[SI:fig:conv\_eta\] Convergence w.r.t. the smearing value $\eta$ of the electron-phonon self-energy of the highest valence and lowest conduction band at $\Gamma$. Interpolated grids use square markers and are labeled as “fine@coarse” to indicate the ${{\mathbf{q}}}$-grids used. Notable is the much smoother convergence of the conduction band, which is the absolute minimum at $\Gamma$. The valence band at $\Gamma$ is below the valence band maximum. ](convergence_eta.pdf)
**k**-independence of self-energy {#SI:sec:kindependence}
=================================
![\[SI:fig:se\_LUMO\] (Left) Electronic band structure of the two lowest conduction bands along $\Gamma\rightarrow\mathrm{Z}$ in the BZ. Each circle indicates an electronic state $n{{\mathbf{k}}}$ for which we explicitly calculated the self-energy $\Upsigma^\mathrm{ep}_{n{{\mathbf{k}}}}$. (Middle) The density of states in the middle is plotted to highlight the correlation to the self-energy as discussed in the main text. (Right) The real and imaginary part of the electron-phonon self-energy of naphthalene for the states circled in the band structure plot on the left. The color of each self-energy matches the corresponding circle. ](sigma_k-independence.pdf){width=".99\textwidth"}
Here we test the ${{\mathbf{k}}}$-independence approximation of the electron-phonon self-energy used in the main text. In \[SI:fig:se\_LUMO\] we compare the frequency-dependent self-energy along the $\Gamma\rightarrow\mathrm{Z}$ path, the direction of the largest dispersion for the conduction bands. Even along this relatively dispersive ${{\mathbf{k}}}$-path, the real and imaginary parts of the self-energy show very little variation, validating the ${{\mathbf{k}}}$-independence approximation. Additionally we can also see that the frequency-dependent self-energy varies very little across the two bands. The combined ${{\mathbf{k}}}$- and $n$-independence leads to the excellent agreement between the mobilities calculated with \[eq:muboltz\] and \[eq:muapprox\].
As laid out in the main text, the two highest valence (lowest conduction) band wavefunctions of the naphthalene crystal, which has two molecules per unit cell, resemble linear combinations of the HOMO (LUMO) of a single naphthalene molecule in the gas phase. This “duality” or pairing of wavefunctions is also called Davydov splitting, or Davydov pairs [@Davydov1964; @Sheka1975]. The electronic states of these Davydov pairs around the band gap (for example HOMO/HOMO-1, HOMO-2/HOMO-3, or LUMO/LUMO+1) interact and mix only very weakly with their neighboring Davydov pairs, since they are energetically far enough apart. Therefore, the resulting wave functions retain their gas-phase-like character throughout the Brillouin zone. In summary, the two wave functions $n$ and $n+1$ of a Davydov pair have a similar molecular orbital character, and are largely independent of ${{\mathbf{k}}}$.
Since the electron-phonon matrix elements are a measure of the overlap of the electronic wave function with the first derivative of the phonon potential, we expect the same $n$/$n+1$ and ${{\mathbf{k}}}$-independence for the matrix elements. Additionally, the separation between Davydov pairs near the band gap is on the order of , and thus at the upper end of our phonon spectrum. Hence, significant contributions to the self-energy, i.e., terms with small denominators in \[eq:Fan\], will only come from scattering within these Davydov pairs. These two factors (${{\mathbf{k}}}$- and $n$-independence of the matrix elements; and well separated bands) lead to the ${{\mathbf{k}}}$- and $n$-independence of the frequency-dependent self-energy. Since both these properties are very typical for organic crystals, we expect to see this approximation to hold for other systems as well.
Electrical mobility
===================
![ Calculated temperature-dependent hole (top) and electron (bottom) mobilities in comparison with experiment (green dots) [@Madelung2000]. The calculated mobilities at using experimental room temperature ([$\Omega_\mathrm{295K}$]{}) lattice parameters (red squares) agree very well with the experimental values. For reference, we also show the temperature dependence of the mobilities for the relaxed structure ([$\Omega_\mathrm{DFT}$]{}, orange dashed). []{data-label="fig:tdep_mobility"}](mobilities_Tdep_comparison_eh.pdf)
---------------------- -------------------- -------------------- ----------------------- -------------------- -------------------- -----------------------
$\mu^\mathrm{h}_a$ $\mu^\mathrm{h}_b$ $\mu^\mathrm{h}_{c*}$ $\mu^\mathrm{e}_a$ $\mu^\mathrm{e}_b$ $\mu^\mathrm{e}_{c*}$
$\mu_\mathrm{Boltz}$ 1.198 2.735 0.239 0.667 0.313 0.209
$\mu(\varepsilon)$ 1.213 2.769 0.241 0.663 0.323 0.209
relative error 0.013 0.012 0.008 0.005 0.033 0.003
---------------------- -------------------- -------------------- ----------------------- -------------------- -------------------- -----------------------
: \[SI:tab:muapprox\]Mobility values calculated with [$\Omega_\mathrm{295K}$]{} at , comparing the values $\mu_\mathrm{Boltz}$ obtained with the self-energy relaxation time approximation \[eq:muboltz\], with $\mu(\varepsilon)$ from \[eq:muapprox\], where we express the mobility as product of four independent functions of energy $
\mu^\mathrm{e,h}_{\alpha} \propto
\int {\mathrm{d}\varepsilon\,}\, D(\varepsilon) f'(\varepsilon) v_{\alpha}^2(\varepsilon) \tau(\varepsilon)$. Mobility values are given in .
![\[SI:fig:muapprox\] Comparing the integrand of the mobility from the explicit summation over $n{{\mathbf{k}}}$ in the self-energy relaxation time approximation in \[eq:muboltz\] ($\mu_\mathrm{Boltz}$, solid transparent line) with the product of the four independent functions $D(\varepsilon)$, $f'(\varepsilon)$, $v_{\alpha}^2(\varepsilon)$, and $\tau(\varepsilon)$ from \[eq:muapprox\] ($\mu(\varepsilon)$, dashed lines). ](mobility_integrand_eh_percent.pdf)
Eigenvalue self-consistent self-energy
======================================
In \[SI:fig:se\_sf\_sc\] we compare the eigenvalue self-consistent (evSC) with the one-shot approach by looking at the self-energies and spectral functions of the two lowest conduction bands (LUMO and LUMO+1) at $\Gamma$. The differences of the frequency-dependent self-energies between the one-shot and self-consistent calculations are quite significant. The main peaks of the real and imaginary parts are red-shifted by almost 0.2 eV, and a second peak develops at around 3.5 eV, above which the self-energy is mostly unchanged in comparison. This directly correlates with the renormalized density of states (DOS), where the lower energy states are renormalized by 0.1-0.2 eV, and the higher states remain unshifted.
While the frequency-dependent self-energies of the LUMO and LUMO+1 states look very similar, both for one-shot and evSC, the effect of the self-consistent approach on their spectral functions is quite different. The quasiparticle peak of the conduction band minimum is slightly red-shifted, and broadened compared to the one-shot. Additionally, the satellites are more pronounced and also red-shifted. Overall, the qualitative shape of the quasiparticle peak remains the same, however.
For the second conduction band LUMO+1, both the position of the main quasiparticle peak as well as the shape of the spectral function are severely altered. The electronic energy of this state falls into a region, where the evSC electron-phonon self-energy is very non-linear, and thus departs from the textbook quasiparticle picture. The renormalization of the main quasiparticle peak changes from for the one-shot calculation to for evSC. Furthermore, instead of the one broad phonon satellite in the one-shot calculation, the evSC spectral function shows two satellites, one at higher and one at lower energies. At other ${{\mathbf{k}}}$-points, the satellite at actually becomes the main quasiparticle peak (not shown here). A similar transition can be more clearly seen for the valence band in \[fig:sf\_full\_bs\], where about half-way between $\Gamma$ and Z, the quasiparticle peak shifts its weight from to . This transition, or seeming quasiparticle “discontinuity” in the conduction bands is less visible because the bands are much more broadened out.
![ Comparing the self-energyies (top) and spectral functions (bottom) of the one-shot with the eigenvalue self-consistent (evSC) self-energy calculation for the lowest (LUMO) and second lowest (LUMO+1) conduction band at $\Gamma$. The DFT eigenvalues are marked with vertical gray dotted lines. []{data-label="SI:fig:se_sf_sc"}](se_sf_reg_vs_SC.pdf)
![\[SI:fig:sc\_conv\] Convergence of the electronic energies using the eigenvalue self-consistent electron-phonon approach for bands around the band gap. The energy difference at the first step is equal to the renormalization obtained from the on-the-mass-shell approximation. ](convergence_SC.pdf)
![\[SI:fig:on\_off\_SC\] Comparing the renormalizations and lifetimes of the evSC calculation with the on-the-mass-shell approximation ($\Upsigma(\varepsilon^0)$) and the quasiparticle solution ($\Upsigma(\varepsilon)$). Valence bands are at ${{\mathbf{k}}}=\mathrm{A}$, and conduction bands at $\Gamma$, coinciding with the VBM and CBM, respectively. The renormalizations (top) are calculated at using [$\Omega_\mathrm{DFT}$]{} lattice parameters. For the lifetimes (bottom) at we used [$\Omega_\mathrm{295K}$]{} to account for lattice expansion. Overall, the on-the-mass-shell approximation agrees better with the evSC approach. ](comparison_ReIm_sc_on_off.pdf)
![\[SI:fig:mobilities\_decomposition\_SC\] Comparing the energy dependent velocities (\[eq:v2approx\]) and lifetimes (\[eq:tauapprox\]) of the hole (top) and electron (bottom) carriers at ([$\Omega_\mathrm{DFT}$]{}) between the one-shot (solid) and SC (dashed) calculations. The high electron velocities around are an artifact caused by the breakdown of the quasiparticle picture near the band splitting. ](v2_tau_comparison_SC_eh.pdf)
input[ms\_aux.tex]{}
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'It is quite common that several different phases exist simultaneously in a system of trapped quantum gases of ultra-cold atoms. One such example is the strongly-interacting Fermi gas with two imbalanced spin species, which has received a great amount of attention due to the possible presence of exotic superfluid phases. By employing novel numerical techniques and algorithms, we self-consistently solve the Bogoliubov de-Gennes equations, which describe Fermi superfluids in the mean-field framework. From this study, we investigate the novel phases of spin-imbalanced Fermi gases and examine the validity of the local density approximation (LDA), which is often invoked in the extraction of bulk properties from experimental measurements within trapped systems. We show how the validity of the LDA is affected by the trapping geometry, number of atoms and spin imbalance.'
address:
- '$^{1}$Department of Physics and Astronomy and Rice Quantum Institute, Rice University, Houston, TX 77005, USA'
- '$^{2}$Department of Physics, University of Cincinnati, Cincinnati, OH 45221, USA'
author:
- 'L. O. Baksmaty$^{1}$, Hong Lu$^{1}$, C. J. Bolech$^{2,1}$ and Han Pu$^{1}$'
title: 'Bogoliubov-de Gennes study of trapped spin-imbalanced unitary Fermi gases'
---
*Keywords*: Fermi gas, Fulde-Ferrel-Larkin-Ovchinnikov state
Introduction
============
Interest in spin-imbalanced Fermi superfluids dates back over a half century to the appearance of the Bardeen-Cooper-Schrieffer (BCS) theory. Several pioneers theoretically considered the fate of a superconductor in the presence of magnetic impurities which could disrupt the spin balance [@clogston; @chandra; @FF_original; @LO_original] prescribed by BCS. In the ensuing decades the BCS picture has emerged as an important paradigm in many branches of physics [@casalbuoni], such as nuclear physics, quantum chromodynamics, and ultra-cold atoms. The ultra-cold atoms, in particular, provide a highly controllable clean system to investigate the effect of spin imbalance on the nature of fermionic pairing. Over the past few years, experiments on polarized Fermi gases at Rice University [@randy1; @randy2; @randy3], MIT [@mit1; @mit2; @mit3; @mit4] and ENS [@ens] have stimulated a flurry of theoretical activity. These efforts represent an important area within the general goal of emulating complicated many-body systems using cold atoms.
All cold atom experiments are necessarily performed in the presence of trapping potentials that hold the atoms together in an inhomogeneous environment. In order to extract the bulk properties of the system (e.g, the equations of state) from measurement on a trapped sample, a local density approximation (LDA) is employed to account for the effects of the trapping potential. The LDA states that the system can be treated locally as a part of an infinite system. In other words, on a length scale which is short in comparison to the spatial variation of the external potential, the external potential $V_{{\rm ext}}({\bf r})$ essentially acts as an offset to the chemical potential. In practice, one defines a local chemical potential $\mu({\bf r})=\mu-V_{{\rm ext}}({\bf r})$, with $\mu$ being the global chemical potential. In the context of cold atoms, the LDA is often an accurate approximation when there are no phase boundaries present. However, because of the large change in the density which occurs from the center to the edge of the trap, there is often more than one phase present and care needs to be taken in application of the LDA. This is especially true at phase boundaries where the LDA usually fails. Even in the circumstances where there is only one phase present in the trap, corrections to the LDA become crucial when the number of particles in the sample is small and finite-size effects are significant [@son].
The trapped polarized Fermi gas represents just such a system — in all experimental observations, phase separation between a region with vanishing spin population and one with finite spin population has been observed. The purpose of this paper is to examine the validity of the LDA as the particle number and aspect ratio of the trap is changed. To this end we employ the Bogoliubov-de Gennes (BdG) equations [@gennes], which is a powerful mean-field tool particularly suitable for inhomogeneous Fermi superfluids, and has been recently adopted by many to study trapped ultra-cold Fermi gases [@liu3D; @machida; @gri; @torma; @Ueda_tezuka].
In a recent paper [@fflo] we analyzed discrepancies observed in pioneering experiments on polarized fermionic superfluids [@mit1; @mit2; @randy1; @randy2] which appear to center on physics arising when the containing trap is highly elongated. In our analysis we found that as the trap becomes increasingly elongated, the solutions of the BdG show a tendency towards metastable behavior which could lead to the observation of states which are not necessarily the ground state. However we also observed that one class of solutions with structure similar to the LDA solution was consistently the lowest in energy within our analysis. Our conclusions have since been confirmed by similar calculations [@Pei] using a density functional formulation [@bulgac] which is more sophisticated than the BdG and accounts for quantum fluctuations and interactions within the normal fluid. In this paper we focus on *this* class of solutions and examine how well such solutions match those obtained from the LDA approximation as a function of particle number and trap aspect ratio.
The content of the paper is organized as follows. In Sec. \[sec:BdG-formulation-and\], we present the BdG formulation and describe the numerical techniques used to solve the BdG equations. In Sec. \[N200\] we discuss our results for a relatively small number of particles and for different trapping geometry. In Sec. \[Nbig\], we focus on a very elongated cigar-like trap but vary the number of atoms. Finally, a concluding remark is presented in Sec. \[sum\].
The Bogoliubov De-Gennes treatment \[sec:BdG-formulation-and\]
==============================================================
Formulation
-----------
We consider a gas of spin-polarized fermionic atoms interacting through a contact potential ($g\sum_{i<j}\delta^{3}(\vec{r}_{i}-\vec{r}_{j})$) and confined to a harmonic trap defined in cylindrical coordinates $(\rho,\phi,z)$ by $V_{\rm ext}(\rho,z)=\frac{m}{2}(\omega_{\bot}^{2}\rho^{2}+\omega_{z}^{2}z^{2})$ with axial and radial frequencies denoted by ($\omega_{z},\omega_{\bot})$. We work at unitarity $(a_{s}\rightarrow\infty)$ and within a cigar-shaped trap with aspect ratio $\alpha=\omega_{\bot}/\omega_{z}$. This system of $N=N_{\uparrow}+N_{\downarrow}$ atoms interacting through contact interaction is described by a Hamiltonian $\hat{H}=\int d\vec{R} \,(H_{0}+H_{I})$ with non-interacting $H_{0}$ and interacting $H_{I}$ portions given by: $$\begin{aligned}
H_{0}(\vec{R}) &= & \sum_{\sigma}\psi_{\sigma}^{\dagger}({-\frac{\hbar^{2}}{2m}\nabla^{2}+V_{\rm ext}
\left(\rho,z\right)-\mu_{\sigma}})\psi_{\sigma} \,,\nonumber \\
H_{I}(\vec{R}) &= & -g\, \psi_{\uparrow}^{\dagger}(\vec{R})\psi_{\downarrow}^{\dagger}(\vec{R})\psi_{\downarrow}(\vec{R})\psi_{\uparrow}(\vec{R}) \, ,\label{eq:basic_hamiltonian}\end{aligned}$$ where $\psi_{\sigma}(\vec{R})$ represents the fermionic field operators, $m$ the mass and $\mu_{\sigma}$ the chemical potential of atomic species with spin $\sigma$. The coupling constant is defined as $g=\frac{4\pi\hbar^{2}a}{m}$. Henceforth, we work in trap units for which: $m=\omega_{z}=\hbar=k_{B}=1$. This implies that energies will be measured in units of $\hbar\omega_{z}$, lengths in units of $l_{0}=\sqrt{\frac{\hbar}{m\omega_{z}}}$ and temperature ($T$) in units of $\hbar\omega_{z}/k_{B}$ .
The Hamiltonian (\[eq:basic\_hamiltonian\]) will be treated within the mean-field BdG approximation for which there are many excellent references [@gennes; @Castin_bcs_theory; @hu1]. Here we simply state the BdG equations for the pair wave functions $u_{j}(\vec{R})$ and $v_{j}(\vec{R})$ which decouple $\hat{H}$:
$$\left[\begin{array}{cc}
{\cal H}_{\uparrow}^{s}-\mu_{\uparrow} & \Delta(\vec{R})\\
\Delta(\vec{R}) & -{\cal H}_{\downarrow}^{s}+\mu_{\downarrow}\end{array}\right]\left[\begin{array}{c}
u_{j}\\
v_{j}\end{array}\right]=E_{j}\left[\begin{array}{c}
u_{j}\\
v_{j}\end{array}\right],\label{mean_field_ham}$$
In the above coupled set of equations, $u_{j}(\vec{R})$ and $v_{j}(\vec{R})$ are two components of the quasi-particle wavefunction associated with energy $E_{j}$. The single particle Hamiltonian ${\cal H}_{\sigma}^{s}$ is defined in our trap units by:$${\cal H}_{\sigma}^{s}(\vec{R})=-{\nabla^2}/{2}+V_{\rm ext}+g\rho_{\bar{\sigma}}-\mu_{\sigma}\label{eq:diagonal_h}$$ and includes the trapping potential, the chemical potential $\mu_{\sigma}$ and the Hartree mean-field potential is given by density $\rho_{\sigma}(\vec{R})=\langle\Psi_{\sigma}^{\dagger} (\vec{R} )\Psi_{\sigma}(\vec{R} )\rangle$. In accordance with fermionic commutation relations [@gennes], the quasi-particle amplitudes are normalized as:$$\begin{aligned}
\int d\vec{R} \,|u_{j}(\vec{R})|^{2}+|v_{j}(\vec{R})|^{2}=1\label{eq:uv_normalization}\end{aligned}$$ and are related to the spin densities through :$$\begin{aligned}
\rho_{\uparrow}(\vec{R}) & = & \sum_{j=1}^{\infty}|u_j(\vec{R})|^{2}f\:(E_{j})\nonumber \\
\rho_{\downarrow}(\vec{R}) & = & \sum_{j=1}^{\infty}|v_j(\vec{R})|^{2}f\:(E_{j}),\label{eq:hartree_h}\end{aligned}$$ where $f(E)=1/(e^{E/k_{B}T}+1)$ is the Fermi-Dirac distribution function. In the unitary limit the Hartree terms $g\rho_{\bar{\sigma}}$ on the diagonal of Eq. (\[mean\_field\_ham\]) do not actually diverge but are unitarity limited [@gupta]. How to incorporating the Hartree term in the unitarity limit is beyond the mean-field BdG formalism. For this reason we ignore this term in our calculations. The paring field or gap paramter is give by $$\Delta(\vec{R}) = g \langle \psi_\uparrow (\vec{R}) \psi_\downarrow (\vec{R}) \rangle = g \sum_j u_j(\vec{R}) v^*_j(\vec{R}) f(E_j) \label{eq:uv_gap_exp}$$ Since the Hartree terms are ignored in our analysis, Eqs. (\[mean\_field\_ham\]), (\[eq:uv\_normalization\]) and (\[eq:uv\_gap\_exp\]) constitute a closed set of nonlinear equations which we solve self-consitently. However as presented above, our formulation has one problem which arises from a nasty side effect of the contact interaction. The contact interaction assumes wrongly that all states are scattered in the same way regardless of their incoming energy and consequently sums in contributions from collisions at arbitrarily high energy which creates an ultra-violet divergence. Hence the gap equation \[eq:uv\_gap\_exp\] needs to be properly regularized as we now discuss.
Regularizing the BdG equations
------------------------------
Due to the assumption of contact interaction, the gap $\Delta$ is a function of the center-of-mass coordinate of the pair, $\vec{R}$. To discuss the regularization, it is more convenient to re-introduce back the relative coordinate $\vec{r}$, with which the gap is defined as $$\Delta (\vec{R},\vec{r}) = \langle\psi_{\uparrow}(\vec{R}+{\vec{r}}/{2})\psi_\downarrow (\vec{R}-{\vec{r}}/{2} )\rangle$$which diverges as $\frac{\Delta(\vec{R})}{2\pi r}$ when $r \rightarrow 0$ [@Castin_bcs_theory]. To regularize Eq. (\[eq:uv\_gap\_exp\]), one simply subtracts off the $1/r$ divergence to obtain the regularized equation [@Castin_bcs_theory]: $$\frac{\Delta(\vec{R})}{g}=\sum_j \, u_{j}(\vec{R})v_{j}^{*}(\vec{R})\,f(E_{j})-\frac{\Delta}{2\pi r}\label{eq:regularized_gap_equation}$$ In practice, the convergene of the sum above is quite slow and we discuss here a numerically efficient way of evaluating $\Delta(\vec{R})$ to sufficient accuracy without undue effort. First an energy cutoff $E_{c}$ is used to break the sum of Eq. (\[eq:uv\_gap\_exp\]) into two pieces as ff.$$\frac{\Delta(\vec{R})}{g}=\sum_{E_{j}<E_{c}}\, u_{j}(\vec{R})v_{j}^{*}(\vec{R})f\:(E_{j})+\frac{\Delta_{c}(\vec{R})}{g}-\frac{\Delta(\vec{R})}{2\pi r}\label{eq:cutoff_equation}$$ The second term $\Delta_{c}(\vec{R})$ is an approximation to the sum above the cutoff using the LDA result for the pairing field [@hu1] which can also be written as :$$\begin{aligned}
\frac{\Delta_{c}(\vec{R})}{g} & = & \frac{\Delta(\vec{R})}{(2\pi)^{3}}\int_{k_c}^{\infty}\frac{d^{3}k}{\sqrt{(\frac{k^{2}}{2}-\mu(\vec{R}))^{2}+\Delta^{2}}}
\label{eq:lda_unregularized_gap}\end{aligned}$$ where $k_c$ is the momentum cutoff related to $E_c$. This leads to a computationaly efficient form of the gap equation: $$\frac{\Delta(\vec{R})}{U_{\rm eff}(\vec{R})}=\sum_{E_{j}<E_{c}}\, u_{j}(\vec{R})v_{j}^{*}(\vec{R}) \,f(E_{j}).\label{eq:numerical_gap}$$ Here we have employed the identity: $$\int_{0}^{\infty}\frac{d\vec{k}}{(2\pi)^{3}}\frac{e^{ikr}}{\frac{k^{2}}{2}}=\frac{1}{2\pi r}\label{eq:divergence_integral}$$ to subsume the LDA approximation of the gap ($\Delta_{c}$) into an effective interaction defined by:$$\frac{1}{U_{\rm eff}(\vec{R})}=\left[\frac{1}{g}-\int_{k_c}^{\infty}d^{3}k
\left(\frac{1}{\sqrt{(\frac{k^{2}}{2}-\mu(\vec{R}))^{2}+\Delta^{2}}}-\frac{1}{\frac{k^{2}}{2}}\right)\right]
\label{eq:effective_interaction}$$ Below this cutoff $E_{c}$, the quasiparticle states are calculated exactly by solving Eqs. (\[mean\_field\_ham\]), (\[eq:numerical\_gap\]) and (\[eq:effective\_interaction\]) self-consistently along with the normalization conditions:$$\begin{aligned}
N_{\sigma} & = & \int \, \rho_{\sigma}(\vec{R}) \,d\vec{R}\label{normalization}\end{aligned}$$ which conserve total particle number $N=N_{\uparrow}+N_{\downarrow}$ and overall polarization $P=(N_{\uparrow}-N_{\downarrow})/N$. The iterative solution of these equations is achieved using a modified Broyden’s approach [@Johnson] which is a nonlinear mixing scheme allowing the formation of polarized regions even if they were not present in the initial condition. In this scheme convergence was achieved when the root mean squared difference between $\Delta$ at different iterations was below some tolerance i.e., $\sqrt{\frac{\sum_{j}(\Delta_{i}^{j}-\Delta_{i+1}^{j})^{2}}{\sum\Delta_{i}^{j2}}}<tol$, where $j$ is the position index and $i$ represents the iteration number. We should point out that at unitarity this method is analogous to a descent technique where the step is optimized through the residuals stored from a few previous steps. As mentioned previouly we choose as our initial condition $\Delta_{\rm LDA}$, the LDA solution to the BdG equations. This is an important point becasue we found in previous work [@fflo] that at larger particle numbers the solution of the BdG can be quite sensitive to the initial conditions.
Special features
----------------
We discretize using a linear triangular finite element mesh in the $\rho$-$z$ plane which anticipates that our results will retain the cylindrical symmetry of the confining potential $V_{ext}$. The accuracy of these calculations are controlled by the density of the trianglular mesh and the cut-off $E_{c}$ used in the hybrid scheme. Both of these are changed in successive solutions until the free-energy or relevant observable converges to a sufficient accuracy. Experience has taught us that this simple renormalization scheme typically converges when the cutoff is of the order $6E_{F}$ (where $E_{F}$ is the Fermi energy) which implies that the number of quasiparticle states to be directly calculated by Eq. (\[mean\_field\_ham\]) is about $6N$. Note that this puts a constraint on the density of the discretizing mesh. Thus, for moderate system sizes, we are still presented with a very large problem. For example, for $N\sim10^{3}$ particles, one essentially needs to calculate $\sim 10^{5}$ quasiparticle states at each iteration.
One important consequence of our finite element discretization is that it yields sparse matrices which are suitable to massively parallel matrix computations. This is of key importance given that the slow convergence of the sum in Eq. (\[eq:numerical\_gap\]) condemns us to calculate a very large number of quasi-particle states. This is true inspite of our efficient hybrid scheme, without which calculation would be prohibitive. It is immediately obvious that these difficulties will increase with the number of particles $N$, and will make the problem impractical for even moderate particle numbers without very careful formulation. In our case these difficulties are inescapable since the issues to be addressed occur in the presence of finite size effects and confinement. Hence it was crucial to develop the ability to perform calculations with realistic particle numbers because it is not [*a priori*]{} obvious how physical properties will scale with system size. At each iteration, we need to find a large number of eigenvalues and eigenfunctions for large matrices. To this end, we use a novel shift-and-invert scheme which we developed independently but is very similar to the one described in Ref. [@zhang].
Briefly, the scheme involves partitioning the sought spectrum amongst groups of processors working independently. The size of the group is determined by the minimum number of processors with enough total memory to store the inverted matrix which is required for building the local Krylov basis. The main challenge here is bookkeeping to prevent over-counting of states and balancing the load amongst the processor groups. It is conceivable that this method could have issues in cases where the equations support huge degenerate subspaces. In our particular formulation we exploited cylindrical symmetry and parity along the long trap axis to reduce the problem. Consequently we only had to contend with accidental degeneracies. A good analysis of these issues may be found in Ref. [@zhang] but a more thorough description of the numerical details will be presented elsewhere. However we note here that this parallelization scheme is very efficient on distriubted computing systems and scales easily to thousands of CPU’s which is as high as we have tested. Potentially it can be used to study much larger systems than we have reported here or in [@fflo].
Validity of the BdG
-------------------
We should devote a few lines here to comment on the validity of the BdG theory at unitarity becasue it is expected that quantum fluctuations and other effects due to the strong interactions could be significant in this regime. The main drawbacks of the BdG is that it fails to account for phase fluctations. At unitarity it has an additional disadvantage in that it also fails to account for interactions within the normal fluid which is unitarily limited [@gupta]. However the BdG is widely expected to yield qualitatively reliable answers for two main reasons. First because of the finite size of these experiments, the trapped gas enjoys protection from fluctuations of arbitrarily low energy or of very long wavelengths. Secondly due to experimental evidence for superfluidity at unitarity, it is quite clear that interactions within the normal fluid are not so great that the order parameter cannot form or will be destroyed. Thus the failure to account for these effects is not expected to change the topology of the phase diagram but at most would slightly shift the phase boundaries. Since our purpose to examine the suitability of the LDA is qualitative, we are confident that the BdG can account for the essential physics. Nevertheless, due to the limitations within the BdG formalism, in particular the neglect of interaction in the normal fluid, our calculation fails to quantitatively locate position of the Clogston limit. In addition, it cannot be applied to study a system with extremely large polarization (i.e., $P \approx 1$) where the polaron physics will dominate [@ens; @pol].
Results for $N=200$ {#N200}
===================
In this section, we focus on a relatively small particle number of $N=200$. As we will show, the system is rather sensitive to the trap geometry. In the following, we first briefly discuss the case of a spherical trap with $\alpha=1$ and then concentrate on elongated cigar-like trapping potentials with $\alpha>1$ and then discuss them in detail.
Spherical trap
--------------
Liu [*et al.*]{} studied a spin-imbalanced Fermi gas confined in a spherical harmonic trap in Ref. [@liu3D]. To benchmark our work, we first did a series of calculations for this geometry and found our results in perfect agreement with those reported in Ref. [@liu3D]. In this case, even though we anticipate only cylindrical geometry, the density profiles always obey the spherical symmetry. Note that the authors of Ref. [@liu3D] solved the one-dimensional (1D) radial equation, hence the spherical symmetry of the cloud is automatically imposed. We refer the readers to Ref. [@liu3D] for details; here we give just a brief description of the key features. The density profiles indicate a phase-separation scenario: a fully paired BCS superfluid core at the trap center surrounded by a fully polarized shell composed of excess majority spins. A thin layer of partially polarized gas forms the interface between the superfluid core and the normal shell. In this intermediate regime, the minority density and the order parameter sharply drop to zero. Here and in other cases, we always found that the profile of the order parameter closely follows that of the density of the minority spin component. Furthermore, in this case, the LDA gives very good agreement with the full BdG calculation even for particle numbers as small as a few hundred.
Cigar trap
----------
![\[ar5c\] Density profile of the atomic clound for $N=200$ in an elongated trap with aspect ratio $\alpha=5$. The upper, middle and lower row are results from different polarizations $P=0.2$, 0.4 and 0.6, respectively. In each row, we have shown (from left to right) the column densities of the majority component $\int dx\,\rho_{\uparrow}$, the minority component $\int dx\,\rho_{\downarrow}$, their difference $\int dx\,(\rho_\uparrow-\rho_{\downarrow})$, and the axial spin density $n_{1D}(z)=\int dxdy\,(\rho_{\uparrow}-\rho_{\downarrow})$.](ar5column){width="9cm"}
So far, all the experiments on spin-imbalanced Fermi gases have been performed in cigar-like traps with aspect ratio $\alpha>1$. For a given particle number, the 1D regime is eventually encountered in this geometry by increasing $\alpha$ and creates the possibility to study the 3D-1D dimensional crossover. Figure \[ar5c\] illustrates several examples of the density profiles for $N=200$ atoms confined in a moderately elongated trap with $\alpha=5$ (this trap aspect ratio is close to what has been used in the MIT experiments). We find it convenient to express our results in terms of the Fermi energy $E_{\textit{F}}=(3N)^{1/3}\alpha^{2/3}$, central number density $(2E_{F})^{3/2}/(6\pi^{2})$, and the Thomas-Fermi radius along the $z$-axis $Z_{\textit{F}}=\sqrt{2E_{\textit{F}}}$ for a single species ideal Fermi gas of $N/2$ particles in a trap with identical parameters. The upper row of Fig. \[ar5c\] shows the density profiles of a system with a relatively small polarization $P=0.2$. Here the axial spin density $n_{1D}(z)$ exhibits a double-horn structure and vanishes near $z=0$. This is a clear violation of the LDA which predicts that $n_{1D}$ should be flat topped [@LDAe]. Fig. \[ar5c\] can be examined in tandem with Fig. \[fig2\] where for a closer inspection, we plot the densities and the order parameter along the axial and radial axis for two different polarizations. Fig. \[fig2\](a) displays results for $P=0.2$. The density profiles along the $z$-axis show clearly a phase separated three-region structure — moving from the center to the edge of the trap, we encounter a fully paired superfluid core, a partially paired intermediate region and a fully polarized normal gas, just like in the previous case of spherical trap. In stark contrast, the density profiles for the two components along the $\rho$-axis are completely overlapped. In fact, this matching of the radial profiles occur for $|z|\le0.1$. As a consequence, the axial spin density vanishes near $z=0$ as shown in the upper row of Fig. \[ar5c\].
![Density and order parameter profiles along the axial and radial axes in a cigar-like trap with $\alpha=5$ for two differnet polarizations: (a) $P=0.2$ and (b) $P=0.6$. []{data-label="fig2"}](AR5){width="10cm"}
That the majority and minority densities overlap along the radial direction can be understood from an argument invoking the surface energy. When induced phase separation occurs, there is an accompanying surface energy associated with the interface between the two phases. The system will then try to minimize the interface in order to reduce the associated energy. For a cigar-like trap as we study here, the superfluid-normal gas interface area can be efficiently reduced if the two spin components match their densities radially. The authors of Ref. [@STmueller; @STStoof] devised phenomenological theories to include the surface term variationally to explain the breakdown of the LDA observed in the Rice experiment [@randy1; @randy2]. In our calculation, the surface energy is automatically included from the self-consistent BdG formulation [@adilet].
As polarization increases, eventually it becomes energetically unfavorable to have this radial overlap. This is illustrated in Fig. \[fig2\](b) for $P=0.6$. Consequently, the axial spin density no longer vanishes near $z=0$ and the LDA becomes more accurate (see the middle and bottom rows of Fig. \[ar5c\]). In addition, it is quite noticeable that, particularly for large $P$, the minority component density has a steeper down turn along the axial axis than along the radial axis. Moreover, in the partially polarized intermediate region, the order parameter has a small oscillation along the axial axis, but not along the radial axis. Similar order parameter oscillations were also found in the spherical trap case [@liu3D]. This is a consequence of the proximity effect which, in the context of superconductor, occurs when a superconductor is in contact with a normal metal, the Cooper pairs from the superconductor diffuse into the normal component.
![\[ar50c\] Same as in Fig. \[ar5c\], but for $\alpha=50$.](ar50column){width="10cm"}
Next, we keep $N$ fixed at 200 but increase the trap aspect ratio to $\alpha=50$ which represents a much more elongated cigar trap and close to what is used in the Rice experiment. A similar display of the column and axial spin density profiles for different polarization as in Fig. \[ar5c\] is shown in Fig. \[ar50c\]. In this very elongated trap, the majority and minority components have their densities matched along the radial axis up to the highest polarization we have calculated which is $P=0.7$, and the minority component has a boxy-looking density profile. This further confirms that the system is able to greatly reduce the effective surface area between the normal state and the superfluid state in anisotropic cigar-like traps. Another marked feature for such an elongated trap is the prominent oscillations of the order parameter along the $z$-axis. As demonstrated in Fig. \[fig4\], these oscillations are quite generic features in such a trap with finite $P$. As $P$ increases, both the amplitude and the spatial extension of the oscillations increase. As shown in Fig. \[fig4\](b), at large polarizations, the axial length of the partially polarized intermediate region becomes comparable to or even larger than that of the BCS core. Accompanied by the oscillation in the order parameter, the density profiles (in particular, the minority density) also exhibit strong oscillations. Such oscillations are reminiscent of the FFLO pairing state predicted by Fulde, Ferrel, Larkin and Ovchinnikov [@FF_original; @LO_original] in which the Cooper pairs possess finite momentum and the order parameter in the bulk develops sinusoidal oscillations that break the spatial translation symmetry.
![$\alpha=50$ for two different polarizations: (a) $P=0.2$ and (b) $P=0.7$. Same units as in Fig. \[fig2\]. We do not show the radial density and order parameter profiles, which look more or less like those in Fig. \[fig2\](a). (c) The surface plot of the order parameter for the case of $P=0.7$ in the $\rho$-$z$ plane, showing strong oscillations with the nodes aligned along the radial direction. (d) The density oscillations for $P=0.7$ leave a strong signal in the doubly integrated axial spin density $n_{1D}$.[]{data-label="fig4"}](AR50)
Our calculations also show that these axial oscillations are aligned along the radial axis, as shown in Fig. \[fig4\](c). We even intentionally started from an initial ansatz of $\Delta$ where the axial oscillations are present but with the nodes mis-aligned in the radial direction, the BdG iterations eventually converge to a state where the nodes are perfectly aligned radially. This radial alignment has important impact in detecting the oscillations in column density profiles where the densities are integrated along one radial axis: Due to the radial alignment, the oscillations are not washed out and can be easily observed, for example, in the doubly integrated axial spin density $n_{1D}$, as illustrated in Fig. \[fig4\](d).
It is interesting to compare our result with the recent work by Bulgac and Forbes [@bulgac] who, using a density functional theory, argued that the FFLO pairing phase occupies a larger phase space region than people previously thought for a 3D homogeneous system. The FFLO state found in Ref. [@bulgac] is also associated with large-amplitude density oscillations, particularly in the minority component. Another perspective on the order parameter oscillations and its potential connection with the FFLO phase is dimensionality. It is well known that the partially polarized phase with FFLO-like oscillations is prominently featured in the phase space of 1D systems [@hu1; @orso; @hu]. That the reduced dimensionality favors such a state can be understood from the argument of Fermi surface nesting or alternatively from the cost of creating domain walls. The cigar-like traps used in our calculation mimic a quasi-1D system and may be the reason that we see pronounced oscillations in our calculation. If this latter explanation is correct, i.e., the partially polarized region featuring FFLO-like oscillations is due to the effective reduction of the spatial dimension, we then expect to see these oscillations diminish as $N$ is increased while the trap aspect ratio is fixed, which makes the system more 3D-like. To confirm this, we now turn to the next section where we keep $\alpha=50$ but vary the total particle number $N$.
Results for large particle numbers at $\alpha=50$ {#Nbig}
=================================================
![Density and order parameter profiles along the axial and the radial axis in a cigar-like trap with $\alpha=50$ for $P=0.3$ but different values of total particle number $N$. Same units as in Fig. \[fig2\].[]{data-label="fig5"}](differentN){width="10cm"}
We now consider a trapped system in a very elongated cigar trap with $\alpha=50$. Fig. \[fig5\] shows the density and order parameter along the axial and radial axis at a fixed polarization $P=0.3$ but different values of total particle number $N$. As one can clearly see, the oscillations in both the order parameter and the density profiles diminish as $N$ is increased and the LDA approximation becomes more and more accurate, which indicates that the FFLO-like region observed above for small $N$ does not represent a bulk 3D phase. Rather, it is a finite-size effect due to the effective reduction of the spatial dimension.
Nevertheless, we have discovered that as $N$ increases, the system exhibits a tendency towards metastability [@fflo]. Numerically, by starting from different initial ansätze for the order parameter $\Delta(\vec{r})$, the BdG solution may converge to different final states. Our calculations show that among these different states, the one that closely resembles the LDA solution always has the lowest energy as long as $N$ is sufficiently large ($N \gtrsim 10^{4}$), but that there may exist several metastable states with energies just slightly larger that violate the LDA. The observed LDA-violating states at Rice are most likely these metastable states. Experimentally, whether the ground state or a metastable state will be realized may depend upon how the evaporative cooling procedure is implemented [@Parish_transport]. This has been confirmed very recently in a new experiment by Hulet group [@hulet].
Conclusion {#sum}
==========
In conclusion, we have carried out a systematic study of a trapped spin-imbalanced Fermi gas in the unitary limit up to a total number $N\sim10^{5}$ atoms. We study a class of solutions which has recently been identified as having the lowest energy [@fflo; @Pei] through a self-consistently solution of the Bogoliubov-de Gennes equations using state-of-the-art numerical techniques. For a given set of trapping parameters, the LDA will eventually become accurate for sufficiently large $N$. However, the validity of the LDA is also sensitive to the trap geometry: Traps with small anisotropy favor the LDA. Our calculations show that for a relatively small number of atoms in a very elongated trapping potential, the system contains three phases: an unpolarized BCS phase, a partially polarized FFLO-like phase and a normal phase. That the FFLO region exists may be understood from the view point of reduced effective spatial dimension. As $N$ is increased while all other parameters remain fixed, the FFLO-like region eventually disappears. A detailed analysis of dimensional crossover will be reported in the future.
This work was supported by a grant from the ARO with funding from the DARPA OLE program, the Welch foundation (C-1669, C-1681) and NSF. We thank the hospitality of KITP where part of the work is carried out.
[34]{} Clogston A M 1962 [*Phys. Rev. Lett.*]{} **9** 266
Chandrasekhar B S 1962 [*Appl. Phys. Lett.*]{} **1** 7
Fulde P and Ferrell R A 1964 [*Phys. Rev.*]{} **135** A550
Larkin A I and Ovchinnikov Y N 1965 [*Zh. Eksp. Teor. Fiz.*]{} **47** 1136 [\[]{}[*Sov. Phys. JETP*]{} **20** 762 [\]]{}
Casalbuoni R and Nardulli G 2004 [*Rev. Mod. Phys.*]{} **76** 263
Patridge G B [*et al.*]{} 2006 [*Science*]{} **311** 492
Patridge G B [*et al.*]{} 2006 [*Phys. Rev. Lett.*]{} **97** 190407
Liao Y-A [*et al.*]{} 2010 [*Nature*]{} **467** 567
Zwierlein M W [*et al.*]{} 2006 [*Science*]{} **311** 492
Zwierlein M W [*et al.*]{} 2006 [*Nature*]{} **442** 54
Schunck C H [*et al.*]{} 2007 [*Science*]{} **316** 867
Shin Y I [*et al.*]{} 2008 [*Nature*]{} **451** 689
Nascimbéne S [*et al.*]{} 2009 [*Phys. Rev. Lett.*]{} **103** 170402
Son D T 2007 [*preprint*]{} arXiv:0707.1851
De Gennes P 1989, [*Superconductivity of Metals and Alloys*]{}, (Addison-Wesley, Reading, MA)
Liu X-J, Hu H and Drummond P 2007 [*Phys. Rev. A*]{} **75** 023614
Mizushima K, Machida K and Ichioka M 2005 [*Phys. Rev. Lett.*]{} **94**, 060404
Ohashi Y and Griffin A 2005 [*Phys. Rev. A*]{} **72** 013601
Kinnunen J, Jensen L M and Törmä P 2006 [*Phys. Rev. Lett.*]{} **96** 110403
Tezuka M, Yanase Y and Ueda M 2010 [*preprint*]{} [arXiv]{}:0811.1650v3
Baksmaty L O [*et al.*]{} 2011 [*Phys. Rev. A*]{} **83**, 023604 (arXiv:1003.4488)
Pei J C, Dukelsky J and Nazarewicz W 2010 [*Phys. Rev. A*]{} **82**, 021603
Bulgac A and Forbes M M 2008 [*Phys. Rev. Lett.*]{} **101** 215301
G. Bruun, Y. Castin, R. Dum, and K. Burnett, Eur. Phys. J. D 7, 433-439 (1999)
Liu X-J, Hu H and Drummond P 2007 [*Phys. Rev. A*]{} **76** 043605
Gupta S [*et al.*]{} 2003 [*Science*]{} **300** 1723
Johnson D D 1988 [*Phys. Rev. B*]{} [**38**]{} 12807
Zhang H [*et al.*]{} 2007 [*ACM Trans. Math. Software*]{} **33** 9
Schirotzek A [*et al.*]{} 2009 [*Phys. Rev. Lett.*]{} [**102**]{} 230402
De Silva T N and Mueller E J 2006 [*Phys. Rev. A*]{} **73** 051602(R)
De Silva T N and Mueller E J 2006 [*Phys. Rev. Lett.*]{} **97** 070402
Haque M and Stoof H T C 2007 [*Phys. Rev. Lett.*]{} **98** 260406
Imambekov A [*et al.*]{} 2006 [*Phys. Rev. A*]{} 74, 053626
Orso G 2007 [*Phys. Rev. Lett.*]{} **98** 070402
Hu H, Liu X-J and Drummond P 2007 [*Phys. Rev. Lett.*]{} **98** 070403
Parish M M and Huse D 2009 [*Phys. Rev. A*]{} **80** 063605
Hulet R G, private communication.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'In this work, we study asymptotic zero distribution of random multi-variable polynomials which are random linear combinations $\sum_{j}a_jP_j(z)$ with i.i.d coefficients relative to a basis of orthonormal polynomials $\{P_j\}_j$ induced by a multi-circular weight function $Q$ defined on ${\Bbb{C}}^m$ satisfying suitable smoothness and growth conditions. In complex dimension $m\geq3$, we prove that $\Bbb{E}[(\log(1+|a_j|))^m]<\infty$ is a necessary and sufficient condition for normalized zero currents of random polynomials to be almost surely asymptotic to the (deterministic) extremal current $\frac{i}{\pi}\partial\overline{\partial}V_{Q}.$ In addition, in complex dimension one, we consider random linear combinations of orthonormal polynomials with respect to a regular measure in the sense of Stahl & Totik and we prove analogous results in this setting.'
address: 'Faculty of Engineering and Natural Sciences, Sabanc[i]{} University, İstanbul, Turkey'
author:
- Turgay Bayraktar
title: On Global universality for zeros of random polynomials
---
Introduction
============
A random *Kac polynomial* is of the form $$f_n(z)=\sum_{j=0}^na_jz^j$$ where coefficients $a_j$ are independent complex Gaussian random variables of mean zero and variance one. A classical result due to Kac and Hammersley [@Kac; @Ham] asserts that normalized zeros of Kac random polynomials of large degree tend to accumulate on the unit circle $S^1=\{|z|=1\}.$ This ensemble of random polynomials has been extensively studied (see eg. [@LO; @HN; @SV; @IZ] and references therein). Recently, Ibragimov and Zaporozhets [@IZ] proved that for independent and identically distributed (i.i.d.) real or complex random variables $a_j$ $$\label{IZc}
\Bbb{E}[\log(1+|a_j|)]<\infty$$ is a necessary and sufficient condition for zeros of random Kac polynomials to accumulate near the unit circle. In particular, under the condition (\[IZc\]) asymptotic zero distribution of Kac polynomials is independent of the choice of the probability law of random coefficients. We refer to this phenomenon as *global universality* for zeros of Kac polynomials.
In [@SZ3], Shiffman and Zelditch remarked that it was an implicit choice of an inner product that produced the concentration of zeros of Kac polynomials around the unit circle $S^1.$ More generally, for a simply connected domain $\Omega\Subset{\Bbb{C}}$ with real analytic boundary $\partial\Omega$ and a fixed orthonormal basis (ONB) $\{P_j\}_{j=1}^{n+1}$ induced by a measure $\rho(z)|dz|$ where $\rho\in\mathcal{C}^{\omega}(\partial \Omega)$ and $|dz|$ denote arc-length, Shiffman and Zelditch proved that zeros of random polynomials $$f_n(z)=\sum_{j=1}^{n+1}a_jP_j(z)\ \text{where}\ a_j \ \text{i.i.d standard complex Gaussians}$$ concentrate near the boundary $\partial\Omega$ as $n\to \infty.$ Furthermore, the empirical measures of zeros $$\frac{1}{n}\sum_{\{z:f_n(z)=0\}}\delta_z$$ converge weakly to the equilibrium measure $\mu_{\overline{\Omega}}$. Recall that for a non-polar compact set $K\subset {\Bbb{C}}$ the *equilibrium measure* $\mu_K$ is the unique minimizer of the *logarithmic energy* functional $$\nu\to \int\int\log\frac{1}{|z-w|}d\nu(z)d\nu(w)$$ over all probability measures supported on $K.$ Later, Bloom [@Bloom2] observed that $\overline{\Omega}$ can be replaced by a regular compact set $K\subset {\Bbb{C}},$ the inner product can be defined in terms of any Bernstein Markov measure (see also [@BloomS] for a generalization of this result to ${\Bbb{C}}^m$ for Gaussian random pluricomplex polynomials). More recently, Pritsker and Ramachandran [@PrK] observed that (\[IZc\]) is a necessary and sufficient condition for zeros of random linear combinations of Szegö, Bergman, or Faber polynomials (associated with Jordan domains bounded with analytic curves) to accumulate near the support of the corresponding equilibrium measure.\
The purpose of this work is to study global universality for normalized zero currents of random multi-variable complex polynomials. Asymptotic zero distribution of multivariate random polynomials has been studied by several authors (see eg. [@SZ; @DS3; @BloomS; @BloomL; @B6; @B7; @B4]). We remark that randomization of the space of polynomials in these papers is different than that of [@IZ; @KZ; @PrK]. Namely, in the former ones each $\mathcal{P}_n$ are endowed with a $d_n:=dim(\mathcal{P}_n)$ fold product probability measure which leads to a sequence of polynomials (with $n^{th}$ coordinate has total degree at most $n$) chosen independently at random according to the $d_n$-fold product measure. On the other hand, the papers [@IZ; @KZ; @PrK] fix a random sequence of scalars for which one considers random linear combinations of a fixed basis for $\mathcal{P}_n$. We adopt the approach of [@IZ; @KZ; @PrK] in the present note.
The setting is as follows: let $Q:{\Bbb{C}}^m\to \Bbb{R}$ be a *weight function* satisfying $$\label{growth}
Q(z)\geq (1+\epsilon)\log\|z\|\ \text{for}\ \|z\|\gg 1$$ for some fixed $\epsilon>0.$ Throughout this note (unless otherwise stated), we assume that the function $Q:{\Bbb{C}}^m\to [0,\infty)$ is of class $\mathscr{C}^2$ and it is invariant under the action of the real torus $\Bbb{S}^m,$ the latter means that $$\label{circular}
Q(z_1,\dots,z_m)=Q(|z_1|,\dots,|z_m|) \ \text{for all} \ (z_1,\dots,z_m)\in \Bbb{C}^m.$$ One can define an associated weighted extremal function $$V_Q(z):=\sup\{u(z):u\in\mathcal{L}({\Bbb{C}}^m), u\leq Q\ \text{on}\ {\Bbb{C}}^m\}$$ where $\mathcal{L}({\Bbb{C}}^m)$ denotes the *Lelong class* of pluri-subharmonic (psh) functions $u$ that satisfies $u(z)-\log^+\|z\|=O(1).$ We also denote by $$\mathcal{L}^+({\Bbb{C}}^m):=\{u\in \mathcal{L}({\Bbb{C}}^m): u(z)\geq \log^+\|z\|+C_u\ \text{for some}\ C_u\in \Bbb{R}\}.$$ Seminal results of Siciak and Zaharyuta (see [@Klimek] and references therein) imply that $V_Q\in\mathcal{L}^+({\Bbb{C}}^m)$ and that $V_{Q}$ verifies $$\label{envelope}
V_Q(z)=\sup\{\frac{1}{\deg p}\log|p(z)|:p\ \text{is a polynomial and}\ \max_{z\in {\Bbb{C}}^m}|p(z)|e^{-deg(p)Q(z)}\leq 1\}.$$ Moreover, a result of Berman [@Berman1 Proposition 2.1] implies that $V_Q$ is of class $\mathscr{C}^{1,1}$.
Next, we define an inner product on the space $\mathcal{P}_n$ of multi-variable polynomials of degree at most $n$ by setting $$\label{n}
\langle f_n,g_n\rangle_n:=\int_{{\Bbb{C}}^m}f_n(z)\overline{g_n(z)}e^{-2nQ(z)}dV_m(z)$$ where $dV_m$ denotes the Lebesgue measure on ${\Bbb{C}}^m$. We also let $\{P_j^n\}_{j=1}^{d_n}$ be the orthonormal basis (ONB) for $\mathcal{P}_n$ obtained by applying Gram-Schmidt algorithm in the Hilbert space $(\mathcal{P}_n, {\langle}\cdot,\rangle_n)$ to the monomials $\{z^J\}_{|J|\leq n}$ where $J=(j_1,\dots,j_m)$ is m-multiindex and we assume that the monomials $\{z^J\}_{|J|\leq n}$ are ordered with respect to lexicographical ordering. Note that since $Q$ is $m-$circular we have $P_j^n(z)=c_J^nz^J$ for some deterministic constant $c^n_J$ and $J\in\Bbb{N}^m.$
Let $a_1,a_2,\dots$ be a sequence of i.i.d. real or complex random variables whose probability law denoted by ${{\bf{P}}}$. Throughout this note, we assume that $a_j$ are non-degenerate, roughly speaking this means that ${{\bf{P}}}[a_j=z]<1$ for every $z\in{\Bbb{C}}$ (see §\[probprelim\].) A *random polynomial* is of the form $$f_n(z)=\sum_{j=1}^{d_n}a_jP_j^n(z)$$ where $d_n:=\dim(\mathcal{P}_n)={n+m\choose n}$. We also let $\mathcal{H}:=\cup_{n=1}^{\infty}\mathcal{P}_n$ and denote the corresponding probability space of polynomials by $(\mathcal{H},{\Bbb{P}}).$
\[main 1\] Let $a_j$ be i.i.d. non-degenerate real or complex random variables satisfying $$\label{A}
\Bbb{E}[\big(\log(1+|a_j|)\big)^m]<\infty.$$ If the dimension of complex Euclidean space $m \geq 3$ then almost surely in $\mathcal{H}$ $$\frac{1}{n}\log|f_n(z)| \xrightarrow[n \to \infty]{} V_{Q}(z)$$ in $L^1_{loc}({\Bbb{C}}^{m}).$ In particular, almost surely in $\mathcal{H}$ $$\frac{i}{\pi}\partial\overline{\partial}(\frac{1}{n}\log|f_n(z)|) \xrightarrow{} \frac{i}{\pi}\partial\overline{\partial}V_{Q}(z)$$ in the sense of currents as $n\to \infty.$
Furthermore, for all dimensions $m\geq 1,$ we have convergence in probability $$\frac{i}{\pi}\partial\overline{\partial}(\frac{1}{n}\log|f_n(z)|) \xrightarrow[]{} \frac{i}{\pi}\partial\overline{\partial}V_{Q}(z)$$ in the sense of currents as $n\to \infty.$
Note that Theorem \[main 1\] provides an optimal condition on random coefficients for a random version of Siciak-Zaharyuta theorem in this context (cf. [@Bloom1; @B6; @BloomL; @B7]). In the univariate case we have $\frac{i}{\pi}\partial\overline{\partial}=\frac{1}{2\pi}\Delta$ where $\Delta$ denotes the Laplacian and we denote the corresponding *equilibrium measure* by $\mu_Q:=\frac{i}{\pi}\partial\overline{\partial}V_Q.$ An important example is $Q(z)=\frac{|z|^2}{2}$ and $\mu_{Q}=\frac{1}{\pi}1_{\Bbb{D}}dz$ where $\Bbb{D}$ denotes closed the unit disc in the complex plane [@SaffTotik pp 245]. Then a routine calculation shows that $$P^{n}_j(z)=\sqrt{\frac{n^j}{2\pi j!}}z^j\ \text{for}\ j=0,1,\dots,n$$ form an ONB for $\mathcal{P}_n.$ A random *Weyl polynomial* is of the form $$W_n(z)=\sum_{j=0}^na_j\sqrt{\frac{n^j}{ j!}}z^j.$$ In particular, Theorem \[main 1\] generalizes a special case of [@KZ Theorem 2.5] to the several complex variables.
Let us denote the Euclidean volume in ${\Bbb{C}}^{m}$ by $Vol_{2m}$ and for an open set $U\subset {\Bbb{C}}^m,$ we define $$\mathcal{V}_{U}:=\frac{1}{(m-1)!}\int_{U} \frac{i}{\pi}\partial\overline{\partial}V_{Q}\wedge (\frac{i}{\pi}\partial\overline{\partial}\|z\|^2)^{m-1}.$$ Next result indicates that in higher dimensions the condition $(\ref{A})$ is also necessary for zero divisors of random polynomials to be almost surely equidistributed with the extremal current $ \frac{i}{\pi}\partial\overline{\partial}V_{Q}$.
\[main\] Let $a_j$ be i.i.d. non-degenerate real or complex valued random variables and assume that the dimension of complex Euclidean space $m \geq 3.$ The logarithmic moment $$\label{A1}
\Bbb{E}[\big(\log(1+|a_j|)\big)^m]<\infty$$ if and only if $$\label{2}
{\Bbb{P}}\Big\{\{f_n\}_{n\geq0}: \lim_{n\to \infty}\frac1nVol_{2m-2}(Z_{f_n}\cap U)=\mathcal{V}_{U}\Big\}=1$$ for every open set $U\Subset ({\Bbb{C}}^*)^m$ such that $\partial U$ has zero Lebesgue measure.
Note that when $m=1$ the volume $Vol_{2m-2}(Z_{f_n}\cap U)$ becomes the number of zeros of $f_n$ in $U$ which we denote by $$\mathcal{N}_n(U,f_n):=\#\{z\in U:f_n(z)=0\}.$$ The following result is an immediate consequence of Theorem \[main 1\] together with Theorem \[main\] and provides a weak universality result for zeros of univariate random polynomials:
\[cor\] Let $a_j$ be i.i.d. non-degenerate real or complex valued random variables. If the logarithmic moment $$\Bbb{E}[\log(1+|a_j|)]<\infty$$ then for every $\epsilon>0$ $$\label{cp} \lim_{n\to \infty}Prob_n\Big\{f_n: |\frac{1}{n}\mathcal{N}_n(U,f_n)-\mu_{Q}(U)\Big|\geq \epsilon\}=0$$ for every open set $U\Subset {\Bbb{C}}^*$ such that $\partial U$ has zero Lebesgue measure.
We remark that the condition (\[cp\]) is called *convergence in probability* in the context of probability theory. Moreover, (\[cp\]) is equivalent to the following statement: for every subsequence $n_k$ of positive integers there exists a further subsequence $n_{k_j} $ such that $\frac{1}{n_{k_j} }\mathcal{N}_{n_{k_j} }(U,f_{n_{k_j}} )\to \mu_{Q}(U)$ with probability one in $\mathcal{H}.$
Next, we consider *random elliptic polynomials* which are of the form $$G_n(z)=\sum_{|J|=n}a_J{n \choose J}^{\frac12}z^J$$ where ${n \choose J}=\frac{n!}{(n-|J|)! j_1!\dots j_m!}$ and $a_J$ are non-degenerate i.i.d. random variables.
Let us denote by $$\mathcal{M}_{U}:=\frac{1}{(m-1)!}\int_{U}\frac{i}{2\pi}\partial\overline{\partial}(\log(1+\|z\|^2))\wedge (\frac{i}{\pi}\partial\overline{\partial}\|z\|^2)^{m-1}.$$ The following result is an analogue of Theorem \[main\] in the present setting (see §\[elliptic\] for details):
\[SU\] Let $a_j$ be i.i.d. non-degenerate real or complex valued random variables and assume that the dimension of complex Euclidean space $m \geq 3.$ The logarithmic moment $$\Bbb{E}[\big(\log(1+|a_j|)\big)^m]<\infty$$ if and only if the zero loci of elliptic polynomials satisfy $$\label{3}
{\Bbb{P}}\Big\{\{G_n\}_{n\geq0}: \lim_{n\to \infty}\frac1nVol_{2m-2}(Z_{G_n}\cap U)=\mathcal{M}_{U}\Big\}=1$$ for every open set $U\Subset ({\Bbb{C}}^*)^m$ such that $\partial U$ has zero Lebesgue measure.
Finally, we consider random linear combinations of univariate orthonormal polynomials of regular asymptotic behavior (cf. [@StahlTotik §3]). Orthogonal polynomials of regular $n^{th}$ root asymptotic behavior are natural generalizations of classical orthogonal polynomials on the real line. More precisely, let $\mu$ be a measure Borel measure with compact support $S_{\mu}\subset {\Bbb{C}}.$ We assume that the support $S_{\mu}$ contains infinitely many points and its logarithmic capacity $Cap(S_{\mu})>0.$ We let $\Omega:=\overline{{\Bbb{C}}}\setminus S_{\mu}$ and $g_{\Omega}(z,\infty)$ denotes the Green function with logarithmic pole at infinity. Then the equilibrium measure of the support $S_{\mu}$ is given by $\nu_{S_{\mu}}:=\Delta g_{\Omega}(z,\infty).$ We say that $\Omega$ is *regular* if $g(z,\infty)\equiv 0$ on $S_{\mu}.$ It is well know that if $\Omega$ is regular then $g(z,\infty)$ is continuous on ${\Bbb{C}}.$ Next, we define the inner product induced by $\mu:$ $${\langle}f,g {\rangle}:=\int_{{\Bbb{C}}}f(z)\overline{g(z)}d\mu$$ on the space of polynomials $\mathcal{P}_n.$ Then one can find uniquely defined orthonormal polynomials $$P_n^{\mu}(z)=\gamma_n(\mu)z^n+\cdots,\ \text{where}\ \gamma_n(\mu)>0\ \text{and}\ n\in\Bbb{N}.$$ We say that $\mu$ is *regular*, denoted by $\mu\in \textbf{Reg}$, if $$\lim_{n\to \infty}\gamma_n(\mu)^{1/n}=\frac{1}{Cap(S_{\mu})}.$$ For a fixed $\mu\in \textbf{Reg},$ we consider random linear combinations of orthonormal polynomials $$f_n(z)=\sum_{j=0}^na_jP^{\mu}_j(z)$$ and we obtain the following generalization:
\[onp\] Let $\mu\in \textbf{Reg}$ such that $\Omega:=\overline{{\Bbb{C}}}\setminus S_{\mu}$ is connected and regular. Assume that the convex hull $Co(S_{\mu})$ has Lebesgue measure zero (hence, $Co(S_{\mu})$ is a line segment). If the logarithmic moment $$\Bbb{E}[\log(1+|a_j|)]<\infty$$ then for every $\epsilon>0$ $$\lim_{n\to \infty}Prob_n\Big\{f_n: |\frac{1}{n}\mathcal{N}_n(U,f_n)-\nu_{S_{\mu}}(U)\Big|\geq \epsilon\}=0$$ for every open set $U\Subset {\Bbb{C}}^*$ such that $\partial U$ has zero Lebesgue measure.
We remark that if $\mu$ is a Bernstein-Markov measure with compact support in ${\Bbb{C}}$ then $\mu\in\textbf{Reg}$ ([@Bloom1 Proposition 3.4]). In particular, any Bernstein-Markov measure $\mu$ supported on a compact subset of the real line falls in the framework of Theorem \[onp\]. The latter class contains classical orthogonal polynomials such as Chebyshev or Jacobi polynomials.
Acknowledgment {#acknowledgment .unnumbered}
==============
We thank N. Levenberg and T. Bloom for their comments on an earlier version of this manuscript. We are grateful to N. Levenberg for pointing out Theorem \[onp\] falls in the framework of this work.
Background {#prelim}
==========
Probabilistic preliminaries {#probprelim}
---------------------------
For a complex (respectively real) random variable $\eta$ we let ${{\bf{P}}}$ denote its probability law and denote its concentration function by $$\mathcal{Q}(\eta,r):=\sup_{z\in \Bbb{C}}{{\bf{P}}}[\eta\in B(z,r)]$$ where $B(z,r)$ denotes the Euclidean ball (respectively interval) centered at $z$ and of radius $r>0.$ We say that $\eta$ is *non-degenerate* if $\mathcal{Q}(\eta,r)<1$ for some $r>0.$ If $\eta$ and $\xi$ are independent complex random variables and $r,c>0$ then we have $$\label{levi}
\mathcal{Q}(\eta+\xi,r)\leq \min\{\mathcal{Q}(\eta,r),\mathcal{Q}(\xi,r)\}\ \text{and} \ \mathcal{Q}(c\zeta,r)=\mathcal{Q}(\zeta,\frac{r}{c}).$$
Let $a_1,a_2,\dots $ be independent and identically distributed (real or complex valued) random variables. The following lemma is standard in the literature and it will be useful in the sequel.
\[lem1\] Let $a_j$ be a sequence of i.i.d. real or complex valued random variables for $j=1,2,\dots$
- If $\Bbb{E}[\big(\log(1+|a_j|)\big)^m]<\infty$ then for each $\epsilon>0$ almost surely $$\label{id}
|a_j|< e^{\sqrt[m]{\epsilon j}}$$ for sufficiently large $j.$
- If $\Bbb{E}[(\log(1+|a_j|))^m]=\infty$ then almost surely $$\limsup_{j\to \infty}|a_j|^{\frac1j}=\infty.$$
For a non-negative random variable $X$ we have $$\label{ineq1}
\sum_{j=1}^{\infty}{{\bf{P}}}[X\geq j]\leq \Bbb{E}[X]\leq 1+\sum_{j=1}^{\infty}{{\bf{P}}}[X\geq j].$$ Letting $X=\frac{1}{\epsilon}(\log(1+|a_1|))^m$ and using the assumption that $a_j$ are identically distributed, we obtain $$\sum_{j=1}^{\infty}{{\bf{P}}}[a_j\in {\Bbb{C}}: |a_j|\geq e^{\sqrt[m]{j\epsilon}}]<\infty.$$ Hence, by independence of $a_j$’s and Borel-Cantelli lemma we have almost surely $$|a_j|<e^{\sqrt[m]{j\epsilon}}$$ for sufficiently large $j$.\
For (ii), we define the event $A^M_j:=\{a_j\in{\Bbb{C}}:|a_j|^{\frac{m}{j}}\geq M\}$ where $M>1$ is fixed. Then by (\[ineq1\]) $$\sum_{j=1}^{\infty}{{\bf{P}}}_n[A^M_j]=\infty$$ and second Borel-Cantelli lemma implies that almost surely $|a_j|^{\frac{m}{j}}\geq M$ for infinitely many values of $j.$ Now, we let $M_n>0$ be a sequence such that $M_n\uparrow \infty.$ Then by previous argument the event $$F_n:=\{|a_j|^{\frac{m}{j}}\geq M_n\ \text{for infinitely many}\ j\}$$ has probability one. Thus letting $F=\cap_{n=1}^{\infty}F_n$ has also probability one and (ii) follows.
Pluripotential Theory
---------------------
### Global extremal function
Let $\Sigma\subset {\Bbb{C}}^m$ be a closed set. Recall that an *admissible weight function* $Q:{\Bbb{C}}^m\to \Bbb{R}$ is a lower semi-continuous function that satisfies
1. $ \{z\in\Sigma: Q(z)<\infty\}\ \text{is not pluripolar}$
2. $\displaystyle \lim_{\|z\|\to \infty}(Q(z)-\log\|z\|)=\infty$ if $\Sigma$ is unbounded.
The *weighted extremal function* associated to the pair $(\Sigma,Q)$ is defined by $$V_{\Sigma,Q}=\sup\{u(z):u\in\mathcal{L}({\Bbb{C}}^m), u\leq Q\ \text{on}\ \Sigma\}.$$ If $\Sigma={\Bbb{C}}^m$ and $Q$ is an admissible weight function we write $V_Q$ for short. We also let $V_{\Sigma,Q}^*$ denote the upper semi-continuous regularization of $V_{\Sigma,Q}$ that is $V_{\Sigma,Q}^*(z):=\displaystyle\limsup_{\zeta\to z}V_{\Sigma,Q}(\zeta).$ It is well known that $V_{\Sigma,Q}^*\in \mathcal{L}^+({\Bbb{C}}^m)$ (see [@SaffTotik Appendix B]). Moreover, for an admissible weight function $Q$ the set $$\{z\in {\Bbb{C}}^m: V_{\Sigma,Q}(z)<V_{\Sigma,Q}^*(z)\}$$ is pluripolar. We also remark that when $Q\equiv 0$ and $\Sigma$ is a non-pluripolar compact set the function $V_ {\Sigma}^*$ is nothing but the pluricomplex Green function of $\Sigma$ (see [@Klimek §5]). We let $B(r)$ denote the ball in ${\Bbb{C}}^m$ centered at the origin and with radius $r>0.$ Then it is well known [@SaffTotik Appendix B] that for sufficiently large r $$V_Q=V_{B(r),Q}\ \text{on}\ {\Bbb{C}}^m$$ for every admissible weight function $Q.$ It also follows from a result of Siciak [@Siciak Proposition 2.16] that if $Q$ is a continuous admissible weight function then $V_Q=V_Q^*$ on ${\Bbb{C}}^m.$ We refer the reader to the manuscript [@SaffTotik Appendix B] for further properties of the weighted global extremal function.
### Bergman kernel asymptotics {#scvprelim}
In the sequel we will assume that $Q:{\Bbb{C}}^m\to\Bbb{R}$ is a $\mathscr{C}^2$ weight function satisfying (\[growth\]) and (\[circular\]). The *Bergman kernel* for the Hilbert space of weighted polynomials $\mathcal{P}_n$ may be defined as $$S_n(z,w):=\sum_{j=1}^{d_n}P_j^{n}(z)\overline{P_j^{n}(w)}$$ where $\{P_j^n\}_{j=1}^{d_n}$ is an ONB for $\mathcal{P}_n$ as in the introduction. The restriction of the Bergman kernel over the diagonal is given by $$S_n(z,z)=\sum_{j=1}^{d_n}|P^{n}_j(z)|^2.$$ It is well known [@BloomL §6] (cf. [@Berman1; @B4]) that $$\frac{1}{2n}\log S_n(z,z)\to V_Q(z)\ \text{locally uniformly on}\ {\Bbb{C}}^m.$$
Proofs
======
By [@BloomL Proposition 4.4] it is enough to prove that almost surely in $\mathcal{H},$ for any subsequence $I$ of positive integers $$(\limsup_{n\in I}\frac1n\log|f_n(z)|)^*=V_{Q}(z)$$ for all $z\in {\Bbb{C}}^m.$ To this end we fix a subsequence $I$ of positive integers.
**Step 1: Proof of upper bound.** Note that by Lemma \[lem1\] for each $\epsilon>0$ there exists $j_0\in\Bbb{N}$ such that almost surely $$\sum_{j=j_0}^{d_n}|a_j|^2\leq d_ne^{2\sqrt[m]{\epsilon d_n}}.$$ Then using $d_n=O(n^m)$ and by Cauchy-Schwarz inequality almost surely in $\mathcal{H}$ $$\begin{aligned}
\limsup_{n\in I}\frac1n\log|f_n(z)| &=& \limsup_{n\in I}\big(\frac1n \log\frac{|f_n(z)|}{\sqrt{S_n(z,z)}}+\frac{1}{2n}\log S_n(z,z)\big)\\
&\leq& \limsup_{n\to \infty}\big(\frac{1}{2n}\log(\sum_{j=1}^{d_n}|a_j|^2) +\frac{1}{2n}\log S_n(z,z)\big)\\
&\leq& \epsilon+V_{Q}(z)\end{aligned}$$ on ${\Bbb{C}}^m.$ Thus, it follows from [@Bloom1 Lemma 2.1] that $(\displaystyle\limsup_{n\in I}\frac1n\log|f_n(z)|)^*\in\mathcal{L}({\Bbb{C}}^m)$ and $$\label{limsup}
F(z):=(\limsup_{n\in I}\frac1n\log|f_n(z)|)^*\leq V_{Q}(z)$$ holds on ${\Bbb{C}}^m$ almost surely in $\mathcal{H}.$\
**Step 2: Proof of lower bound.** In order to get the lower bound first we prove the following lemma which is a generalization of [@B8 Proposition 2.1]:
\[orthogonal\] For every $\epsilon>0$ and $z\in ({\Bbb{C}}^*)^m$ there exists $\delta>0$ such that for sufficiently large $n\in\Bbb{N}$ $$\#\{j\in\{1,\dots,d_n\}: P^n_j(z)>e^{n(V_{Q}(z)-3\epsilon)}\}\geq \delta d_n.$$
We denote the probability measures $\mu_n:=\frac{1}{b_n}e^{-2nQ(z)}dV_m$ where the normalizing constants $b_n:=\int_{{\Bbb{C}}^m}e^{-2nQ(z)}dV_m$. It follows that the sequence of measures $\{\mu_n\}_{n=1}^{\infty}$ satisfies large deviation principle (LDP) on ${\Bbb{C}}^m$ with the rate function $\mathcal{I}(z)=2[Q(z)-\displaystyle\inf_{w\in{\Bbb{C}}^m}Q(w)]$ (see e.g. [@DeSt 1.1.5]). More precisely, for $A\subset {\Bbb{C}}^m$ letting $$\mathcal{I}(A):=\inf_{z\in A}\mathcal{I}(z)$$ we have $$\limsup_{n\to \infty}\frac1n\log\mu_n(K)\leq -\mathcal{I}(K)\ \text{and}\ \liminf_{n\to\infty}\frac1n\log\mu_n(U)\geq -\mathcal{I}(U)$$ for every closed set $K\subset {\Bbb{C}}^m$ and every open set $U\subset {\Bbb{C}}^m.$
Next, we define $$c^n_{nT}:=(\int_{{\Bbb{C}}^m}|z^T|^{2n}e^{-2nQ(z)}dV_m)^{-\frac12}$$ where $T\in [0,1]^m$ is a multi-index and $z^T=z_1^{t_1}\cdots z_m^{t_m}$. Then by Varadhan’s lemma [@DeSt Theorem 2.1.10] and (\[growth\]), for every such $T=(t_1,\dots,t_m)$ $$\begin{aligned}
-\lim_{n\to \infty}\frac1n\log c^n_{nT} &=&\sup_{r\in\Bbb{R}^m_+}(\sum_{j=1}^mt_j\log r_j-Q(r_1,\dots,r_m))\\
&=& \sup_{S\in {\Bbb{R}^m}}({\langle}S,T{\rangle}-Q(e^{s_1},\dots,e^{s_m}))\\
&=:&u(T).\end{aligned}$$ Let us denote by $$\Phi(S):=Q(e^{s_1},\dots,e^{s_m})$$ where $S=(s_1,\dots,s_m)\in {\Bbb{R}^m}$ and Legendre-Fenchel transform of $\Phi$ is by definition given by $$\begin{aligned}
\Phi^{\star}(T):&=&\sup_{S\in {\Bbb{R}^m}}({\langle}S,T{\rangle}-\Phi(S))\\
&=&\sup_{S\in {\Bbb{R}^m}_{\geq0}}({\langle}S,T{\rangle}-\Phi(S)).\end{aligned}$$ where the second equality follows from $Q\geq 0$. Since $u(T)=\Phi^{\star}(T)$ for $T\in [0,1]^m$ the function $u(T)$ is a lower-semicontinuous convex on $[0,1]^m$.
On the other hand, denoting by $\Psi(S):=V_Q(e^{s_1},\dots,e^{s_m})$ since $\Psi$ is a $\mathcal{C}^{1,1}$ convex function we have $$\Psi(S)=\Psi^{\star\star}(S).$$ Thus, for every $\epsilon>0$ and $S\in {\Bbb{R}^m}$ there exists $T_0\in R^m_{\geq0}$ such that $$\Psi(S)-\epsilon<\langle S,T_0\rangle-\Psi^{\star}(T_0)\leq \langle S,T_0\rangle-\Phi^{\star}(T_0)$$ where the latter inequality follows from the inequality $V_Q\leq Q$ on ${\Bbb{C}}^m.$ Moreover, it follows from [@Rockafellar Theorem 23.5] and $V_Q\in \mathcal{C}^{1,1}({\Bbb{C}}^m)$ that $T_0=\nabla\Psi(S)$ and hence by using $V_Q\in \mathcal{L}$ we conclude that $T_0\in [0,1]^m.$ Thus, for every $\epsilon>0$ and $S\in {\Bbb{R}^m}$ there exists $T_0\in [0,1]^m$ such that $${\langle}S,T_0{\rangle}-u(T_0)>V_Q(e^{s_1},\dots,e^{s_m})-\epsilon$$ and by lower-semicontinuity of $u$ there exists a product of intervals $\mathcal{J}\subset [0,1]^m$ containing $T_0$ such that the Lebesgue measure $|\mathcal{J}|>0$ and $${\langle}S,T{\rangle}-u(T)>V_Q(e^{s_1},\dots,e^{s_m})-2\epsilon\ \ \text{for every}\ T\in \mathcal{J}.$$ Now, for fixed $z\in({\Bbb{C}}^*)^m$ letting $S=(\log|z_1|,\dots,\log|z_m|)$ then for sufficiently large $n$ we have $$\frac1n \log (c^n_{Tn}|z^T|^{n})>V_{Q}(z)-3\epsilon$$ for every $T\in \mathcal{J}.$ Finally, letting $\mathcal{J}_n:=\{J\in\Bbb{N}^m: |J|\leq n\ \text{and}\ \frac{1}{n}J\in \mathcal{J}\}$ where $\frac1nJ:=(\frac{j_1}{n},\dots,\frac{j_m}{n})$ we see that for sufficiently large $n$ we have $$\#\mathcal{J}_n\geq \frac{d_n}{2}|\mathcal{J}|$$ where $|\mathcal{J}|$ denotes Lebesgue measure of $\mathcal{J}\subset {\Bbb{R}^m}.$
Now, we turn back to proof of the lower bound. For fixed $z\in({\Bbb{C}}^*)^m$ and for every $\epsilon>0$ by Lemma \[orthogonal\] there exists a product interval $\mathcal{J}\subset [0,1]^m$ such that $$P_j^n(z)>e^{n(V_{Q}(z)-\epsilon)}$$ where $P_j^n(z)=C_J^nz^J$ and $J\in \mathcal{J}_n:=\{|J|\leq n: \frac{1}{n}J\in \mathcal{J}\}$. Next, we define the random variables $$X_n:=\sum_{j\in \mathcal{J}_n}a_j\alpha_j\ \text{and}\ Y_n:=\sum_{j\not\in \mathcal{J}_n}a_j\alpha_j$$ where $$\alpha_j:=e^{-n(V_{Q}(z)-\epsilon)}P_j^n(z).$$ Then by (\[levi\]) and sufficiently large $n$ we have $$\label{es1}
Prob_n[f_n:|f_n(z)|<e^{n(V_{Q}(z)-2\epsilon)}]\leq \mathcal{Q}(X_n+Y_n, e^{-\epsilon n})\leq \mathcal{Q}(X_n,e^{-\epsilon n}).$$ Now, it follows from Kolmogorov-Rogozin inequality [@Esseen] and $\alpha_j>1$ that $$\label{es2} \mathcal{Q}(X_n,e^{-\epsilon n})\leq C_1(\sum_{J\in \mathcal{J}_n}(1- \mathcal{Q}(a_j\alpha_j,e^{-\epsilon n}))^{-\frac12}\leq C_2 |\mathcal{J}_n|^{-\frac12}\leq C_3(d_n)^{-\frac12}.$$ Hence combining (\[es1\]) and (\[es2\]) we obtain: for every $z\in ({\Bbb{C}}^*)^m$ there exists $C_{\epsilon}>0$ such that $$\label{es3}
Prob_n[f_n:\frac1n\log|f_n(z)|<V_{Q}(z)-\epsilon]\leq \frac{C_{\epsilon}}{\sqrt{n^m}}.$$ Since $m\geq 3,$ it follows from Borel-Cantelli lemma and (\[es3\]) that with probability one in $\mathcal{H}$ $$\label{liminf}
\liminf_{n\to \infty}\frac1n\log|f_n(z)|\geq V_Q(z).$$ Thus, we conclude that for each $z\in({\Bbb{C}}^*)^m$ there exits a subset $\mathcal{C}_z\subset\mathcal{H}$ of probability one such that that for every sequence $\{f_n\}_{n\in\Bbb{N}}\in \mathcal{C}_z$ $$\label{lim}
F(z)=(\limsup_{n\in I}\frac1n\log|f_n(z)|)^*=V_Q(z)$$ Next, we fix a countable dense subset $D:=\{z_j\}_{j\in\Bbb{N}}$ in ${\Bbb{C}}^m$ such that $z_j\in({\Bbb{C}}^*)^m$ and (\[lim\]) holds. Then, we define $$\mathcal{C}:=\cap_{j=1}^{\infty}\mathcal{C}_{z_j}.$$ Note that $\mathcal{C}\subset \mathcal{H}$ is also of probability one. Since $V_{Q}(z)$ is continuous on ${\Bbb{C}}^m$ we have $$V_{Q}(z)=\lim_{z_j\in D, z_j\to z}V_{Q}(z_j)\leq \limsup_{z_j\in D,z_j\to z}F(z_j)\leq F(z)$$ where the second inequality follows from (\[liminf\]) and the last one follows from upper-semicontinuity of $F(z).$ We deduce that for every $\{f_n\}_{n\in\Bbb{N}}\in \mathcal{C}$ $$F(z)=V_Q(z)$$ for every $z\in({\Bbb{C}}^*)^m.$ Since $\{z\in{\Bbb{C}}^m: z_1\cdots z_m=0\}$ has Lebesgue measure zero, by a well-known property of psh functions we conclude that $$F(z)=V_Q(z)$$ for every $z\in {\Bbb{C}}^m.$ This completes the proof for dimensions $m\geq3$.
On the other hand, it follows from [@BloomL Proposition 4.4], Step 1, (\[es3\]) and the preceding argument that for every $\epsilon>0,$ open set $U\Subset {\Bbb{C}}^m$ and sufficiently large $n$ $$Prob_n[f_n\in\mathcal{P}_n: \|\frac1n\log|f_n|-V_Q\|_{L^1(U)}\geq \epsilon]\leq \frac{C_{\epsilon}}{\sqrt{n^m}}$$ which gives the second assertion.
First, we prove that (\[A\]) is a sufficient condition for (\[2\]). We fix an open set $U\Subset ({\Bbb{C}}^*)^m$ such that $\partial U$ has zero Lebesgue measure. Let us denote by $$\Theta:=\frac{1}{(m-1)!}\frac{i}{\pi}\partial\overline{\partial}V_{Q}\wedge (\frac{i}{2}\partial\bar{\partial}\|z\|^2)^{m-1}.$$ For $\delta>0$ arbitrary, we fix real valued smooth functions $\varphi_1,\varphi_2$ such that $0\leq \varphi_1\leq \chi_U\leq \varphi_2\leq 1$ and $$\int_U\Theta-\delta\leq \int_{{\Bbb{C}}^m}\varphi_1\Theta\leq \int_{{\Bbb{C}}^m}\varphi_2\Theta\leq \int_{\overline{U}}\Theta+\delta.$$ Now, letting $$\psi_j:=\frac{\varphi_j}{(m-1)!}(\frac{i}{2}\partial\bar{\partial}\|z\|^2)^{m-1}$$ for $j=1,2$ by Wirtinger’s theorem we have $$Vol_{2m-2}(Z_{f_n}\cap U)\leq \int_{Z_{f_n}}\psi_2.$$ Then by Theorem \[main 1\] $$\begin{aligned}
\limsup_{n\to \infty}\frac1n Vol_{2m-2}(Z_{f_n}\cap U) &\leq &\int_{{\Bbb{C}}^m} \varphi_2\Theta\\
&\leq& \int_{\overline{U}}\Theta+\delta.\end{aligned}$$ Similarly one can obtain $$\liminf_{n\to \infty}\frac1nVol_{2m-2}(Z_{f_n}\cap U)\geq \int_{U}\Theta-\delta.$$ Since $\delta>0$ is arbitrary the assertion follows.
Next, we prove that (\[A\]) is a necessary condition for (\[2\]). We will prove the assertion by contradiction. Assume that $$\Bbb{E}[\big(\log(1+|a_j|)\big)^m]=\infty.$$ By assumption $U\Subset ({\Bbb{C}}^*)^m$ so we have $0<b_n:=\min_{j=1,\dots,d_n}\inf_{z\in U}|P^n_{j}(z)|.$ For $\epsilon>0$ small we let $$t_n:=\big(\frac{e^{n(M_Q+\epsilon)}}{b_n}\big)^m$$ where $M_Q:=\sup_{\overline{U}}V_{Q}.$ Then by the argument in the proof of Lemma \[lem1\] (ii) for each $n\in \Bbb{N}_+$ the set $$F_n:=\{|a_{j}|^{m/j}\geq t_n\ \text{for infinitely many}\ j\}$$ has probability one. This implies that $$F:=\cap_{n=1}^{\infty} F_n$$ has also probability one. Thus, we may assume that for infinitely many values of $n$ there exists $j_n\in\{1,\dots,d_n\}$ such that $$\label{ff}
\max_{j=1,\dots ,d_n} |a_j|^{\frac1j}= |a_{j_n}|^{\frac{1}{j_n}}\ \text{and} \ |a_{j_n}|\geq t_n^{j_n/m}.$$ For simplicity of notation let us assume $j_n=d_n.$ Now, we will show that the random polynomial $f_n(z)=\sum_{j=1}^{d_n}a_jP_j^{n}(z)$ has no zeros in $U$ for infinitely many values of $n$. Denoting $a':=(a_j)_{j=1}^{d_n-1},$ by Cauchy-Schwarz inequality, uniform convergence of the Bergman kernel on $\overline{U}$ and (\[ff\]) we have $$\begin{aligned}
|\sum_{j=1}^{d_n-1}a_jP_j^{n}(z)|
&\leq& \|a'\| S_n(z,z)^{\frac12} \\
&\leq & \sqrt{d_n}|a_{d_n}|^{\frac{d_n-1}{d_n}} \exp(n(V_{Q}(z)+\frac{\epsilon}{2}))\\
&\leq & |a_{d_n}|^{\frac{d_n-1}{d_n}} \exp(n(M_Q+\epsilon))\\
&= & \frac{ \exp(n(M_Q+\epsilon))}{|a_{d_n}|^{\frac{1}{d_n}}} |a_{d_n}|\\
&<& b_n|a_{d_n}|
\end{aligned}$$ for infinitely many values of $n$. Hence, $$\sup_{z\in U} |\sum_{j=1}^{d_n-1}a_jP_j^{n}(z)| <\inf_{z\in U}|a_{d_n}P_{d_n}^n(z)|.$$
Generalizations and Concluding remarks
======================================
Elliptic Polynomials {#elliptic}
--------------------
Recall that a *random elliptic polynomial* in ${\Bbb{C}}^m$ is of the form $$G_n(z)=\sum_{|J|\leq n}a_J{n \choose J}^{\frac12}z^J$$ where ${n \choose J}=\frac{n!}{(n-|J|)! j_1!\dots j_m!}$ and $a_J$ are non-degenerate i.i.d. random variables. These polynomials induced by taking $Q(z)=\frac12\log(1+\|z\|^2)$ i.e. the potential of the standard Fubini-Study Kähler metric on the complex projective space ${\Bbb{C}}\Bbb{P}^m.$ In this case, the scaled monomials ${N \choose J}^{\frac12}z^J$ form an ONB with respect to the inner product $$\langle F_n,G_n\rangle_n:=\int_{{\Bbb{C}}^m}F_n(z)\overline{G_n(z)}\frac{dV_m(z)}{(1+\|z\|^2)^{n+m+1}}$$ Moreover, since $Q(z)$ is itself a Lelong class of psh function the weighted extremal function in this setting is given by $$V_Q(z)=Q(z)=\frac12\log(1+\|z\|^2).$$ Specializing further, if the coefficients $a_J$ are standard i.i.d. complex Gaussians this ensemble is known as $SU(m+1)$ polynomials and their zero distribution was studied extensively among others by [@BBL; @SZ].
Since the proof is very similar to that of Theorems \[main 1\] and \[main\] we explain the modifications in the present setting.
By [@BloomL Proposition 4.4] it is enough to prove that almost surely in $\mathcal{H},$ for any subsequence $I$ of positive integers $$F(z):=(\limsup_{n\in I}\frac1n\log|f_n(z)|)^*=V_{Q}(z)\ \text{for all}\ z\in {\Bbb{C}}^m$$
In order to prove the upper bound $F(z)\leq V_Q(z),$ we use the same argument as in Thorem \[main 1\] together with the Bergman kernel asymptotics. Namely, letting $S_n(z,z):=\sum_{|J|\leq n} {n \choose J}|z^{2J}| $ a routine calculation gives $$\frac{1}{2n}\log S_n(z,z)\to \frac12\log(1+\|z\|^2)$$ locally uniformly on ${\Bbb{C}}^m$ (see eg. [@SZ]). On the other hand, for the lower bound (\[liminf\]), we need an analogue of Lemma \[orthogonal\]. Note that $Q(z)=\frac12\log(1+\|z\|^2)$ is a multi-circular weight function whose infimum is 0 attained at $z=0.$ Then proceeding as in the proof Lemma \[orthogonal\], one can show that the sequence of measures $\mu_n:=\frac{1}{a_n}e^{-2nQ(z)}dV_m$ verifies a LDP with rate function $\mathcal{I}(z)=2Q(z).$ This result and Kolmogorov-Rogozin inequality allow us to prove an analogue of (\[es3\]) in the present setting. This together with the argument in the first part of the proof of Theorem \[main\] finish the proof of sufficiency of (\[A\]). In order to prove necessity, we use the Bergman kernel asymptotics and we apply the same argument as in the second part of the proof of Theorem \[main\].
Regular Orthonormal Polynomials
-------------------------------
We proceed as in the proof of Theorems \[main 1\] and \[main\]. To this end we fix a subsequence $n_k$ of positive integers. It follows from [@StahlTotik Theorem 3.1(ii) ] that $$\label{st}
\lim_{n\to\infty}\frac1n\log |P^{\mu}_n(z)|= g_{\Omega}(z,\infty)$$ holds locally uniformly on ${\Bbb{C}}\setminus Co(S_{\mu}).$ Denoting the Bergman kernel by $$S_n(z,z):=\sum_{j=0}^{n}|P_j^{\mu}(z)|^2$$ we infer that $$\frac{1}{2n}\log S_n(z,z)\to g_{\Omega}(z,\infty)$$ locally uniformly on ${\Bbb{C}}\setminus Co(S_{\mu}).$ Thus, by Lemma \[lem1\] and Cauchy-Schwarz inequality almost surely in $\mathcal{H}$ we have $$\limsup_{n_k\to \infty}\frac{1}{n_k}\log|f_{n_k}(z)|\leq g_{\Omega}(z,\infty)$$ for every $z\in {\Bbb{C}}\setminus Co(S_{\mu}).$
In order to prove the lower bound, we use the local uniform convergence (\[st\]) which replaces Lemma \[orthogonal\]. This in turn together with Kolmogorov-Rogozin inequality give $$Prob_n[f_n:\frac1n\log|f_n(z)|<g_{\Omega}(z,\infty)-\epsilon]\leq \frac{C_{\epsilon}}{\sqrt{n}}$$ for every $z\in{\Bbb{C}}^*\setminus Co(S_{\mu}).$ Then applying the argument in Theorem \[main\] using the assumption $Co(S_{\mu})$ has Lebesgue measure zero we obtain the assertion.
Almost sure convergence in lower dimensions
-------------------------------------------
In order to get almost sure convergence in Theorems \[main 1\] and \[main\] for complex dimensions $m\leq 2$ we need a stronger form of Kolmogorov-Rogozin inequality. More precisely, for a fixed unit vector ${u^{(n)}}\in{\Bbb{C}}^n,$ i.i.d. real or complex random variables $a_j$ for $ j=1,\dots ,n$ and $\epsilon\geq 0$ we consider the *small ball probability* $$p_{\epsilon}({u^{(n)}}):={{\bf{P}}}_n[\{a^{(n)}:|{\langle}{a^{(n)}},{u^{(n)}}{\rangle}|\leq\epsilon\}]$$ where ${{\bf{P}}}_n$ is the product probability measure induced by the law of $a_j's$ and ${\langle}{a^{(n)}},{u^{(n)}}{\rangle}:=\sum_{j=1}^na_j{u^{(n)}}_j.$ In order to obtain the lower bound in Theorem \[main 1\] we need for every $\epsilon>0$ $$\sum_{n\geq 1}p_{e^{-\epsilon n}}(u^{(d_n)})<\infty$$ for every unit vector $u^{(d_n)}\in{\Bbb{C}}^{d_n}$.
We remark that if the random variables $a_j$ are standard (real or complex) Gaussians then the probability $p_{\epsilon}({u^{(n)}})\sim \epsilon.$ In particular, $p_{\epsilon}(u^{(n)})$ does not depend on the direction of the vector ${u^{(n)}}.$ However, for most other distributions, $p_{\epsilon}({u^{(n)}})$ does depend on the direction of ${u^{(n)}}.$ For instance if $a_j$ are Bernoulli random variables (i.e. taking values $\pm1$ with probability $\frac12$) then $p_{0}((1,1,0,\dots,0))=\frac12$ on the other hand, $p_{0}((1,1,\dots,1))\sim n^{-\frac12}.$ Determining small ball probabilities is a classical theme in probability theory. We refer the reader to the manuscripts [@FriS; @TaoV; @RV1; @RV2] and references therein for more details.\
Another interesting problem is to find a necessary and sufficient condition for almost sure convergence of normalized zero currents when the space of polynomials $\mathcal{P}_n$ is endowed with $d_n$-fold product probability measure. A sufficient condition was obtained in [@B6]. Namely, let $a_j^n$ be iid random variables whose probability ${{\bf{P}}}$ has a bounded density and logarithmically decaying tails i.e. $$\label{tailc}
{{\bf{P}}}\{a_j\in{\Bbb{C}}:\log |a_j|>R\}= O(R^{-\rho})\ \text{as}\ R\to \infty\ \text{for some}\ \rho>m+1.$$ We consider random polynomials of the form $f_n(z)=\sum_{j=1}^{d_n}a^n_jP^{n}_j(z)$. If (\[tailc\]) holds then almost surely normalized zero currents $\frac1n[Z_{f_n}]$ converges weakly to the extremal current $\frac{i}{\pi}\partial\overline{\partial}V_Q.$
### Higher codimensions
In [@B6 Theorem 1.2] (see also [@B7]) it is proved that if the coefficients of random polynomials $f_n(z)=\sum_{j=1}^{d_n}a^n_jP^n_j(z)$ are i.i.d random variables whose distribution law verifies (\[tailc\]) then almost surely normalized empirical measure of zeros $$\frac{1}{n^m}\sum_{\{z\in{\Bbb{C}}^m:f^1_n(z)=\dots=f^m_n(z)=0\}}\delta_z$$ of $m$ i.i.d. random polynomials $f_n^1,\dots,f_n^m$ converges weakly to the weighted equilibrium measure $(\frac{i}{\pi}\partial\overline{\partial}V_{\Sigma,Q}^*)^m$. In the present paper, we have observed that for codimension one we no longer need $a_j$ to have a density with respect to Lebesgue measure. For instance, $a_j$ can be discrete such as Bernoulli random variables. It would be interesting to know if [@B6 Theorem 1.2] or a weaker form of it (eg. convergence with high probability) also generalizes to the setting of discrete random variables.
[XXXXX]{}
T. Bayraktar, *Expected number of real roots for random linear combinations of orthogonal polynomials associated with radial weights*, Potential Anal. DOI: 10.1007/s11118-017-9643-9.
T. Bayraktar, *Equidistribution of zeros of random holomorphic sections*, Indiana Univ. Math. J. **65** (2016), no. 5, 1759–1793.
T. Bayraktar, *Asymptotic normality of linear statistics of zeros of random polynomials*, Proc. Amer. Math. Soc. **145** (2017), no. 7, 2917–2929.
T. Bayraktar, *Zero distribution of random sparse polynomials*, Michigan Math. J. **66** (2017), no. 2, 389–419.
E. Bogomolny, O. Bohigas, and P. Leboeuf, *Quantum chaotic dynamics and random polynomials*, J. Statist. Phys. **85** (1996), no. 5-6, 639–679.
R. Berman, *Bergman kernels for weighted polynomials and weighted equilibrium measures of [$\Bbb C^n$]{}*, Indiana Univ. Math. J. **58** (2009), no. 4, 1921–1946.
T. Bloom and N. Levenberg, *Random [P]{}olynomials and [P]{}luripotential-[T]{}heoretic [E]{}xtremal [F]{}unctions*, Potential Anal. **42** (2015), no. 2, 311–334.
T. Bloom, *Random polynomials and [G]{}reen functions*, Int. Math. Res. Not. (2005), no. 28, 1689–1708.
T. Bloom, *Random polynomials and (pluri)potential theory*, Ann. Polon. Math. **91** (2007), no. 2-3, 131–141.
T. Bloom and B. Shiffman, *Zeros of random polynomials on [$\Bbb C^m$]{}*, Math. Res. Lett. **14** (2007), no. 3, 469–479.
J.-D. Deuschel and D. W. Stroock, *Large deviations*, Pure and Applied Mathematics, vol. 137, Academic Press, Inc., Boston, MA, 1989.
T.-C. Dinh and N. Sibony, *Distribution des valeurs de transformations méromorphes et applications*, Comment. Math. Helv. **81** (2006), no. 1, 221–258.
C. G. Esseen, *On the concentration function of a sum of independent random variables*, Z. Wahrscheinlichkeitstheorie und Verw. Gebiete **9** (1968), 290–308.
O. Friedland and S. Sodin, *Bounds on the concentration function in terms of the [D]{}iophantine approximation*, C. R. Math. Acad. Sci. Paris **345** (2007), no. 9, 513–518.
J. M. Hammersley, *The zeros of a random polynomial*, Proceedings of the [T]{}hird [B]{}erkeley [S]{}ymposium on [M]{}athematical [S]{}tatistics and [P]{}robability, 1954–1955, vol. [II]{} (Berkeley and Los Angeles), University of California Press, 1956, pp. 89–111.
C. P. Hughes and A. Nikeghbali, *The zeros of random polynomials cluster uniformly near the unit circle*, Compos. Math. **144** (2008), no. 3, 734–746.
I. Ibragimov and D. Zaporozhets, *On distribution of zeros of random polynomials in complex plane*, Prokhorov and Contemporary Probability Theory, Springer, 2013, pp. 303–323.
M. Kac, *On the average number of real roots of a random algebraic equation*, Bull. Amer. Math. Soc. **49** (1943), 314–320.
M. Klimek, *Pluripotential theory*, London Mathematical Society Monographs. New Series, vol. 6, The Clarendon Press, Oxford University Press, New York, 1991, Oxford Science Publications.
Z. Kabluchko and D. Zaporozhets, *Asymptotic distribution of complex zeros of random analytic functions*, Ann. Probab. **42** (2014), no. 4, 1374–1395.
J. E. Littlewood and A. C. Offord, *On the number of real roots of a random algebraic equation. [III]{}*, Rec. Math. \[Mat. Sbornik\] N.S. **12(54)** (1943), 277–286.
I. Pritsker and K. Ramachandran, *Equidistribution of zeros of random polynomials*, J. Approx. Theory **215** (2017), 106–117.
R. T. Rockafellar, *Convex analysis*, Princeton Landmarks in Mathematics, Princeton University Press, Princeton, NJ, 1997, Reprint of the 1970 original, Princeton Paperbacks.
M. Rudelson and R. Vershynin, *The [L]{}ittlewood-[O]{}fford problem and invertibility of random matrices*, Adv. Math. **218** (2008), no. 2, 600–633.
M. Rudelson and R. Vershynin, *Smallest singular value of a random rectangular matrix*, Comm. Pure Appl. Math. **62** (2009), no. 12, 1707–1739.
J. Siciak, *Extremal plurisubharmonic functions in [${\bf C}^{n}$]{}*, Ann. Polon. Math. **39** (1981), 175–211.
H. Stahl and V. Totik, *General orthogonal polynomials*, Encyclopedia of Mathematics and its Applications, vol. 43, Cambridge University Press, Cambridge, 1992.
E. B. Saff and V. Totik, *Logarithmic potentials with external fields*, Grundlehren der Mathematischen Wissenschaften \[Fundamental Principles of Mathematical Sciences\], vol. 316, Springer-Verlag, Berlin, 1997, Appendix B by Thomas Bloom.
L. A. Shepp and R. J. Vanderbei, *The complex zeros of random polynomials*, Trans. Amer. Math. Soc. **347** (1995), no. 11, 4365–4384.
B. Shiffman and S. Zelditch, *Distribution of zeros of random and quantum chaotic sections of positive line bundles*, Comm. Math. Phys. **200** (1999), no. 3, 661–683.
B. Shiffman and S. Zelditch, *Equilibrium distribution of zeros of random polynomials*, Int. Math. Res. Not. (2003), no. 1, 25–49.
T. Tao and V. H. Vu, *Inverse [L]{}ittlewood-[O]{}fford theorems and the condition number of random discrete matrices*, Ann. of Math. (2) **169** (2009), no. 2, 595–632.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'This paper proposes an efficient method for the simultaneous estimation of the state of a quantum system and the classical parameters that govern its evolution. This hybrid approach benefits from efficient numerical methods for the integration of stochastic master equations for the quantum system, and efficient parameter estimation methods from classical signal processing. The classical techniques use Sequential Monte Carlo (SMC) methods, which aim to optimize the selection of points within the parameter space, conditioned by the measurement data obtained. We illustrate these methods using a specific example, an SMC sampler applied to a nonlinear system, the Duffing oscillator, where the evolution of the quantum state of the oscillator and three Hamiltonian parameters are estimated simultaneously.'
author:
- 'Jason F. Ralph'
- Simon Maskell
- Kurt Jacobs
title: 'Multi-parameter estimation along quantum trajectories with Sequential Monte Carlo methods'
---
Introduction
============
Stochastic master equations provide a model for the evolution of open quantum systems subject to continuous measurements [@Bel1999; @Wis2010; @Jac2014]. The trajectories that the stochastic master equations generate represent the evolution of the state of an individual quantum system, conditioned on a particular measurement record. In theoretical studies, the measurement record is a simulated sequence corresponding to a particular realization of the evolution. However, recent experiments that implement continuous quantum measurements have demonstrated that the evolution of individual quantum systems can be reconstructed from experimental data [@Mur2013; @Web2014; @Six2015; @Cam2016]. The generation of such trajectories in real time during the measurement process will be an important step towards state-dependent feedback control of individual quantum systems [@Bel1999; @Wis2010; @Jac2014]. Feedback control has been demonstrated in quantum systems using the output measurement record as an input signal to the control system in optical [@Smi2002; @Bra2012], opto-mechanical [@Kub2009; @Wil2015; @Sud2017], and mesoscopic superconducting systems [@Vij2012; @Ris2012]. In most of these examples, while the direct use of the measurement record in the control system demonstrates the utility of quantum feedback control, it is limited by the fact that the evolution of the underlying state of the system is not included in the generation of the controls. State-dependent control is more flexible and can include quantities that are estimated from, but are not directly measured in experiments.
Given a measurement record, a Stochastic Master Equation (SME) provides an estimate of the quantum state at each point in time, and – in the case of mixed states – an indication of the uncertainty associated with this state in terms of an estimate of its purity. The SME is derived by taking a single quantum system and coupling it weakly to environmental degrees of freedom that mediate a continuous measurement process. The continuous measurement of the coupled system is realized by continuously measuring the state of the environment, which can be modeled as a sequence of projective measurements on successive environmental degrees of freedom. The simplest example of this process is the measurement of the electromagnetic radiation emitted by the system, in which case the environment is the electromagnetic field [@Wis2010; @Jac2014]. In the most common form of the SME, a Markovian condition is applied, meaning that the environment carries information away from the system but does not by itself feed this information back to affect the system at a later time. The resulting evolution of the system is continuous and stochastic, with the stochastic term arising from the effect of the sequence of measurements on the combined system.
In simulations, an SME is used to analyze the properties of an open system under the action of a continuous sequence of measurement operators, using a realization for the noise process and calculating the evolution of the system conditioned on this realization. When interpreting experiments, the SME is used to reconstruct the estimate of the quantum state as a function of time from the given measurement record provided by the experiment. In this regard, the SME is very similar to classical state estimation techniques (often referred to as ‘target tracking’ or ‘object tracking’ [@Bla1986; @Bar2001; @Ral2010]), which are used to interpret sequences of classical noisy sensor measurements to form a coherent picture of the world. These classical techniques have been developed to interpret sensor data (radar or sonar signals, and sequences of images) where objects are moving against noisy backgrounds. The motion of the objects may be unpredictable or uncooperative, their identity may not be known from the measurements, they may be occluded for periods of time, and individual objects may not be fully resolved by the sensor. Classical state estimation techniques provide methods to solve all of these problems and ambiguities.
With a continuous quantum measurement, the measurement record contains information about the evolution of the particular quantum state (a [*quantum trajectory*]{}) but the properties of this trajectory also contain information about the classical parameters that govern the dynamics of the system: the classical parameters in the Hamiltonian and the strength of the coupling to the environment. In this paper, we demonstrate how the stochastic master equation can be augmented with techniques drawn from classical state estimation, Sequential Monte Carlo (SMC) methods, to estimate several Hamiltonian parameters efficiently alongside the quantum trajectories.
We begin our presentation by first reviewing other approaches to Hamiltonian parameter estimation, and the development of a set of Hybrid SMEs for the quantum evolution and the Kushner-Stratonovich equation for the classical parameters [@Ral2011] in sections II and III, respectively. In section IV, we then discuss how efficient classical parameter estimation techniques [@Gre2017] can be applied to the solution of the classical aspects of the Hybrid SME. Section V introduces an example system, the Duffing oscillator, which contains a number of relevant experimental parameters, and Section VI presents results for the simultaneous estimation of the quantum trajectories for the oscillator and up to three Hamiltonian parameters. Section VII discusses how such methods may be useful in practical systems and draws conclusions from the results presented.
Hamiltonian Parameter Estimation
================================
The problem of estimating the dynamical parameters of a quantum system has been studied previously by a number of authors, including those who have adopted a continuous measurement approach. This estimation process is often referred to as Hamiltonian parameter estimation. This is because the basic description of quantum dynamics is encapsulated by the Hamiltonian operator, which determines the equations of motion, and values of the classical parameters in the Hamiltonian determine the specifics of the evolution.
A standard method for determining the dynamics of a quantum system is to prepare it many times in a range of different initial states, allow it to evolve, and then measure it before re-preparation. The results of the measurements can then be combined in a tomographic-like process to obtain the equation of motion for linear Schrödinger evolution [@Chu1997]. An alternative approach, and the one in which we are interested here, is to prepare the system only once and to continually monitor its subsequent evolution to build a picture of its dynamics. A full description of the problem involves starting with a prior probability density for the parameters one wishes to determine and then using Bayes’ theorem to continually update this probability density from the stream of measurement results as they are obtained. A number of authors have considered this problem [@Gam2001; @Ver2001; @Sto2004; @Tsa2009a; @Tsa2009b; @Tsa2010; @Tsa2011; @Ral2011; @Neg2013; @Ber2015; @Bas2015; @Cor2017]. This is of particular interest when the parameters of a system change slowly with time, and one wishes to be able to track the variations in the parameters. It is also relevant to the problem of using quantum systems as probes to measure time-varying classical fields (such as gravity waves [@LIGO2016] and magnetic fields [@Yan2017]), as these fields appear as parameters in the Hamiltonian.
As discussed in the introduction, a dynamical equation referred to as the stochastic master equation (SME) can be used to track the evolution of a quantum system from the results of a continuous measurement so long as the dynamical parameters of the system are known. If they are not known then the full estimation problem involves both the SME and a Kushner-Stratonovich equation that evolves the probability density for the parameters of the system. The combined set of dynamical equations has been referred to as a [*Hybrid SME*]{} [@Ral2011]. The first papers on the subject of Hamiltonian parameter estimation via continuous measurements were concerned mainly with deriving the Hybrid SME and applying it to the estimation of a single parameter [@Gam2001; @Ver2001]. Subsequently, Tsang and collaborators considered the more general problem of smoothing in which a time-varying parameter (a signal or wave-form) is estimated from all the measurement results obtained, and determined the ultimate limits to this procedure [@Tsa2009a; @Tsa2009b; @Tsa2010; @Tsa2011; @Ber2015]. An alternative and interesting approach to the problem was proposed recently by Bassa *et al. *[@Bas2015]. While most of the related work on parameter estimation employs continuous measurements, this approach considered a sequence of instantaneous measurements, and employed a discrete version of the Hybrid SME where several classical parameter values were encoded in an expanded quantum state.
A major problem with the Hybrid SME is that it is highly demanding from a computational point of view; in order to evolve the Kushner-Stratonovich equation for the probability density describing the observer’s knowledge of the parameters, the SME must be evolved for every value of the parameters for which this density is appreciable. The grid of points for which the SME must be evolved becomes large very quickly as the number of parameters increases. Two previous papers have put forward methods aimed at addressing this difficulty. Ralph *et al. *[@Ral2011] and Cortez *et al. *[@Cor2017] considered the estimation of a single frequency parameter, and presented methods to bypass the Kushner-Stratonovich equation. These papers estimate the natural oscillation frequency of a qubit directly from the measurement record. This approach has many benefits in terms of computational efficiency, but it has the disadvantage of not providing a simultaneous estimate of the quantum state of the system – which would be provided by the full solution of the Hybrid SME. Here, we will explore the use of a potentially more powerful technique in which the probability density is replaced with a finite set of samples of the parameters that are evolved instead. The examples given below typically use 50-100 quantum states and the equivalent of thousands to millions of classical parameter values. The purpose is again to reduce the number of copies of the quantum state that must be evolved in parallel using the SME, but we will apply this method to the challenging problem of estimating multiple parameters simultaneously.
Hybrid Stochastic Master Equations
==================================
The simultaneous estimation of the quantum state of a system and the classical Hamiltonian parameters that govern its evolution was considered in Ref. [@Ral2011], where an approach was presented based on a set of parallel SMEs, each using a different set of parameter values contained in a vector $\underline{\lambda}$, which have an associated probability. The final mixed state is then constructed by averaging over the probabilities for the classical parameters. The probabilities associated with the different parameter vectors evolve via a Kushner-Stratonovich equation and are conditioned on the continuous measurement record [@Ral2011].
For a quantum system subject to a continuous measurement, with a known set of Hamiltonian parameters, the evolution of the quantum state, $\rho_c(t)$, conditioned on the measurement record, $y(t)$, is given by the stochastic master equation [@Bel1999; @Wis2010; @Jac2014]. In general, the interaction with the environment can be represented by a set of system operators which are coupled to environmental degrees of freedom, some of which are not measured $\hat{V}_j$ $(j = 1\dots m)$ (‘unprobed’ operators), and some of which are measured and generate the continuous weak measurement $\hat{L}_r$ $(r = 1\dots m')$. In an ideal case, the measurement record is 100% efficient, with all of the available information being reflected in the measurement record. Unfortunately, real measurements are rarely ideal and the continuous measurement record is often corrupted with extraneous (classical) noise sources. These extraneous degrees of freedom can be characterized by an efficiency parameter for the measurement operators, $\hat{L}_r$ has an efficiency $\eta_r$. Specifically, $\eta_r$ is the fraction of the total noise power due to the quantum measurement as opposed to power contained in the other extraneous noise sources.
For unprobed operators $\hat{V}_r$ and measurement operators $\hat{L}_r$, the general form for the SME is given by, $$\begin{aligned}
\label{sme1}
d\rho_c&=&- i \left[\hat{H},\rho_c\right]dt \nonumber\\
&&+\sum_{j=1}^{m} \left\{ \hat{V}_{j} \rho_c \hat{V}^{\dagger}_{j} -\frac{1}{2}\left(\hat{V}^{\dagger}_{j} \hat{V}_{j} \rho_c
+ \rho_c \hat{V}^{\dagger}_{j} \hat{V}_{j} \right)\right\}dt \nonumber\\
&&+\sum_{r=1}^{m'} \left\{ \hat{L}_{r} \rho_c \hat{L}^{\dagger}_{r} -\frac{1}{2}\left(\hat{L}^{\dagger}_{r} \hat{L}_{r} \rho_c
+ \rho_c \hat{L}^{\dagger}_{r} \hat{L}_{r} \right)\right\}dt \nonumber\\
&&+ \sum_{r=1}^{m'} \sqrt{\eta_r}\left(\hat{L}_{r}\rho_c+\rho_c \hat{L}^{\dagger}_{r}-\mathrm{Tr}(\hat{L}_{r}\rho_c+\rho_c \hat{L}^{\dagger}_{r}) \right)dW_{r} \nonumber\\\end{aligned}$$ where $\hat{H}$ is the Hamiltonian of the system, $dt$ is an infinitesimal time increment, and the measurement record for each of the measurement operators $\hat{L}_{r}$ during a time step $t\rightarrow t+dt$ is given by, $y(t+dt)-y(t)=dy_r(t)= \sqrt{\eta_j}\mathrm{Tr}(\hat{L}_{r}\rho_c+\rho_c \hat{L}^{\dagger}_{r}) dt+dW_{r}$. We will take $dW_{r}$ to be a real Wiener increment such that $dW_{r}=0$ and $dW_{r} dW_{r'} = \delta_{rr'}dt$ for simplicity, but this is not strictly necessary. More general forms of complex increments may also be used [@Wis2005].
Where the evolution of a quantum system is governed by a set of Hamiltonian parameters that are not known exactly, we can describe the parameters in terms of a classical probability density, $P(\underline{\lambda})$, where $\underline{\lambda}=(\lambda_1, \lambda_2,...)$. The system is then described by a set of SMEs, one for each set of possible parameter values, $$\begin{aligned}
\label{sme1a}
d\rho_{c,\underline{\lambda}}&=&- i \left[\hat{H}(\underline{\lambda}),\rho_{c,\underline{\lambda}}\right]dt \nonumber\\
&&+\sum_{j=1}^{m} \left\{ \hat{V}_{j} \rho_{c,\underline{\lambda}} \hat{V}^{\dagger}_{j} -\frac{1}{2}\left(\hat{V}^{\dagger}_{j} \hat{V}_{j} \rho_{c,\underline{\lambda}}
+ \rho_{c,\underline{\lambda}} \hat{V}^{\dagger}_{j} \hat{V}_{j} \right)\right\}dt \nonumber\\
&&+\sum_{r=1}^{m'} \left\{ \hat{L}_{r} \rho_{c,\underline{\lambda}} \hat{L}^{\dagger}_{r} -\frac{1}{2}\left(\hat{L}^{\dagger}_{r} \hat{L}_{r} \rho_{c,\underline{\lambda}}
+ \rho_{c,\underline{\lambda}} \hat{L}^{\dagger}_{r} \hat{L}_{r} \right)\right\}dt \nonumber\\
&&+ \sum_{r=1}^{m'}
\sqrt{\eta_r}\left(\begin{array}{l}
\hat{L}_{r}\rho_{c,\underline{\lambda}}+\rho_{c,\underline{\lambda}} \hat{L}^{\dagger}_{r} \\
-\mathrm{Tr}(\hat{L}_{r}\rho_{c,\underline{\lambda}}+\rho_{c,\underline{\lambda}} \hat{L}^{\dagger}_{r}) dW_{r}
\end{array}\right)\end{aligned}$$
The evolution of the probability density $P(\underline{\lambda})$ is governed by a Kushner-Stratonovich stochastic differential equation, derived in [@Ral2011] for a single efficient measurement operator and in the absence of additional unprobed environmental operators. Any unprobed environmental operators affect the evolution of the individual SMEs but they do not play a role in the evolution of $P(\underline{\lambda})$. However, the equation given in [@Ral2011] generalizes naturally to include measurement inefficiencies and is given by, $$\begin{aligned}
\label{KS1}
dP_r(\underline{\lambda}) = &\sqrt{\eta_r}(\mathrm{Tr}(\hat{L}_r\rho_{c,\underline{\lambda}}+\rho_{c,\underline{\lambda}}\hat{L}^{\dagger}_r)-\mathrm{Tr}(\hat{L}_r\rho_c+\rho_c\hat{L}^{\dagger}_r)) \nonumber \\
&\times(dy_r(t)-\sqrt{\eta_r}\mathrm{Tr}(\hat{L}_r\rho_c+\rho_c\hat{L}^{\dagger}_r) dt)P(\underline{\lambda}) \end{aligned}$$ where $dP_r(\underline{\lambda})$ is the update to the probability density due to a measurement increment $dy_r(t)$ corresponding to the measured operator $\hat{L}_{r} $, such that $$\label{dP}
P(dy_r(t)|\underline{\lambda})=\frac{e^{(-(dy_r(t)-\sqrt{\eta}\mathrm{Tr}(\hat{L}_r\rho_{c,\underline{\lambda}}+\rho_{c,\underline{\lambda}}\hat{L}^{\dagger}_r))^2/dt)}}{\sqrt{2\pi dt}}$$ and the full conditional density matrix $\rho_c$ is given by, $$\label{full_rho}
\rho_c = \int_{\underline{\lambda}}P(\underline{\lambda})\rho_{c,\underline{\lambda}} d{\underline{\lambda}}$$
Sequential Monte Carlo Methods
==============================
Sequential Monte Carlo methods originate in the field of multi-target tracking [@Gor1993]), but have been adopted and generalized to form a set of very efficient methods for parameter and state estimation in classical signal processing and nonlinear filtering. SMC methods are sometimes referred to as [*particle filters*]{}, but particle filters are a special case of the general approach. An SMC method relies on the approximation of a continuous probability distribution by a finite set of points (or particles) which sample the parameter space. The importance of each sample point changes in response to (is conditioned by) the measurements associated with the parameters being estimated, and the sample points can be periodically resampled to concentrate sampling towards regions of higher relative probability.
In particle filters, the sample points are allowed to evolve according to some dynamical process, generating a time dependent history or a track within the parameter space. In the example presented in this paper, the parameters are selected to be constant and another SMC method is more suitable. We adopt an approach used recently for parameter estimation in classical differential equations [@Gre2017], which is an example of a [*Sequential Monte Carlo sampler*]{}. This approach is particularly well-suited to the estimation of fixed parameters; however, the SMC sampler used here still embodies all of the key features of a general SMC method: sampling, conditioning/updating, and resampling. A number of very approachable tutorials and introductions to particle filters and general SMC methods have been published. For example, a comprehensive guide to SMC methods and their applications is available in [@Dou2001], a mathematical introduction is given in [@Cap2007], and a widely cited tutorial to particle filters and SMC methods is contained in [@Aru2002].
Formally, an SMC method approximates a (classical) expectation for a function $h(x)$ over a probability distribution $p(x)$ defined on some parameter space $\Lambda$, $x \in \Lambda$, given by $$\bar{h} = \int h(x)p(x)dx$$ using a finite sum of a set of points $x^{(i)}$ ($i = 1\ldots N$) drawn from $p(x)$, which is known as the [*target distribution*]{}. The expectation value for an arbitrary function can be approximated by, $$\bar{h} \simeq \frac{1}{N}\sum_{i =1}^{N}h(x^{(i)})$$ The larger the number of sample points, the better the approximation – in fact, under reasonable assumptions, the variance of the error in $\bar{h}$ can be shown to scale as 1/N in any number of dimensions [@Pre2007]. The problem is that, in most practical cases, the probability distribution is unknown. It needs to be estimated from a sequence of measurements. To do this, another distribution, the [*proposal distribution*]{} $q(x)$, is introduced such that [@Cap2007], $$\begin{aligned}
\bar{h}& = &\int h(x)q(x)\frac{p(x)}{q(x)}dx=\int h(x)q(x)w(x)dx \nonumber \\
&&\simeq\sum_{i =1}^{N}\frac{w^{(i)}}{\sum_{j =1}^{N}w^{(j)}}h(x^{(i)}) \end{aligned}$$ where $w(x)=p(x)/q(x)$ and the $w^{(i)}$’s are (unnormalized) weights associated with each sample point. In our case, each sample point is associated with a parameter value or a vector of values for each of the parameters being sought, $w^{(i)} \leftrightarrow \underline{\lambda}^{(i)}$, where $\underline{\lambda}^{(i)} \in \Lambda$. Initially, the sample points are randomly selected from a prior distribution, which covers the entire range of possible parameter values, and are given a uniform weight. As new measurements are added, the accuracy of the estimated quantity $\bar{h}$ is improved by updating the weights associated with the particles to reflect the new information that the measurement contains. Some weights are increased and some weights are reduced when the measurement supports or contradicts the corresponding sample point, respectively.
The values $w^{(i)}$ are referred to as ‘weights’ rather than probabilities because, although they are related to probabilities, they are not necessarily normalized after each time step and do not necessarily sum to one. In practice, it is convenient to normalize the weights after each time step. Here, we denote the normalized weights by $\tilde{w}^{(i)}$ [@Gre2017]. When sample points have very low weight, and hence very low probability, they can be removed and replaced with alternative particles, but this resampling process must be done carefully so as to ensure that the statistical quantities remain unbiased and will converge efficiently to the desired values.
The proposal distribution should be simple to calculate and different choices of $q(x)$ are used in different variants of the SMC approach [@Dou2001; @Cap2007; @Aru2002]. The secret to working with SMC methods is to pick a suitable proposal distribution to solve the problem in a robust manner using limited computational resources. In particular, a good choice of proposal distribution allows classical parameters to be estimated significantly more efficiently than when using an enumerative or grid based method [@Aru2002], as was used for a one parameter Hybrid SME problem in [@Ral2011; @Bas2015].
For our Hybrid SME problem, we start by selecting an initial set of sample points in the parameter space using a prior distribution and initialize an SME (\[sme1a\]) for each of the sample points. For the examples shown below, the quantum state of the system is initialized to be a thermal mixed state, and the prior distribution is chosen to be uniform over some finite range within which the true parameter values are known to lie. An accurate initial prior distribution can significantly reduce the number of particles required by the SMC sampler, but in many situations the prior is not well defined. Once the points have been selected, the weights are initialized with $w^{(i)}_0 = \tilde{w}^{(i)}_0 = 1/N$.
For each time step, the individual SMEs are integrated using the increment (\[sme1a\]) found using the parameter value $\underline{\lambda}^{(i)}$ associated with the particle. A corresponding measurement probability is found from (\[dP\]) and used to update the (unnormalized) particle weight $w^{(i)}_{k-1,r} \rightarrow w^{(i)}_{k,r}$ using [@Gre2017] $$\label{weights}
w^{(i)}_{k,r}=p(dy_r(t_k)|\underline{\lambda}^{(i)})w^{(i)}_{k-1,r}$$ for the $k$’th measurement from measurement operator $\hat{L}_r$ at time $t_k$. All particles are updated after the measurement increment and then weights are normalized.
Resampling to generate new particles only occurs when the distribution of weights amongst the particles is such that the effective sample size (or the effective number of particles) $N_{eff} = 1/(\sum_i (\tilde{w}^{(i)})^2)$ falls below some threshold value – indicating that the weight is being concentrated in a relatively small number of particles and a significant number of particles have low weight and do not contribute to the estimates; a problem known in the SMC literature as [*sample impoverishment*]{} or [*weight degeneracy*]{} [@Cap2007]. It is known that the variance of the weight distributions across different realizations of the SMC sampler is guaranteed to grow with each time step [@Aru2002]. However, since a given realization (i.e. one run of the algorithm) does not have access to the ensemble of all possible realizations, it is convenient to monitor something that can be computed from a single realization. The effective sample size is well established as such a quantity [@Cri2002] and it can be considered to be a noisy measurement of the (inverse of the) variance. Between resampling events, the variance of the weight distributions will (on average) increase and the effective sample size will decrease. While the precise threshold value used is a somewhat arbitrary choice for the algorithm designer, it is common (across the vast range of applications of SMC samplers and particle filters) to consider threshold values between $N/10$ and $N/2$. In the cases shown below the threshold value for $N_{eff}$ was set to be $N/2$ [@Gre2017].
When the particles are resampled, the new candidate values $\tilde{\underline{\lambda}}$ are sampled from the distribution formed from the current particle weights. The particles with the highest weights are more likely to be selected, although the particles with relatively low weights still have a chance of being selected. The new particle parameter values are then selected using the distribution $q(\tilde{\underline{\lambda}}|\underline{\lambda}^{(i)})={\cal N}(\tilde{\underline{\lambda}};\underline{\lambda}^{(i)},\Sigma)$, where ${\cal N}(x;\mu_{x},S)$ is a normal distribution with mean $\mu_{x}$ and covariance $S$, and $\Sigma$ is related to the covariance matrix associated with the current particle weights, $\Sigma_k$. The role of $q(\tilde{\underline{\lambda}}|\underline{\lambda}^{(i)})$ is to select new points, $\tilde{\underline{\lambda}}^{(i)}$, around the current particles with large weights, but not at exactly the same point. In this paper, we use a [*defensive strategy*]{} [@Gre2017; @Hes1995], where 90% of resampled points use a covariance which is 10% of the current covariance, $\Sigma=0.1\Sigma_k$, and 10% of the resampled points use the full covariance matrix $\Sigma=\Sigma_k$. This allows for small perturbations in parameter space around the high weight sample points, including the correlations between different parameters seen in the covariance matrix, and a small number of large excursions, to explore more of the parameter space than is currently being covered by the sample points with large weight. There are two specific design considerations relevant to the choice of the distribution, $q$. The first is to ensure that having more samples will give rise to more accurate estimates of quantities of interest, which is manifest empirically as robustness. Put simply, this demands that samples are proposed in a way that explores possible but potentially low probability states. The second is to ensure that the SMC method is computationally efficient, i.e. that it gets as accurate an estimate as is possible with a given number of samples. This demands that samples are placed in high probability areas. A defensive proposal is an advanced, but relatively standard technique, used in particle filters and SMC samplers, that combines robustness with efficiency by having two elements to the proposal, one that is designed to ensure the sampler is robust and the other that is designed to ensure that it is efficient.
When the new sample points have been selected, they are initially assigned the weight $\check{w}^{(i)}_0 = 1/N$ and then the unnormalized weight for the new candidate points is calculated reusing the entire record of measurement increments, $$\label{resample}
\check{w}^{(i)}_k=\frac{\prod_{r,k}p(dy_r(t_k)|\tilde{\underline{\lambda}}^{(i)})}{\prod_{r,k}p(dy_r(t_k)|\underline{\lambda}^{(i)})}\check{w}^{(i)}_0$$ Once all of the new weights have been recalculated they can be renormalized, and the integration of the Hybrid SME can continue as normal. Where the new sample weights are still degenerate and the effective number of particles is still below the threshold value, the resampling (and recalculation) needs to be performed again.
\[process\] ![\[fig:SMCSteps\] (Color online) Schematic process, showing the main steps in the SMC Sampler for a one parameter example.](SME_SMC_Fig1 "fig:"){width="1.0\hsize"}
To summarize, the SMC method used here can be described in the following five steps:
(a) Initialize the individual density matrices $\rho_{c,\underline{\lambda}^{(i)}}$ using thermal mixed states, and select a set of classical parameter values (sample points or particles) $\underline{\lambda}^{(i)}$ using a uniform distribution covering the full range of possible parameter values, and assign uniform weights to each of these $\tilde{w}^{(i)}_0 = 1/N$.
(b) Evolve the quantum state using the individual SMEs (\[sme1a\]), using the corresponding classical parameters and updating their weights using (\[weights\]). Continue this evolution until $N_{eff}$ drops below the threshold value.
(c) If $N_{eff}$ is below the threshold value, the classical parameter values are resampled using a cumulative probability distribution calculated from the particle weights. This resampling creates a new set of particles/sample points, where the classical parameters are selected around the ‘parent’ values. The defensive strategy introduces small perturbations around the parent values and the occasional large perturbation to explore a wider parameter space – the new weights associated with each of the new parameter values/sample points are uniformly distributed at this point.
(d) Once the new values have been selected, the complete evolution of each quantum state is recalculated using new initial thermal states and the individual SMEs (using the same measurement record), and the uniform weights from step (c) are recalculated using (\[resample\]).
(e) Return to step (b) with evolution of the quantum state and weights determined by the individual SMEs and the weight update (\[weights\]), until $N_{eff}$ drops below the threshold value again, at which point the resampling step (c) and the re-weighting step (d) are again required.
A schematic example of the estimation process for a one dimensional parameter example is shown in Figure 1. In this example, it is possible to see that the initial uniform weighting of the particles evolves so that the relatively large number of particles below a parameter value of $1.0$ carry very little weight, and the distribution of particles immediately after resampling is concentrated more towards the values above 1.0. The re-weighted parameter values shown in (d) represent a better approximation to the underlying probability distribution than those shown in (b), which contains significant gaps towards the peak of the distribution. For a more detailed description of the implementation, a full description of the SMC method is given as pseudo-code in [@Gre2017].
The recalculation over the entire measurement history is an unfortunate, but necessary, computational cost in the SMC sampler. Recalculating the weights for the entire history of measurement increments will often take a significant amount of time. However, the need to regularly resample the entire set of particles reduces as the distribution of the particles improves to reflect the information contained in the measurements [@Gre2017]. This means that the computational load introduced is biased towards the start of the calculation of a quantum trajectory. In addition, for resampled points very close to the parent particles, some approximations are possible based on the fact that the ratio between the products in (\[resample\]) is very close to one. It is not possible to remove the recalculation entirely however without constraining the resampled parameter values and therefore not exploring the full parameter space.
Example System – Duffing Oscillator
===================================
The properties of the quantum trajectories generated by the Duffing oscillator have been studied extensively in terms of the appearance of chaotic behavior from quantum systems in the classical limit [@Sch1995; @Bru1996; @Bru1997; @Hab1998; @Bha2000; @Sco2001; @Bha2003; @Eve2005; @Eas2016; @Pok2016; @Ral2017], but it is also a model used for a number of other practical systems where quantum effects in classical nonlinear systems are of interest. For example, it has been used to describe the motion of a levitated particle in an electromagnetic trap [@Gie2013; @Gie2015], and is the basis for the analysis of the properties of vibrating beam accelerometers [@Aik2001; @Mes2009; @Agr2013]. The Hamiltonian for the Duffing oscillator can be written in the general form, using dimensionless position and momentum operators $\hat{q}$ and $\hat{p}$, $$\label{DuffHam}
\hat{H}(\underline{\lambda}) = \frac{1}{2}\hat{p}^2+\frac{1}{2}\omega^2\hat{q}^2+\frac{1}{4}\mu\hat{q}^4+g\cos(t)\hat{q}+\frac{\Gamma}{2}(\hat{q}\hat{p}+\hat{p}\hat{q})$$ where the vector $\underline{\lambda}=(\omega,\mu,g)$ contains the three Hamiltonian parameters of interest: the natural (linear) oscillation frequency $\omega$, the nonlinear coefficient $\mu$, and the strength of the external driving term $g$. The measurement is applied via a Linblad operator $\hat{L}=\sqrt{2\Gamma}\hat{a}$, and $\hat{a}$ is the harmonic oscillator lowering operator so that $\hat{q}=(\hat{a}^{\dagger}+\hat{a})/\sqrt{2}$ and $\hat{p}=i(\hat{a}^{\dagger}-\hat{a})/\sqrt{2}$ with $[\hat{a},\hat{a}^{\dagger}]=1$ and $\hbar=1$. We fix the measurement strength so that $\Gamma=0.125$ for all of the results presented here. The final term in the Hamiltonian is included because, in combination with the dissipative measurement process, it generates linear damping in momentum. This is a useful numerical addition because it keeps the phase space contained, thereby restricting the numbers of states required in the simulation, without affecting the underlying physics.
The numerical integration of the individual SMEs uses a method developed by Rouchon and colleagues [@Ami2011; @Rou2015] specifically for stochastic master equations. This method has been demonstrated to provide significant benefits in terms of accuracy versus computational resources when compared to standard methods, such as Milstein’s method [@Mil1995], for both systems involving small numbers of basis states [@Rou2015] and large numbers of basis states [@Ral2016]. We also employ a moving basis method used by Schack, Brun and Percival [@Sch1995; @Bru1996] to shift basis states to be centred on the current expectation value of the state. Although not strictly necessary [@Sch1995; @Bru1996], we shift the basis after each time step. This comes at a computational cost but it also ensures that the number of basis states employed is minimized. Once the evolution of the individual SMEs has been calculated, using the appropriate set of parameters, the combined density operator is calculated by averaging over all of the individual states, weighted appropriately by the particle weights.
The increment to the state $\rho_{{\mbox{\scriptsize c}},\underline{\lambda}}^{(n)}$ for the time step from $t_n = n\Delta t$ to $t_{n+1} = (n+1)\Delta t$ is calculated using $$\label{sme2}
\rho_{{\mbox{\scriptsize c}},\underline{\lambda}}^{(n+1)}= \frac{\hat{M}_{n,\underline{\lambda}}\rho_{{\mbox{\scriptsize c}},\underline{\lambda}}^{(n)}\hat{M}_{n,\underline{\lambda}}^{\dagger}+(1-\eta)\hat{L}\rho_{{\mbox{\scriptsize c}},\underline{\lambda}}^{(n)}\hat{L}^{\dagger}\Delta t }
{\mathrm{Tr}\left[\hat{M}_{n,\underline{\lambda}}\rho_{{\mbox{\scriptsize c}},\underline{\lambda}}^{(n)}\hat{M}_{n,\underline{\lambda}}^{\dagger}+ (1-\eta)\hat{L}\rho_{{\mbox{\scriptsize c}},\underline{\lambda}}^{(n)}\hat{L}^{\dagger}\Delta t \right]}$$ where $\Delta \rho_{{\mbox{\scriptsize c}},\underline{\lambda}}^{(n)} = \rho_{{\mbox{\scriptsize c}},\underline{\lambda}}^{(n+1)}- \rho_{{\mbox{\scriptsize c}},\underline{\lambda}}^{(n)}$ and $\hat{M}_{n,\underline{\lambda}}$ is given by $$\begin{aligned}
\label{Mn1}
\hat{M}_{n,\underline{\lambda}} &=& I-\left(i\hat{H} +\frac{1}{2} \hat{L}^{\dagger}\hat{L}\right)\Delta t +\frac{\eta}{2}\hat{L}^2(\Delta W(n)^2-\Delta t) \nonumber\\
&& +\sqrt{\eta}\hat{L}\left(\sqrt{\eta}\mathrm{Tr}[\hat{L}\rho_{{\mbox{\scriptsize c}},\underline{\lambda}}^{(n)}+\rho_{{\mbox{\scriptsize c}},\underline{\lambda}}^{(n)}\hat{L}^{\dagger}]\Delta t +\Delta W(n)\right) \nonumber\end{aligned}$$ where the $\Delta W$’s are independent Gaussian variables with zero mean and a variance equal to $\Delta t$. Once the increment has been calcluated, center of the basis is moved to the new location of the state in phase space, as given by the expectation values of the phase space operators, $(q_{(n+1),\underline{\lambda}}, p_{(n+1),\underline{\lambda}}) = (\mathrm{Tr}[\hat{q}\rho_{{\mbox{\scriptsize c}},\underline{\lambda}}^{(n+1)}], \mathrm{Tr}[\hat{p}\rho_{{\mbox{\scriptsize c}},\underline{\lambda}}^{(n+1)}])$, using the displacement operator [@Sch1995; @Bru1996], $$\hat{D}(p_{(n+1),\underline{\lambda}},q_{(n+1),\underline{\lambda}})=\exp\left( i(p_{(n+1),\underline{\lambda}} \hat{q}- q_{(n+1),\underline{\lambda}} \hat{p})\right)$$ and the conditioned state in the shifted basis is given by $$\rho_{{\mbox{\scriptsize c}},\underline{\lambda}}^{(n+1)} \rightarrow \hat{D}(p_{(n+1),\underline{\lambda}},q_{(n+1),\underline{\lambda}})
\rho_{{\mbox{\scriptsize c}},\underline{\lambda}}^{(n+1)}\hat{D}(p_{(n+1),\underline{\lambda}},q_{(n+1),\underline{\lambda}})^{\dagger}$$
Results
=======
The Duffing Hamiltonian (\[DuffHam\]) has four classical parameters but we will fix the measurement strength so that $\Gamma=0.125$ and we will concentrate on the estimation of the other three parameters: the linear oscillator frequency $\omega$, the coefficient of the nonlinear term $\mu$, and the magnitude of the drive term $g$. The estimated values for these three parameters are denoted by $\tilde{\omega}$, $\tilde{\mu}$, and $\tilde{g}$ respectively. For all of the examples shown below, the actual values for parameter values were set to be $\omega = 1.2$, $\mu = 0.15$, and $g=3.0$. The numerical integration of the SMEs was performed using time steps $\Delta t = 2\pi/500$ so that there were 500 steps per period of the drive term. The individual SMEs for each particle/sample point used a moving basis with 15 harmonic oscillator states, and the composite state was calculated by combining the density matrices from the individual SMEs using (\[full\_rho\]), using a moving basis with 60 harmonic oscillator states.
Figure 2 shows two examples for the estimation of the linear oscillator frequency $\tilde{\omega}$. The examples correspond to the same stochastic record (i.e. the same realization) but with different measurement efficiencies. The blue lines correspond to the case where the measurement is 100% efficient (with $\eta = 1$). This shows a rapid convergence to the actual value, $\omega= 1.2$, within about 50-100 periods/cycles of the drive term. The 3 sigma errors predicted for the estimate are also shown, together with the resampling events as blue circles. The convergence is fairly rapid and the estimate is relatively stable once converged. The red line on the same figure shows an example where the measurement is inefficient, corresponding to a measurement efficiency of 40% or $\eta = 0.4$ (chosen to match the estimated efficiency reported in [@Web2014]). In this case, the convergence is much slower, indicating that the measurement record contains less information upon which a parameter estimate can be constructed. In this case, the estimated parameter value only stabilizes after around 150-200 cycles of the drive term, and the larger estimated errors indicate this increased uncertainty. In both cases shown, there are slight variations in the estimated values (seen around 200-250 cycles) but these are relatively small and are well within the estimated errors. In addition, where the estimation process takes longer, the number of resampling events (red circles) tends to increase and they often occur later in the process than the corresponding resampling events for efficient measurements, leading to increased computational demands to recalculate the weights after resampling. In addition to the estimates, Figure 2 also shows the purity of the full estimated quantum state for both cases as an inset. For efficient measurements, the conditioned quantum state purifies very rapidly (1-2 periods of the drive term) and remains pure throughout the estimation process. For inefficient measurements, the conditioned quantum state purifies somewhat but then the purity fluctuates between 0.8 and 0.9. The state remains mixed because information about the quantum state is being corrupted by extraneous noise. This is a characteristic of inefficient measurements in quantum systems, and it is not affected by, and does not itself affect, the classical parameter estimation process.
\[Fig\_2\] ![\[fig:LinearParameter\] (Color online) Examples of estimated values for the linear parameter ($\tilde{\omega}$) using SMC sampler with efficient measurements ($\eta = 1.0$, solid blue line) and inefficient measurements ($\eta = 0.4$, solid red line) with an actual linear parameter value $\omega = 1.2$ (solid black line) and 101 sample points (other parameters are given in the text). Three standard deviation errors are indicated in each case with dotted lines, and the resampling points are indicated by circles along the solid black line. The inset figure shows the purity values for the estimated state in each case.](SME_SMC_Fig2_wInset "fig:"){width="1.05\hsize"}
Figure 3 shows the evolution of the effective number of particles $N_{eff}$ as a function of time for the examples shown in Figure 2. The resampling events are marked on Figure 2 as large dots, but they are also seen in Figure 3 as large jumps in $N_{eff}$ after the resampling. The data in this figure is useful when optimizing the resampling parameters. It provides information regarding the average number of particles being used. An efficient SMC process would expect to have rapid fluctuations in $N_{eff}$ in the initial phases of the estimation process, with frequent resampling, which would become more gradual drops in $N_{eff}$ as the estimates improve. As time increases, and more measurements are added, the resampling events become less frequent, as is shown in Figures 2 and 3.
\[Fig\_3\] ![\[fig:LinearParameter2\] (Color online) Examples of the effective number of particles $N_{eff}$ for the estimates shown in Fig.2 for efficient measurements ($\eta = 1.0$, solid blue line) and inefficient measurements ($\eta = 0.4$, solid red line).](SME_SMC_Fig3 "fig:"){width="1.05\hsize"}
The estimation of the frequency of the linear oscillator term is relatively straightforward, and this is also found to be the case for the magnitude of the drive term $g$. Estimating the coefficient of the nonlinear term $\mu$ is more challenging however. When the external drive is very small, the Duffing oscillator will appear to be approximately linear and estimating the degree of nonlinearity is problematic. As the amplitude of the drive is increased, the system will explore more of the nonlinear potential and $\mu$ will become easier to estimate. This fact is reflected in the results obtained. For the parameter values selected, the drive term is sufficiently strong to explore the nonlinearity of the potential, but not sufficiently strong so as to require very large numbers of basis states or to make the estimation process easy compared to the other two parameters. An example of the estimation of the nonlinear coefficient is shown in Figure 4, where the convergence to a stable value takes much longer than either example shown in Figure 2, requiring over 500 periods of the drive term to stabilize the estimated value (note the different x-axis compared to Figure 2).
\[Fig\_4\] ![\[fig:NonLinearParameter\] (Color online) An example of estimated values for the nonlinear parameter ($\tilde{\mu}$) using SMC sampler with efficient measurements ($\eta = 1.0$, solid blue line) with an actual nonlinear parameter value $\mu = 0.15$ (solid black line) and 101 sample points (other parameters are given in the text). Three standard deviation errors are indicated in each case with dotted lines, and the resampling points are indicated by circles along the solid black line.](SME_SMC_Fig4 "fig:"){width="1.05\hsize"}
\[Fig\_5\] ![\[fig:MultiParameter\] (Color online) An example of values for all three parameters ($\tilde{\omega}$, $\tilde{\mu}$, and $\tilde{g}$) estimated simultaneously using SMC sampler with efficient measurements ($\eta = 1.0$, solid blue line) and 1001 sample points (other parameters are given in the text). Three standard deviation errors are indicated in each case with dotted lines, and the resampling points are indicated by circles along the solid black line.](SME_SMC_Fig5 "fig:"){width="1.05\hsize"}
Each of the examples shown in Figures 2 and 4 show the estimation of one parameter, the other parameters are assumed to be known. The estimation of one parameter is relatively straightforward and a value can be found using a grid-based method (as was the case in [@Ral2011] and [@Bas2015]). The number of particles required for the estimation of $\tilde{\omega}$ and $\tilde{\mu}$ is around 101 sample points in each of the SMC examples shown above. The number of resampling events is around 4-6 in the cases shown in Figures 2 and 4, and the maximum number of quantum trajectories that would need to be calculated is approximately equivalent to 200-300 trajectories on a fixed grid Hybrid SME. The expected errors for a fixed grid approach are related to the grid spacing, which is related to the initial range over which these grid points are initially distributed. For the cases considered here, with an initial distribution of points for a parameter value $\lambda$ between $0.5\lambda \leq \lambda^{(i)} \leq 1.5\lambda$. Assuming that the actual values of $\lambda$ are uniformly distributed across each interval, the expected error for a fixed grid approach with $N_{grid}$ points would be limited by $\sigma_{\lambda,grid} > (\lambda/N_{grid})/\sqrt{12}\simeq 0.1\%-0.15\%\lambda$. This value is achievable only in the long time limit and the actual error is likely to be significantly larger than this. In the examples given above, the SMC sampler produces parameter estimates with errors approaching this limit within a few hundred cycles. There is therefore a small but potentially significant benefit in using the SMC sampler method for one parameter estimation.
Moving from single to multiple parameter estimation presents a serious problem for grid-based methods. The number of points required scales exponentially in the number of dimensions to achieve the same accuracy. The error from a grid-based approach results from approximating an integral of functions in $D$ dimensions, where the error is $O((N_{grid})^{(-1/D)})$. The error for an SMC sampler comes from approximating the integral directly (using Monte-Carlo integration) and therefore is $O(1/N_{grid})$ whatever value $D$ takes [@Pre2007]. (See reference \[40\] for proofs for the convergence of SMC and particle filter based methods). So, for $D=1$, the two approaches offer similar scaling of error with $N_{grid}$, in higher dimensions, an SMC sampler will asymptotically outperform a grid-based method as $N_{grid}$ tends to infinity. Of course, differences in constants of proportionality mean that a computational benefit from using the SMC sampler in a small number of dimensions (number of parameters) is not guaranteed. Estimating all three parameters in our example, at a level of accuracy equivalent to the one parameter examples above, would require around ten million grid points, $(300)^3 = 9\times 10^6$. With SMC methods, this number is dramatically reduced.
Figure 5 shows an example of the simultaneous estimation of all three parameters using 1001 sample points. The values for $\tilde{\omega}$ and $\tilde{g}$ still converge rapidly whilst $\tilde{\mu}$ takes longer to establish a stable estimate. When comparing this with a grid based method, we note that the number of trajectories is larger than for the single parameter case and the number of resampling events is also increased, approximately 20 in the case shown in Figure 5. This is equivalent to a run-time for approximately 10,000 trajectories on a fixed grid. Using the same assumptions as before, this would give errors limited by $\sigma_{\underline{\lambda},grid} > (\lambda/\sqrt[3]{N_{grid}})/\sqrt{12}\simeq 1.5\% \lambda$. The errors found using the SMC sampler described above are nearly an order of magnitude smaller than this limit for one of the three parameters ($\omega$) and comparable for the remaining two parameters ($\mu$ and $g$). There is an additional benefit, in that the sample points not only provide estimates of the parameter values, they also provide information regarding the correlations between the different parameters. For the example shown in Figure 5, the mean vector and the estimated covariance matrix ($S$) are given by $$\left(\begin{array}{c} \tilde{\omega} \\ \tilde{\mu} \\ \tilde{g} \end{array} \right) = \left(\begin{array}{c} 1.1981 \\ 0.1557 \\ 3.0874 \end{array} \right)$$ $$\small{
S = \left( \begin{array}{ccc} 4.0308 \times 10^{-3} & -1.6449 \times 10^{-6} & 9.2270 \times 10^{-6} \\ -1.6449 \times 10^{-6} & 3.9795 \times 10^{-4} & -2.8254 \times 10^{-6} \\ 9.2270 \times 10^{-6} & -2.8254 \times 10^{-6} & 1.1586 \times 10^{-2} \end{array}\right)}$$ Note also that in Figure 5, the standard deviation of the linear parameter ($\tilde{\omega}$) is larger than in Figure 2 and the convergence is slower than for the single parameter case. This is partly due to the larger uncertainty generally in the three unknown parameters, and in part due to the slower convergence of the nonlinear parameter ($\tilde{\mu}$). The coupling between the parameters, shown by the non-negligible correlations shown in the covariance matrix, means that uncertainty in the nonlinear parameter increases the standard deviation of the other two parameters.
The use of an SMC sampler to estimate the Hamiltonian parameter values directly from the quantum trajectories is more efficient than an equivalent grid-based method but it still presents a computational challenge. Solving a single SME can be simplified using a stochastic integration method designed specifically for SMEs, like Rouchon’s method [@Ami2011; @Rou2015], and using efficient numerical tools, like moving basis states [@Sch1995; @Bru1996]. However, solving many simultaneous SMEs to determine the evolution of the particle weights still requires significant computational resources. The number of combinations of parameter values explored using the SMC sampler is significantly less than that required by a conventional grid-based method, but each sample point explored requires the full trajectory to calculated, or recalculated after resampling. The number of SMEs required to be calculated can be said to be [*relatively*]{} small but it is still not a trivial exercise. In their favor, SMC methods are amenable to parallelization [@Gre2017], since the evolution of SME and the recalculation of each trajectory after resampling are largely independent processes and can be distributed simply across a number of processors. However, at present, it is more likely that this type of technique is more likely to be used for post-processing experimental data rather than as part of an on-line closed-loop control system.
Conclusions
===========
Continuous quantum measurements, and their associated stochastic master equations (SMEs), provide a means to monitor the dynamical evolution of a quantum system and to provide an estimate of the underlying quantum state. In addition, the quantum trajectories resulting from the integration of stochastic master equations contain useful information about the parameters that govern the evolution of the system. Hybrid stochastic master equations provide a means to extract the information regarding these classical parameters. Hybrid SMEs involve running many parallel SMEs, each one having a different value for the parameter (or parameters). The classical probabilities attached to the individual SMEs and the associated parameter values can then be found by integrating a Kushner-Stratonovich equation. This classical estimation process is numerically costly, and is even more so when estimates are required for multiple parameters. This paper has demonstrated how such estimates can be found using a technique taken from classical state estimation and nonlinear filtering, a Sequential Monte Carlo (SMC) sampler. The SMC sampler used in this paper has been demonstrated to allow the simultaneous estimation of three Hamiltonian parameters, together with their statistical correlation and the associated quantum trajectories, in a computationally tractable form, with a relatively small number of candidate parameter values and parallel SMEs.
Even with such methods, the computational task in solving the Hybrid SME is formidable, and is currently beyond the point where it could be used as part of a closed-loop quantum control system. At present, the strength of such techniques is in the ability to post-process experimental measurement data to verify the quantum states used in an experiment but also to provide an independent, in-situ means to check the parameters that govern their evolution.
*Acknowledgments:* JFR would like to thank the US Army Research Laboratories (contract no. W911NF-16-2-0067). JFR would also like to thank Hendrik Ulbricht and Peter Barker for helpful and informative discussions.
[99]{}
V. P. Belavkin [*Reports on Mathematical Physics*]{} [**45**]{}, 353 (1999), and references contained therein. H. M. Wiseman, G. J. Milburn, ‘Quantum Measurement and Control’ (Cambridge University Press, Cambridge, 2010). K. Jacobs ‘Quantum Measurement Theory and Its Applications’ (Cambridge University Press, Cambridge, 2014). K. W. Murch, S. J. Weber, C. Macklin, I. Siddiqi, [*Nature*]{} [**502**]{}, 211 (2013). S. J. Weber, A. Chantasri, J. Dressel, A. N. Jordan, K. W. Murch, I. Siddiqi, [*Nature*]{} [**511**]{}, 570 (2014). P. Six, P. Campagne-Ibarcq, L. Bretheau, B. Huard, P. Rouchon, [*54th IEEE Conference on Decision and Control Conference (CDC)*]{} (2015). P. Campagne-Ibarcq, P. Six, L. Bretheau, A. Sarlette, M. Mirrahimi, P. Rouchon, B. Huard, [*Physical Review X*]{} [**6**]{}, 011002 (2016). P. Smith, J. E. Reiner, L. A. Orozco, S. Kuhr, H. M. Wiseman, [*Physical Review Letters*]{} [**89**]{}, 133601 (2002). S. Brakhane, W. Alt, T. Kampschulte, M. Martinez-Dorantes, R. Reimann, S. Yoon, A. Widera, D. Meschede, [*Physical Review Letters*]{} [**109**]{}, 173601 (2012). Kuban A. Kubanek, M. Koch, C. Sames, A. Ourjoumtsev, P. W. H. Pinkse, K. Murr, G. Rempe, [*Nature*]{} [**462**]{}, 898, (2009). D.J. Wilson, V. Sudhir, N. Piro, R. Schilling, A. H. Ghadimi, and T.J. Kippenberg, [*Nature*]{}, [**524**]{}, 325 (2015). V. Sudhir, D.J. Wilson, R. Schilling, H. Sch[ü]{}tz, S. A. Fedorov, A. H. Ghadimi, A. Nunnenkamp, and T.J. Kippenberg, [*Physical Review X*]{} [**7**]{}, 011001 (2017). R. Vijay, C. Macklin, D.H. Slichter, S.J. Weber, K.W. Murch, R. Naik, A. N. Korotkov, I. Siddiqi, [*Nature*]{} [**490**]{}, 77 (2012). D. Rist[è]{},C.C. Bultink, K.W. Lehnert, and L. DiCarlo, [*Physical Review Letters*]{} [**109**]{}, 240502 (2012). J. F. Ralph, K. Jacobs, C. D. Hill, [*Physical Review A*]{} [**84**]{}, 052119 (2011). P. L. Green, S Maskell, [*Mechanical Systems and Signal Processing*]{}, [**93**]{}, 379Ð396 (2017). I. L. Chuang, M. A. Nielsen, [*Journal Modern Optics*]{} [**44**]{}, 2455 (1997). J. Gambetta, H. M. Wiseman, [*Physical Review A*]{} [**64**]{}, 042105 (2001). F. Verstraete, A. C. Doherty, H. Mabuchi, [*Physical Review A*]{} [**64**]{}, 032111 (2001). J. K. Stockton, J. M. Geremia, A. C. Doherty, H. Mabuchi, [*Physical Review A*]{} [**69**]{}, 032109 (2004). M. Tsang, [*Physcial Review Letters*]{} [**102**]{}, 250403 (2009). M. Tsang, [*Physical Review A*]{} [**80**]{}, 033840 (2009). M. Tsang, [*Physical Review A*]{} [**81**]{}, 013824 (2010). M. Tsang, H.M. Wiseman, C.M. Caves, [*Physical Review Letters*]{} [**106**]{}, 90401 (2011). A Negretti, K. [M[o]{}lmer]{}, [*New Journal Physics*]{} [**15**]{}, 125002 (2013). D.W. Berry, M. Tsang, M.J.W. Hall, H.M. Wiseman, [*Physical Review X*]{} [**5**]{}, 031018 (2015). H. Bassa, S.K. Goyal, S.K. Choudhary, H. Uys, L. [Dió]{}si, T. Konrad, [*Physical Review A*]{} [**92**]{}, 032102 (2015). L. Cortez, A. Chantasri, L.P. García-Pintos, J. Dressel, A.N. Jordan, [*Physical Review A*]{} [**95**]{}, 012314 (2017). B.P. Abbott, R. Abbott, T.D. Abbott, *et al.*, [*Physical Review Letters*]{} [**116**]{}, 061102 (2016). F. Yang, A.J. Koll[á]{}r, S.F. Taylor, R.W. Turner, B.L. Lev, [*Physical Review Applied*]{} [**7**]{}, 034026 (2017). N. Gordon, D. Salmond, A. F. Smith, [*IEE Proceedings F Radar Signal Processing*]{}, [**140**]{}, 107Ð113 (1993). S. Blackman, ‘Multiple Target Tracking with Radar Applications’ (Artech House, 1986) Y. Bar-Shalom, X.R. Li, T. Kirubarajan, ‘Estimation with Applications to Tracking and Navigation’ (Wiley & Sons, 2001). J. F. Ralph, ‘Target Tracking’ in ‘Encyclopedia of Aerospace Engineering’, Vol.5, Ch 251, eds. R. Blockley, W. Shyy (Wiley & Sons, 2010). H. M. Wiseman, A. Doherty, [*Physical Review Letters*]{} [**94**]{}, 070405 (2005). A. Doucet, N. De Freitas, and N. Gordon, Eds., ‘Sequential Monte Carlo Methods in Practice’ (Springer, New York, 2001). O. Cappé, S. J. Godsill, E. Moulines, [*Proceedings of the IEEE*]{} [**95**]{}, 899 (2007). M. Arulampalam, S. Maskell, N. Gordon, T. Clapp, [*IEEE Transactions on Signal Processing*]{} [**50**]{}, 241-254, (2002). W. H. Press, S. A. Teukolsky, W. T. Vetterling, B. P. Flannery, “Section 7.9.1 Importance Sampling" in “Numerical Recipes: The Art of Scientific Computing (3rd ed.)" (Cambridge University Press, New York, 2007). D. Crisan, A. Doucet. [*IEEE Transactions on signal processing*]{} [**50**]{}, 736-746 (2002). T. Hesterberg, [*Technometrics*]{}, [**37**]{}, 185-194 (1995). R. Schack, T. A. Brun, I. C. Percival. [*Journal of Physics A: Mathematical and General*]{} [**28**]{}, 5401 (1995). T. A. Brun, I. C. Percival, R. Schack, [*Journal of Physics. A: Mathematical and General*]{} [**29**]{} 2077 (1996). T. A. Brun, N. Gisin, P. F. O’Mahony, M. Rigo, [*Physics Letters A*]{} [**229**]{} 267-272 (1997). S. Habib, K. Shizume, W. H. Zurek, [*Physical Review Letters*]{} [**80**]{}, 4361 (1998). T. Bhattacharya, S. Habib, K. Jacobs, [*Physical Review Letters*]{} [**85**]{} 4852 (2000). A. J. Scott, G. J. Milburn, [*Physical Review A*]{} [**63**]{}, 042101 (2001). T. Bhattacharya, S. Habib, K. Jacobs, [*Physical Review A*]{} [**7**]{} 042103 (2003). M. J. Everitt, T. D. Clark, P. B. Stiffell, J. F. Ralph, A. R. Bulsara, C. J. Harland. [*New Journal of Physics*]{} [**7**]{} 64 (2005). J. K. Eastman, J. J. Hope, A. R. R. Carvalho. [*Emergence of chaos controlled by quantum noise*]{}, arXiv:1604.03494 (2016). B. Pokharel, P. Duggins, M. Misplon, W. Lynn, K. Hallman, D. Anderson, A. Kapulkin, A. K. Pattanayak, [*Dynamical complexity in the quantum to classical transition*]{}, arXiv:1604.02743 (2016). J. F. Ralph, K. Jacobs, M. J. Everitt, [*Physical Review A*]{} [**95**]{}, 012135 (2017). J. Gieseler, L. Novotny, R. Quidant, [*Nature Physics*]{}, [**9**]{}, 806 (2013). J. Gieseler, L. Novotny, C. Moritz, C. Dellago, [*New Journal of Physics*]{}, [**17**]{}, 045011 (2015). M. Aikele, K. Bauer, W. Ficker, F. Naubauer, U. Prechtel, J. Schalk, H. Seidel, [*Sensors and Actuators A*]{} [**90**]{}, 161-167 (2001). R. M. C. Mestrom, R. H. B. Fey, H. Nijmeijer, [*IEEE/AMSE Transactions on Mechatronics*]{} [**14**]{}, 423-433 (2009). D. K. Agrawal, J. Woodhouse, A. A. Seshia, [*IEEE Trans. on Ultrasonics, Ferroelectrics and Frequency Control*]{}, [**60**]{}, 1646-1659 (2013). H. Amini, M. Mirrahimi, P. Rouchon, [*in Proc. 50th IEEE Conf. on Decision and Control*]{}, pp. 6242-6247 (2011). P. Rouchon, J. F. Ralph, [*Physical Review A*]{} [**91**]{}, 012118 (2015). G.N. Milstein, ‘Numerical Integration of Stochastic Differential Equations’ (Springer, Berlin, 1995). J. F. Ralph, K. Jacobs, J. Coleman, [*Physical Review A*]{} [**94**]{}, 032108 (2016).
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'We investigate how the integrability conditions for conformal anomalies constrain the form of the effective action in even-dimensional quantum geometry. We show that the effective action of four-dimensional quantum geometry (4DQG) satisfying integrability has a manifestly diffeomorphism invariant and regularization scheme-independent form. We then generalize the arguments to six dimensions and propose a model of 6DQG. A hypothesized form of the 6DQG effective action is given.'
---
qqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqq = qqqqqqqqqqqqq\
hep-th/0012053\
[*Institute of Particle and Nuclear Studies, High Energy Accelerator Research Organization (KEK),*]{}\
[*Tsukuba, Ibaraki 305-0801, Japan*]{}
Introduction
============
Since quantum geometry (QG) is defined by functional integrations over the metric fields, diffeomorphism invariance in QG can equivalently be described as an invariance under any change of the background metric. This background-metric independence includes an invariance under a conformal change of the background metric. Thus, in even-dimensional QG well-defined on the background metric [@p]–[@or], conformal anomalies [@cd]–[@ii] play an important role. Therefore, to preserve diffeomorphism invariance we must formulate an even-dimensional QG while considering that conformal anomalies always exist [@p]–[@or].
Background-metric independence in two dimensions implies that QG can be described as a conformal field theory [@kpz; @dk]. This idea can be generalized to an arbitrary numbers of even dimensions [@amm; @hs; @h2; @kmm; @or]. However, as recently studied [@hs; @h2], this generalization is not simple, because the traceless mode becomes dynamical in higher dimensions, so that higher-dimensional QG can no longer be described as a free theory. Furthermore, it has been found that the integrability condition of the conformal anomaly [@bcr; @bpb] introduces a strong constraint on even-dimensional QG [@r; @ft1; @h2].
In this paper we further consider how the integrability condition of the conformal anomaly affects even-dimensional QG. We also settle the problem of the regularization scheme dependence and show that the effective action has a manifestly diffeomorphism invariant and regularization scheme-independent form.
This paper is organized as follows. In the next section we present the fundamental idea of how to preserve diffeomorphism invariance in even-dimensional QG and review how such an idea is realized in exactly solvable 2DQG [@kpz; @dk; @s]. In $D \geq 4$ dimensions, the integrability condition of the conformal anomaly not only restricts interactions of matter fields to conformally invariant ones, but also reduces the number of indefinite coefficients in the gravity sector [@h2]. How the integrability condition affects 4DQG is rediscussed in section 3. We then show that the effective action can be written in a diffeomorphism invariant and scheme independent form. A generalization to six dimensions [@kmm; @or] is discussed in section 4. We show that Duff’s scheme [@duff1] is also useful to tame the trivial anomalies in six dimensions [@ds]–[@bcn]. We then propose a model of 6DQG that is based on the arguments concerning integrability made in the 4DQG case. Many indefinite coefficients that result from the existence of many curvature invariants are fixed by enforcing the integrability, and a hypothesized scheme-independent form of 6DQG effective action is given. Section 5 is devoted to conclusions and discussion.
We use the curvature convention in which $R_{\mu\nu}=R^{\lam}_{~\mu\lam\nu}$ and $R^{\lam}_{~\mu\s\nu}=\pd_{\s} \Gamma^{\lam}_{~\mu\nu} -\cdots$.
Conditions of Diffeomorphism Invariance
=======================================
In this section we briefly explain how to realize diffeomorphism invariance in even-dimensional QG.
QG is defined by functional integration over the metric field as Z= , \[z\] where $I$ is an invariant action and $X$ is a matter field. In this paper we consider a conformal scalar without self-interactions, for example. The measure of the metric field is defined by the invariant norm <dg,dg>\_g = d\^D x g\^g\^ (dg\_dg\_+udg\_dg\_) , where $D=2n$ and $u >-1/D$. This measure can be orthogonally decomposed into the conformal mode and the traceless mode as && <d , d >\_g = d\^D x (d )\^2 , \[cmoc\]\
&& <d h, d h>\_g = d\^D x tr (\^[-h]{} d \^h)\^2 . \[tmoc\] Here, the metric is decomposed as $g_{\mu\nu}=\e^{2\phi}\bg_{\mu\nu}$ and $\bg_{\mu\nu}=(\hg \e^h)_{\mu\nu}$, where $tr(h)=0$ [@kkn; @hs; @h2].
This definition is manifestly diffeomorphism invariant/background-metric independent. However, it is not well-defined because the measures of the metric fields defined by (\[cmoc\]) and (\[tmoc\]) have a metric dependence, represented by $\sq{g}$, in the measures themself, so that we must integrate this dependence when we quantize the conformal mode $\phi$.
Instead, we consider measures defined on the background metric as && <d , d >\_ = d\^D x (d )\^2 , \[moc\]\
&& <d h, d h>\_ = d\^D x tr (\^[-h]{} d \^h)\^2 . \[moh\] This replacement, however, violates diffeomorphism invariance. In fact, these norms conformally change under a general coordinate transformation generated by $\dl g_{\mu\nu} =g_{\mu\lam}\nabla_{\nu}\xi^{\lam}
+g_{\nu\lam}\nabla_{\mu}\xi^{\lam}$, which can be decomposed as && = \_\^ + \^\_ ,\
&& \_ = \_\_\^ +\_\_\^ -\_\_\^ , \[gct\] where the relation $\bnabla_{\lam} \xi^{\lam}= \hnabla_{\lam} \xi^{\lam}$ is used. Therefore, these measures produce conformal anomalies [@fujikawa] under the general coordinate transformation.
As a lesson from 2DQG [@dk; @s; @h1], in order to preserve diffeomorphism invariance, we must add an action $S$ as Z= , \[zz\] where the measures of the metric fields are now defined by (\[moc\]) and (\[moh\]).
Let us now briefly see how background-metric independence constrains the theory provided by (\[zz\]). Background-metric independence for the traceless mode represents the condition that $\hg$ and $h$ always appear in the combination $\bg=\hg\e^h$ in (\[zz\]) [@hs]. This condition guarantees, at most, that the effective action has an invariant form on the metric $\bg$.
Background-metric independence for the conformal mode requires that $S$ satisfies the Wess-Zumino condition [@wzc], defined by S(, ) =S(, )+S(-,\^[2]{}) . \[wz\] Such an action is obtained by integrating the conformal anomaly within the interval $[0,\phi]$. Hence it satisfies the initial condition $S(0,\bg)=0$ and has a local form. In this paper we call this local action the Wess-Zumino action, because condition (\[wz\]) is essential in the arguments concerning diffeomorphism invariance. In two dimensions it is usually called the Liouville action [@p]. The well-known non-local forms of the integrated conformal anomaly are called Polyakov action [@p] and Riegert action [@r] in two and four dimensions, respectively. Why we distinguish between the local and the non-local actions becomes clear below.
Although the Wess-Zumino condition fixes the form of $S$, some overall coefficients remain to be determined. These coefficients should be determined from the requirement of diffeomorphism invariance in a self-consistent manner. The process used to determine them is as follows.
Under the general coordinate transformation, $\dl I=0$, while the Wess-Zumino action is not invariant and produces a conformal anomaly. This property results from condition (\[wz\]). Diffeomorphism invariance is now realized dynamically in such a manner that $\dl S$ cancels conformal anomalies calculated with loop effects of the combined theory, ${\cal I}=S+I$. In other words, we consider the regularized 1PI effective action $\Gamma$ of the combined theory ${\cal I}$ and require $\dl \Gamma=0$ to determine $S$. This means that the tree action ${\cal I}$ is not manifestly invariant, but by including loop effects the effective action becomes an invariant form on the metric $g$.
Here, it is worth commenting on the difference in the Wess-Zumino action defined by (\[wz\]) and the non-local Polyakov/Riegert action. The former produces conformal anomalies under a general coordinate transformation, while the non-local Polyakov/Riegert action, which appears in the effective action due to loop effects, is generally defined by the condition that it produces conformal anomalies under a conformal change.
As an exercise, let us first discuss 2DQG coupled $N$ conformal scalars. The tree action in the conformal gauge is given by [@dk; @s; @h1] = d\^2 x ( \^\_ \_ + ) + I\_[GF+FP]{} + I\_M(X,) , where $I_M$ is the invariant action of the $N$ free scalars. The gauge-fixing term and the Faddeev-Popov (FP) ghost action are given by [@kato] I\_[GF+FP]{} = d\^2 x ( -i B\_(\^ -\^) +2 \^ b\_ \_ c\^ ) , where the reparametrization ghost $c^{\mu}$ is a contravariant vector. $B_{\mu\nu}$ and the anti-ghost $b_{\mu\nu}$ are covariant symmetric traceless tensors. The coefficient $b$ is uniquely determined by diffeomorphism invariance.
Consider the effective action of 2DQG, which has the form = [I]{}(,X,) + W() , where $W$ is a loop effect that depends only on $\bg$ because the measure is now defined on $\bg$. The condition of diffeomorphism invaraiance, $\dl \Gamma=0$, is now given by - d\^2 x + \_ W() =0 , where $\dl_{\om}\bg_{\mu\nu}=2\om \bg_{\mu\nu}$ and $\om=-\half \hnabla_{\lam}{\xi}^{\lam}$. Here, $\dl W = \dl_{\om}W$ because $W$ does not depend on the conformal mode $\phi$. The second term on the l.h.s is just the conformal anomaly of the theory ${\cal I}$.
From one-loop calculations using the tree action ${\cal I}$, we obtain the well-known non-local Polyakov action [@p], W() = d\^2 x , where $N$ comes from scalar matter fields and this becomes $N-26$ through the effect of the ghosts. The change in the coefficient from $N-26$ to $N-25$ is due to a contribution from the conformal mode.
As mentioned above, diffeomorphism invariance determines the coefficient $b$ uniquely as [@dk] b= . \[b\] Using the relation && - d\^2 x ( \^\_ \_ + ) + d\^2 x\
&& =d\^2 x R R , the effective action can be reexpressed in a manifestly invariant form, = d\^2 x R R + I\_M (X,g) . Here, we have used the fact that the matter action is conformally invariant, so that $I_M (X,\bg)=I_M (X,g)$.
4D Quantum Geometry
===================
Recently, we showed that there is a model of diffeomorphism invariant 4DQG [@hs; @h2]. This model has many advantages in physics. In particular, it is renormalizable and asymptotically free. Also, it is capable of solving the cosmological constant problem dynamically without any fine-tuning, [@t1; @am] and it naturally describes our four dimensional universe at the low-energy region and for large $N$. [^1] However, the unitarity problem remains unsolved. In this paper we do not discuss the unitarity problem, which is expected to be solved dynamically [@t2; @ft2; @kaku; @h2].
Tree action
-----------
The tree action of 4DQG [@hs] is given by a proper combination of the Wess-Zumino action [@r; @ft1] and the invariant action required by the integrability conditions discussed in [@h2] and also in the following subsection 3.3 as &&[I]{}= d\^4 x { + a [|F]{} + 2 b \_4 + b ( [|G]{}- )\
&& + (2a+2b+3c)[|R]{}\^2 + [L]{}\_[GF+FP]{} } + I\_[LE]{}(X,g) , \[tree\] where ${\cal L}_{GF+FP}$ contains the gauge-fixing term and the FP ghost Lagrangian defined below. The term $I_{LE}$ represents lower derivative actions which include actions of conformally invariant matter fields, the Einstein-Hilbert action, and the cosmological constant term. The lower derivative gravitational actions are treated in the perturbation of the massive constants [@am; @ao; @h2].
The invariants $F$ and $G$ are defined by F &=& R\_R\^-2R\_R\^ +R\^2 ,\
G &=& R\_R\^-4R\_R\^ +R\^2 . In four dimensions they are the square of the Weyl tensor and the Euler density, respectively. The operator $\Delta_4$ is the conformally covariant fourth-order operator [@r], \_4 = \^2 + 2 R\^\_\_ -R + (\^R)\_ , which satisfies $\Delta_4 = \e^{-4\phi}\bDelta_4 $ locally for a scalar.
In the above, we introduce the dimensionless coupling $t$ only for the traceless mode as $\bg_{\mu\nu}=(\hg \e^{th})_{\mu\nu}$ and consider the perturbation of $t$. The kinetic term of the conformal mode comes from the Wess-Zumino action. Since the invariant $R^2$ terms in the Wess-Zumino action and the invariant action cancel out in our model, the self-interactions of $\phi$ appear only in the lower derivative actions in the exponential form, which can be treated exactly, order by order in $t$ [@h2].
The gauge-fixing term and the FP ghost action are given by [@ft2; @bjs]. \_[GF+FP]{} = 2i B\^ N\_ \^ - B\^ N\_ B\^ -2i [c]{}\^ N\_ \^ [**\_B**]{} h\^\_[ ]{} , \[gfix\] where $\chi^{\nu}=\hnabla^{\lam}h^{\nu}_{~\lam}$, and $N_{\mu\nu}$ is a symmetric second-order operator. The BRST transformations are given by && [**\_B**]{} h\^\_[ ]{} = i { \^ c\_ +\_ c\^ - \^\_[ ]{} \_ c\^ + t c\^ \_ h\^\_[ ]{}\
&& + h\^\_[ ]{} ( \_ c\^ - \^ c\_ ) + h\^\_[ ]{} ( \^ c\_ - \_ c\^ ) + } ,\
&& [**\_B**]{} = i t c\^ \_ + i \_ c\^ , \[brst\]\
&& [**\_B**]{} [c]{}\^ = B\^ , B\^ = 0 ,\
&& [**\_B**]{} c\^ = i t c\^\_ c\^ . The first two of these equations are obtained by replacing $\xi^{\mu}/t$ in the equation for general coordinate transformation, (\[gct\]), with the contravariant vector ghost field $ic^{\mu}$. The kinetic term of the ghost action then becomes $t$ independent. This BRST transformation is nilpotent. Using this transformation, the gauge-fixing term and the FP ghost action can be written as ${\cal L}_{GF+FP}= 2i{\bf \dl_B}
\{ {\tilde c}^{\mu}N_{\mu\nu} (\chi^{\nu}
+ \frac{i}{2}\zeta B^{\nu}) \}$ [@ku].
The important property of this tree action is that it transforms under the general coordinate transformation (\[gct\]) as = d\^4 x { -a ( + ) -b - c } , \[vta\] where = - \_ \^ . \[om4\] In the case of the BRST transformation, $\xi^{\mu}$ is replaced by $itc^{\mu}$. [^2] The $\bBox \bR$ terms in (\[vta\]) depend on the regularization scheme. We use here Duff’s scheme [@duff1] of dimensional regularization characterized by the equations && \_ d\^D x F = (D-4) d\^D x ( F + R ) ,\
&& \_ d\^D x G = (D-4) d\^D x G . When we define the tree action ${\cal I}$, it is taken into account that Duff’s scheme will be used subsequently for computing loop effects of the effective action. As shown below, the scheme-dependent terms cancel out, and we obtain a scheme-independent effective action.
Effective action
----------------
As investigated in [@h2], [^3] the regularized effective action of the theory ${\cal I}$ has the following form: = [I]{}(X,, ) +V\_[NS]{}(,) + W\_F(,) + W\_G() +W\_[R]{}() . \[eff4\] Here, the first term on the r.h.s. is the tree action. $V_{NS}$, $W_F$, $W_G$ and $W_{\Box R}$ come from loop diagrams. The former represents corrections to the Wess-Zumino action, and the latter three represent corrections to the traceless mode $h$.
Let us first consider corrections to the traceless mode. Here, $W_F$ is the part that is associated with the conformally invariant counterterm of $\bF$; it can be determined by computing two-point diagrams of the traceless mode. In Duff’s scheme, it has the following scale-dependent form: W\_F(,) = d\^4 x { - \_ ( ) \^ - \^2 } . Here, the appearance of the $\bR^2$ term is due to our use of Duff’s scheme. $C$ is the Weyl tensor, and $\Delta^C_4 = \Box^2 + \cdots$ is an appropriate conformally covariant operator for the Weyl tensor. Although the explicit form of $\Delta^C_4$ is unknown, it is known that there is a function $W_F$ that satisfies the equation [@ddi; @ds; @deser] W\_F(,) =\_ W\_F(,) = d\^4 x ( + ) , where $\dl_{\om}\bg_{\mu\nu}=2\om \bg_{\mu\nu}$, with (\[om4\]). Thus, $W_F$ produces the type-B anomaly in the classification of [@ds].
The term $W_G$ in (\[eff4\]) is the part that is associated with the conformally invariant counterterm of $\bG$. It is called the non-local Riegert action, which produces the type-A anomaly, or the Euler density in the classification of [@ds], and it has the form W\_G() = d\^4 x { \_4 [|[G]{}]{} - \^2 } , where = G - R . As stated above, $W_G$ produces the type-A anomaly as W\_G() = \_ W\_G () = d\^4 x . \[vwg\] The $\bR^2$ term is needed to realize equation (\[vwg\]), which guarantees that $W_G$ does not have any contribution to two-point diagrams of the traceless mode $h$ in the flat background. This is consistent with the direct loop calculations of two-point diagrams of $h$. Hence, $W_G$ is related to $h^3 $ vetex corrections in the flat background.
The coefficients $f$ and $e$ are scheme independent. They can be expanded by the renormalized coupling $t_r$ as f=f\_0 + f\_1 t\_r\^2 + , e=e\_0 + e\_1 t\_r\^2 + . Here, $f_0$ and $e_0$ have already been computed using one-loop diagrams as f\_0 = --+ , e\_0 = + - , where the first term in each coefficient comes from $N$ conformal scalar fields [@duff1]. The second and the last terms come from the traceless mode [@ft2] and the conformal mode [@amm], respectively. The coefficients $f_1$ and $e_1$ are given by functions of $a$ and $b$, to which not only two-loop diagrams, but also one-loop (but order $t_r^2$) diagrams contribute [@h2].
The beta function for the coupling $t_r$ is given by $\b=\fr{f}{2} t_r^3$. Since $f_0$ is negative, 4DQG is asymptotically free. Here, note that, although background-metric independence implies an invariance under any confromal change of the background metric, the usual $\b$ function does not need to vanish. This is due to the fact that there exists a conformal anomaly, or the Wess-Zumino action.
The last term in (\[eff4\]) is a scheme-dependent part, defined by W\_[R]{}() = - d\^4 x \^2 . It is unknown whether this term is really necessary or not. In any case, the coefficient $u$ is at most order $t^2$, so that $u= u_1t_r^2 +\cdots$.
As computed in [@h2], the correction $V_{NS}$ is scale-independent, and it merely changes coefficients $a$ and $b$ in the tree action into ${\tilde a}=a(1+v_a)$ and ${\tilde b}=b(1+v_b)$, where $v_a$ and $v_b$ are order $t_r^2$ at the one-loop level. The implications of this fact are discussed in the following subsection.
Now, the conditions for diffeomorphism invariance are given by the following equations [@h2]: = f , = e , c = u . \[bmi4\] Since $f_1$ and $e_1$ are functions of $a$ and $b$, while $f_0$ and $e_0$ are constants independent of $a$ and $b$, we can solve these equations perturbatively, order by order in $t_r$. Note that the one-loop coefficients of $v_a$ and $v_b$ are related to the order $t_r^2$ coefficients, $f_1$ and $e_1$, of $W_F$ and $W_G$. This is reasonable, because the Wess-Zumino action originally comes from the measure, and thus is essentially a quantum effect. Thus, one-loop contributions given by quantizing the Wess-Zumino action are related to two-loop contributions.
Substituting the solutions of (\[bmi4\]) into the regularized effective action, the $\bR^2$ terms cancel out, and we obtain the scheme-independent and manifestly invariant effective action, [^4] = d\^4 x { - C\_ ( ) C\^ + } + I\_[LE]{}(X,g) . Here, the Weyl action $F$ is absorbed into the scale, $\mu$.
Two-loop integrability
----------------------
Here, we summarize the conditions of diffeomorphism invariance discussed in ref. [@h2].
The condition that a theory can be made diffeomorphism invariant is that in the effective action, there is no action which produces a term that does not appear in the variation of the tree action $\dl {\cal I}$, (\[vta\]). Namely, diffeomorphism invariance implies that the action W\_[R\^2]{}(,) = d\^4 x ( ) \[wr2\] is not allowed, because this action produces $\bR^2$ under a general coordinate transformation. Further, a scale-dependent action including the conformal mode $\phi$, for example V\_S (,,) = d\^4 x \_4 ( ) , \[wphi\] is not allowed, because this action cannot be absorbed into the Wess-Zumino action by changing coefficients $a$ and $b$, and it produces a term that is not in $\dl {\cal I}$ under a general coordinate transformation.
In general, parts of the effective actions, other than that which produces the type-B anomaly, must be indpendent of the scale $\mu$, as $V_{NS}$, $W_G$ and $W_{\Box R}$. The vanishing of $r$ and $s$, at least to order $t^2_r$, is demonstrated in a previous paper [@h2]. We give this demonstration in the following.
First, we expand $r$ and $s$ as $r=r_0 + r_1 t^2_r +\cdots$ and $s=s_0 + s_1 t^2_r +\cdots$. The vanishing of $r_0$ is guaranteed in our model because at this order, only conformally invariant vertices contribute to the one-loop diagrams. This is a consequence of the fact that the invariant $R^2$ terms with the coefficients $a$, $b$ and $c$ cancel out, so that self-interactions of the conformal mode $\phi$ do not appear in the tree action ${\cal I}$, except in the lower derivative terms, such as the cosmological constant in the exponential form. This fact also implies $s_0=0$, because there are no diagrams that contribute to $s_0$.
The vanishing of $s_1$ is proved directly by showing the finiteness of the self-energy diagram of $\phi$ [@h2]. Here, the fact that there are no interactions of $R^2$ is also essential. Note that we cannot explain this result by using conformal invariance, because conformal invariance does not forbid that there exists the counterterm of $\phi \bDelta_4 \phi$. It can be explained only by diffeomorphism invariance/background-metric independence.
The background-metric independence for the conformal mode implies that $W_{R^2}$ and $V_S$ are related in such a manner that $s=0$ implies $r=0$. Thus, $r_1=0$ is shown indirectly.
A more direct demonstration of $r_1=0$ is as follows. Since there are no self-interactions of $\phi$, two-loop diagrams that contribute to $f_1$, $g_1$ and $r_1$ can be derived from the conformally invariant vertices of $2b\phi \bDelta_4 \phi$ and $\fr{1}{t^2}\bF$, so that the contributions of two-loop diagrams to $r_1$ vanish. However, there are contributions from one loop (but order $t_r^2$) diagrams, which include the vertices of $a\bF\phi$, $b(\bG-\fr{2}{3}\bBox\bR)\phi$ and $\fr{1}{32}(2a+2b+3c)\bR^2$. Here, because these vertices, with the exception of the first one, are non-conformally invariant, we must pay attention to such one-loop contributions.
As shown in [@hs; @h1], the variation in the one-loop contributions to the effective action of our model is given by \_W\^[(1)]{}() = -2 [Tr]{}(\^[-]{}) , \[kernel\] where $\eps$ is a cutoff. The matrix operator ${\cal K}$ is defined by the kinetic term $\half \Phi^t {\cal K} \Phi$ on an arbitrary background-metric $\hg$, where $\Phi=(\phi, h^{\mu}_{~\nu}, X)$. The $t$-independent diagonal parts give the coefficients $f_0$ and $e_0$. The off-diagonal parts, as well as the $t$-dependent diagonal parts, give contributions of order $t^2$. Note that, unlike in the case of matter fields, we do not use the condition of conformal invariance for gravitational fields to derive this expression. We merely use the facts that ${\cal K}$ is a fourth-order operator and there are no self-interactions of the conformal mode. If there are the invariant $R^2$ term with the coefficients $a$, $b$ and $c$, we cannot describe $\dl_{\om}W^{(1)}$ in such a simple form, because we do not introduce the coupling for the conformal mode $\phi$. That $\dl_{\om}W^{(1)}$ is expressed in the simple form such as the r.h.s. of (\[kernel\]) is a general property of $2n$-th order operators in $2n$ dimensions, and such a quantity has been shown to be integrable [@hs; @h2]. Thus, our model satisfies $r_1=0$.
In four dimensions, integrability places strong constraints on QG. It seems that there is no 4DQG other than ours that overcomes the integrability conditions. Thus, 4DQG may be fixed uniquely according to the conformal matter content.
6D Quantum Geometry
===================
In this section we show that the arguments concerning integrability in 4DQG can be generalized to the six dimensional case. Since there are many curvature invariants in six dimensions, many indefinite coefficients appear in the definition of 6D action. However, we show below that many of them are fixed by the integrability.
Duff’s scheme in six dimensions
-------------------------------
Recently, six-dimensional conformal anomalies have been studied in detail [@bpb]–[@ii]. In this subsection we summarize the results of these studies and then show that we can also apply Duff’s scheme to the six-dimensional case.
In six dimensions there are 17 independent curvature invariants. We here use the following bases [@bpb; @bcn]: && K\_1=R\^3 , K\_2=RR\_R\^ , K\_3= RR\_R\^ ,\
&& K\_4=R\_\^[ ]{}R\_\^[ ]{}R\_\^[ ]{} , K\_5=R\_R\_R\^ , K\_6= R\_R\^\_[ ]{}R\^ ,\
&& K\_7=R\_\^[ ]{}R\_\^[ ]{}R\_\^[ ]{} , K\_8=R\_R\^R\_[ ]{}\^[ ]{} , K\_9= RR ,\
&& K\_[10]{}=R\_R\^ , K\_[11]{}=R\_R\^ , K\_[12]{}=R\^\_\_ R ,\
&& K\_[13]{}=(\_R\_)\^R\^ , K\_[14]{}=(\_R\_)\^R\^ ,\
&& K\_[15]{}=(\_R\_)\^R\^ , K\_[16]{}=R\^2 , K\_[17]{}=\^2 R .
The results for conformal anomalies are summarized as follows. There are ten independent integrable curvature invariants [@bcn]. They provide a basis for the conformal anomalies in six dimensions. In the classification of ref. [@ds], the type-A anomaly is unique and given by the Euler density, G\_6 = -K\_1 +12 K\_2 -3 K\_3 -16 K\_4 +24 K\_5 +24 K\_6 -4 K\_7 -8 K\_8 . Here, we normalize it as G\_6 = - \_[\_1 \_1 \_2 \_2 \_3 \_3]{} \^[\_1 \_1 \_2 \_2 \_3 \_3]{} R\^[\_1 \_1]{}\_[ \_1 \_1]{} R\^[\_2 \_2]{}\_[ \_2 \_2]{} R\^[\_3 \_3]{}\_[ \_3 \_3]{} .
There are three type-B anomalies. They are locally conformally invariant in six dimensions: F\_1 &=& K\_1 -K\_2 +K\_3 +K\_4 -K\_5 -K\_6 + K\_8 ,\
F\_2 &=& K\_1 -K\_2 +K\_3 +K\_4 -K\_5 -3 K\_6 + K\_7 ,\
F\_3 &=&- K\_1 + K\_2 - K\_3 -K\_4 +6 K\_5 +2 K\_7 -8K\_8\
&& + K\_9 -6K\_[10]{} +6 K\_[11]{} +3 K\_[13]{} -6K\_[14]{} +3K\_[15]{} . Here, $F_1$ and $F_2$ correspond to two independent combinations of the Weyl tensors, $C_{\a\mu\nu\b}C^{\mu\lam\s\nu}C_{\lam~~\s}^{~\a\b}$ and $C_{\a\b}^{~~\mu\nu}C_{\mu\nu}^{~~\lam\s}C_{\lam\s}^{~~\a\b}$, respectively. $F_3$ gives the kinetic term of the traceless mode, which is expressed, up to a total derivative term, as $ C_{\mu\a\b\gm} ( \Box \dl^{\mu}_{~\nu} + 4 R^{\mu}_{~\nu}
- \fr{6}{5}R \dl^{\mu}_{~\nu} ) C^{\nu\a\b\gm}$.
The other six combinations are given by && M\_5 = 6K\_6 -3K\_7 +12K\_8 +K\_[10]{} -7K\_[11]{} -11K\_[13]{} +12K\_[14]{} -4K\_[15]{} ,\
&& M\_6 = -K\_9 +K\_[10]{} +K\_[12]{} +K\_[13]{} ,\
&& M\_7 = K\_4 +K\_5 -K\_9 +K\_[12]{} +K\_[14]{} ,\
&& M\_8 = -K\_9 +K\_[11]{} +K\_[12]{} +K\_[15]{} ,\
&& M\_9 = K\_[16]{} ,\
&& M\_[10]{} = K\_[17]{} . These are classified as trivial conformal anomalies.
In order to treat the trivial anomalies $M_5 \cdots M_{10}$, unambiguously, we use dimensional regularization. Consider the conformal variations of the functions $G_6$, $F_1$, $F_2$ and $F_3$ defined by the combinations listed above. In $D$ dimensions we obtain the equations \_ d\^D x G\_6 = (D-6) d\^D x G\_6 and \_ d\^D x F\_i = (D-6) d\^D x ( F\_i +\_[n=5]{}\^[10]{}z\_[i,n]{} M\_n ) (i=1,2,3) , \[duff6\] where && \[z\_[1,5]{}, z\_[1,6]{}, z\_[1,7]{}, z\_[1,8]{}, z\_[1,9]{}, z\_[1,10]{}\] = \[, -, , , , 0\] ,\
&& \[z\_[2,5]{}, z\_[2,6]{}, z\_[2,7]{}, z\_[2,8]{}, z\_[2,9]{}, z\_[2,10]{}\] = \[-, -, -, -, -, 0\] ,\
&& \[z\_[3,5]{}, z\_[3,6]{}, z\_[3,7]{}, z\_[3,8]{}, z\_[3,9]{}, z\_[3,10]{}\] = \[1 , , 3, , , \] . Here, note that the r.h.s. of equation (\[duff6\]) is expanded in terms of $F_i$ itself and the trivial conformal anomalies. This equation suggests that Duff’s scheme also works well in six dimensions.
Tree action
-----------
Let us first look for a conformally covariant sixth-order operator in six dimensions [@kmm]. It can be expanded in terms of the 21 independent operators, apart from the $\Box^3$ term, as &&\_6 = \^3 + v\_1 R\^\_\_ +v\_2 R \^2 + v\_3 (\^R\^)\_\_\_\
&& +v\_4 (\^R)\_ +v\_5 (\^\^R)\_\_ +v\_6 (R\^)\_\_\
&& +v\_7 (R) +v\_8 R\^\_[ ]{}R\^\_\_ +v\_9 R\_R\^\
&& +v\_[10]{} R\^R\^[ ]{}\_[ ]{}\_\_ +v\_[11]{} R\^R\^\_[ ]{} \_\_ +v\_[12]{} R\^R\_\
&& +v\_[13]{} RR\^\_\_ +v\_[14]{} R\^2 +v\_[15]{} (\^R)\_ +v\_[16]{} R\_(\^R\^)\_\
&& +v\_[17]{} R\^(\_R\_)\_ +v\_[18]{} R\_(\^R\^)\_ +v\_[19]{} R\_(\^R\^)\_\
&& +v\_[20]{} R\^(\_R)\_ +v\_[21]{} R(\^R)\_ , From the requirement that $\dl_{\phi}(\sq{g}\Delta_6 Y)= 0$ is satisfied locally for a scalar $Y$, the coefficients are determined as follows: && v\_1 = 4 , v\_2 =-1 , v\_3 = 4 , v\_4 = 0 , v\_5 = 0 , v\_6 = 4 ,\
&& v\_7 = - , v\_8 = \_1 , v\_9 = \_2 , v\_[10]{}= \_1 , v\_[11]{}= 6-\_1 ,\
&& v\_[12]{}= -1 +\_1 -\_2 , v\_[13]{}= -2 +\_1 , v\_[14]{}= -\_1 +\_2 ,\
&& v\_[15]{}= , v\_[16]{}= \_1 +4\_2 , v\_[17]{}= -\_1 , v\_[18]{}= 6 +\_1 ,\
&& v\_[19]{}= -2 -\_1 -2\_2 , v\_[20]{}= 1-\_1 , v\_[21]{}= - +\_1 +\_2 . In six dimensions, $\Delta_6$ is not unique, as the two constants $\zeta_1$ and $\zeta_2$ are not determined by the conformal property alone. The terms with these arbitrary constants are collected, using the Weyl tensor, in the forms $\zeta_1 \nb^{\mu}(C_{\mu\a\b\gm}C^{\nu\a\b\gm}\nb_{\nu})$ and $\zeta_2 \nb^{\lam}(C_{\a\b\gm\dl}C^{\a\b\gm\dl}\nb_{\lam})$, respectively [@kmm].
Next, we look for a combination of $G_6$ and $M_n$ that satisfies the following conformal property locally: \_ { ( G\_6 -\_[n=5]{}\^[10]{} w\_n M\_n ) } = 6 \_6 . \[cpog\] This equation determines the coefficients $w_n$ uniquely for each $\Delta_6$ with $\zeta_1$ and $\zeta_2$ as && w\_5 = 1+\_1 , w\_6 =11+\_1-3\_2 , w\_7 = -6 -\_1 ,\
&& w\_8 = 1+\_1 +3\_2 , w\_9 = -+\_1 +\_2 , w\_[10]{} = .
Using equation (\[cpog\]), the Wess-Zumino action defined by integrating the conformal anomalies within the interval $[0,\phi]$ are expressed in the form &&S(,)\
&&= d\^6 x \^\_[0]{} d { \^[3]{}\_[i=1]{} a\_i ( F\_i +\_[n=5]{}\^[10]{}z\_[i,n]{} M\_n ) + b G\_6 + \_[10]{}\^[5]{}c\_n M\_n }\
&&= d\^6 x { \^[3]{}\_[i=1]{}a\_i \_i + 3b \_6 +b (\_6 -\_[n=5]{}\^[10]{} w\_n \_n ) }\
&& + \^[10]{}\_[n=5]{} d\^6 x ( L\_n - \_n ) . Here, the $L_n$ are local functions given by integrating the $M_n$ as \_ d\^6 x L\_n = d\^6 x M\_n such that && L\_5 = K\_1 -K\_2 +K\_6 , L\_6 = K\_1 -K\_2 ,\
&& L\_7 = K\_1 -K\_2 +K\_3 -K\_5 -K\_6 ,\
&& L\_8 = K\_1 -K\_3 , L\_9 = -K\_1 , L\_[10]{} = K\_1 -K\_9 .
As discussed in the case of 4DQG, the integrability condition suggests that the sixth-order parts of the invariant action $I$ should be chosen such that the invariant $L_n$ terms cancel out in the sum ${\cal I}=S+I$. Hence, we obtain a 6DQG tree action analogous to that in 4DQG as &=& d\^6 x { - ( \_3 +\_1 \_1 +\_2 \_2 ) + \^[3]{}\_[i=1]{}a\_i \_i \
&& + 3b \_6 +b (\_6 -\_[n=5]{}\^[10]{} w\_n \_n )\
&& - \^[10]{}\_[n=5]{} ( \_[i=1]{}\^[3]{}a\_i z\_[i,n]{} +b w\_n +c\_n ) \_n } + I\_[LE]{}(X,g) . Here, we introduce the dimensionless coupling $t$, as in 4DQG. In six dimensions, two extra dimensionless constants $\a_1$ and $\a_2$, in addition to $\zeta_1$ and $\zeta_2$ in $\Delta_6$ and $w_n$, appear. These constants are not fixed by the arguments of the integrability. The constants $t$, $\a_1$ and $\a_2$ are renormalized, but $\zeta_1$ and $\zeta_2$ may not be.
Under a general coordinate transformation, this action changes according to = d\^6 x { -\^[3]{}\_[i=1]{} a\_i ( \_i +\_[n=5]{}\^[10]{}z\_[i,n]{} \_n ) - b \_6 - \_[10]{}\^[5]{}c\_n \_n } , where =-\_\^ . \[om6\]
Effective action
----------------
It is expected that the effective action of this model has the form = (X,,) + W\_[G\_6]{}() + \^[3]{}\_[i=1]{}W\_[F\_i]{}(,) + \^[10]{}\_[n=5]{} W\_[M\_n]{}() , where the tilde on ${\cal I}$ denotes the inclusion of finite corrections to the Wess-Zumino action described by $V_{NS}$ in the four-dimensional model. Here, $W_{G_6}$ is the generalization of the non-local and scale-independent Polyakov-Riegert action [@a; @deser]. We find its complete form as W\_[G\_6]{} () = d\^6 x { \_6 \_6 + \_[n=5]{}\^[10]{} w\_n \_n } , where \_6 = G\_6 -\_[n=5]{}\^[10]{} w\_n M\_n . This produces the type-A anomaly, W\_[G\_6]{}() = \_ W\_[G\_6]{} () = d\^6 x \_6 , where $\dl_{\om}\bg_{\mu\nu}=2\om\bg_{\mu\nu}$, with (\[om6\]). This equation is realized for arbitrary values of $\zeta_1$ and $\zeta_2$. These constants, as well as $e$, are determined according to matter content.
The action $W_{F_i}$, which produces the type-B anomaly in Duff’s scheme, is defined by W\_[F\_i]{}(,) =f\_i ( W\^\_[F\_i]{}(,) + \_[n=5]{}\^[10]{} d\^6 x \_n ) . The $\bL_n$ terms appear in Duff’s scheme. $W^{\pp}_{F_i}$ is a scale-dependent part defined through the equation \_ W\^\_[F\_i]{}(,) = d\^6 x \_i . It is known that the coefficients $e$ and $f_i$ are independent of the regularization scheme.
The remaining action, $W_{M_n}$, is a scheme-dependent part defined by W\_[M\_n]{}() = d\^6 x \_n . This action produces a trivial anomaly, $\bM_n$. As in 4DQG, it is unknown whether this action is really necessary or not. Since the vertices of the tree action at zeroth order in $t$ is conformally invariant, the coefficients $u_n$ will be at most order $t^2$.
The conditions for diffeomorphism invariance are now given by =f\_i , =e , c\_n=u\_n , where the tildes on $a_i$ and $b$ indicate the inclusions of corrections to the Wess-Zumino action. As in 4DQG, the scheme-dependent terms, $\bL_n$, cancel out, and the final expression takes the invariant and scheme-independent form = d\^6 x \_6 \_6 +\_[i=1]{}\^[3]{} f\_i W\^\_[F\_i]{}(g,)+ I\_[LE]{}(X,g) . The matter contributions to the coefficients $e$ and $f_i$ are computed in refs. [@bft; @g; @ii].
Conclusions and Discussion
==========================
In this paper we have discussed how the integrability conditions for conformal anomalies constrain the form of the effective action of even-dimensional QG. We showed that the effective action of 4DQG satisfying such integrability conditions has a manifestly diffeomorphism invariant and scheme-independent form. We then generalized the arguments to six dimensions and proposed a model for 6DQG. The expected scheme-independent form of the effective action was presented.
Now, the role of conformal anomalies in even-dimensional QG is naturally understood in terms of background-metric independence/diffeomorphism invariance. In $D=2n ~(\geq 4)$ dimensions, unlike the case for 2DQG, there is no critical matter content where the Wess-Zumino action vanishes. Thus, $2n$-dimensional QG is to be necessarily $2n$-th order, because of diffeomorphism invariance.
Background-metric independence does not require the vanishing of the usual beta functions in $D\geq 4$ dimensions, though it implies invariance under any conformal change of the background metric. This is due to the fact that there exist conformal anomalies, or the Wess-Zumino action, in even dimensions. We believe that conformal invariance in physics should be re-interpreted in terms of diffeomorphism invariance. In this case, the problem of dependence on the regularization scheme would disappear.
In odd dimensions, because there is no conformal anomaly, background-metric independence seems to require the theory to be finite. In three dimensions the Einstein-Hilbert$+$cosmological constant action is written in the Chern-Simons action, and its quantum theory is expected to be topological [@witten]. However, for $D \geq 5$, it is unknown whether odd-dimensional QG exists or not. Since in odd dimensions, we cannot introduce a dimensionless coupling constant, it seems necessary to make the theory super-renormalizable.
There is another approach to QG based on dynamical triangulation in two [@wei; @bk; @adj] and four dimensions [@migdal; @bbkptt; @efhty; @adj]. It is expected that our model is obtained in the continuum limit of such a simplicial QG. In this paper we do not discuss quantum corrections of the lower-derivative grvitational actions. The anomalous dimensions of the gravitational constant and the cosmological constant are needed to compare the two methods [@am; @h2]. A project involving detailed comparison in 4DQG between them has started [@efhty].
Finally, we comment on dimensional regularization. Because dimensional regularization violates conformal invariance in general, it is not a suitable regularization for a theory in which conformal invariance plays an important role. Nevertheless, dimensional regularization is still useful, because this violation is quite small and it is expected to give correct results for sufficiently higher order loops [@hath].
There is an assertion that, when using dimensional regularization, we can regularize QG defined by (\[z\]) in a manifestly diffeomorphism invariant way if we take great care concerning the conformal mode dependence [@kkn]. At present, the relation between this approach and ours is unknown. Detailed analyses of this relation are important to prove renormalizability to all orders.
The beautiful relations obtained among integrable curvature invariants in $D$ dimensions seem to suggest the validity of dimensional regularization. Our model, at least up to order $t_r^2$, gives correct results because of the finiteness of the self-energy diagrams of $\phi$, which implies that our model is rather insensitive to the conformal mode dependence. Whether or not the derived effective action at higher order is acceptable will be decided by the condition that it possesses a scheme-independent form and does not contain terms that violate diffeomorphism invariance, such as (\[wr2\]) and (\[wphi\]).
[**Acknowledgements**]{}
This work was supported in part by the Grant-in-Aid for Scientific Research from the Ministry of Education, Science and Culture of Japan.
[99]{} A. Polyakov, Phys. Lett. [**103B**]{} (1981) 207; Mod. Phys. Lett. [**A2**]{} (1987) 893. V. Knizhnik, A. Polyakov and A. Zamolodchikov, Mod. Phys. Lett. [**A3**]{} (1988) 819. J. Distler and H. Kawai, Nucl. Phys. [**B321**]{} (1989) 509;\
F. David, Mod. Phys. Lett. [**A3**]{} (1988) 1651. N. Seiberg, Prog. Theor. Phys. Suppl. [**102**]{} (1990) 319;\
J. Polchinski, Proc. of String 1990 (Texas A& M, march 1990); Nucl. Phys. [**B346**]{} (1990) 253;\
K. Hamada, Phys. Lett. [**B324**]{} (1994) 141; Nucl. Phys. [**B413**]{} (1994) 278. K. Hamada, Phys. Lett. [**B300**]{} (1993) 322;\
K. Hamada and A. Tsuchiya, Int. J. Mod. Phys. [**A8**]{} (1993) 4897. H. Kawai and M. Ninomiya, Nucl. Phys. [**B336**]{} (1990) 115;\
H. Kawai, Y. Kitazawa and M. Ninomiya, Nucl. Phys. [**B393**]{} (1993) 280; Nucl. Phys. [**B404**]{} (1993) 684;\
T. Aida, Y. Kitazawa, H. Kawai and M. Ninomiya, Nucl. Phys. [**B427**]{} (1994) 158;\
T. Aida, Y. Kitazawa, J. Nishimura and A. Tsuchiya, Nucl. Phys. [**B444**]{} (1995) 353;\
H. Kawai, Y. Kitazawa and M. Ninomiya, Nucl. Phys. [**B467**]{} (1996) 313;\
T. Aida and Y. Kitazawa, Nucl. Phys. [**B491**]{} (1997) 427. R. Riegert, Phys. Lett. [**134B**]{} (1984) 56. E. Fradkin and A. Tseytlin, Phys. Lett. [**134B**]{} (1984) 187. E. Tomboulis, Nucl. Phys. [**329**]{} (1990) 410. I. Antoniadis and E. Mottola, Phys. Rev. [**D45**]{} (1992) 2013;\
I. Antoniadis, P. Mazur and E. Mottola, Phys. Lett. [**B323**]{} (1994) 284, [**B394**]{} (1997) 49. I. Antoniadis, P. Mazur and E. Mottola, Nucl. Phys. [**B388**]{} (1992) 627. S. Odintsov, Z. Phys. [**C54**]{} (1992) 531;\
I. Antoniadis and S. Odintsov, Phys. Lett. [**B343**]{} (1995) 76. K. Hamada and F. Sugino, Nucl. Phys. [**B553**]{} (1999) 283. K. Hamada, Prog. Theor. Phys. [**103**]{} (2000) 1237 hep-th/9912098; hep-th/0005063. D. Karakhanian, R. Manvelian and R. Mkrtchian, Mod. Phys. Lett. [**A11**]{} (1996) 409 hep-th/9411068;\
T. Arakelian, D. Karakhanian, R. Manvelian and R. Mkrtchian, Phys. Lett. [**B353**]{} (1995) 52. S. Odintsov and A. Romeo, Mod. Phys. Lett. [**A9**]{} (1994) 3373 hep-th/9410191. D. Capper and M. Duff, Nuovo Cimento [**23A**]{} (1974) 173. S. Deser, M. Duff and C. Isham, Nucl. Phys. [**B111**]{} (1976) 45. M. Duff, Nucl. Phys. [**B125**]{} (1977) 334;\
[*Twenty years of the Weyl anomaly*]{}, Class. Quant. Grav. [**11**]{} (1994) 1387 hep-th/9308075 and references therein. K. Fujikawa, Phys. Rev. [**D23**]{} (1981) 2262; Phys. Rev. Lett. [**44**]{} (1980) 1733. L. Bonora, P. Cotta-Ramusino and C. Reina, Phys. Lett. [**B126**]{} (1983) 305. L. Bonora, P. Pasti and M. Bregola, Class. Quant. Grav. [**3**]{} (1986) 635. S. Deser and A. Schwimmer, Phys. Lett. [**B309**]{} (1993) 279 hep-th/9302047. D. Anselmi, Nucl. Phys. [**B567**]{} (2000) 331 hep-th/9905005;\
Phys. Lett. [**B476**]{} (2000) hep-th/9908014. S. Deser, Phys. Lett. [**B479**]{} (2000) 315 hep-th/9911129. F. Bastianelli, S. Frolov and A. Tseytlin, JHEP [**0002**]{} (2000) 013 hep-th/0001041. F. Bastianelli, G. Cuoghi and L. Nocetti, [*Consistency conditions and trace anomalies in six dimensions*]{} hep-th/0007222. P. Gilkey, J. Diff. Geom. [**10**]{} (1975) 601;\
D. Toms, Phys. Rev. [**D26**]{} (1982) 2713;\
T. Parker and S. Rosenberg, J. Diff. Geom. [**25**]{} (1987) 199. S. Ichinose and N. Ikeda, J. Math. Phys. [**40**]{} (1999) 2259 hep-th/9810256;\
F. Bastianelli and O. Corradini, hep-th/0010118. J. Wess and B. Zumino, Phys. Lett. [**37B**]{} (1971) 95. M. Kato and K. Ogawa, Nucl. Phys. [**B212**]{} (1983) 443;\
K. Itoh, Nucl. Phys. [**B342**]{} (1990) 449. E. Tomboulis, Phys. Lett. [**70B**]{} (1977) 361; Phys. Lett. [**97B**]{} (1980) 77;\
B. Hasslacher and E. Mottola, Phys. Lett. [**99B**]{} (1981) 221. E. Fradkin and A. Tseytlin, Nucl. Phys. [**B201**]{} (1982) 469; Phys. Lett. [**104B**]{} (1981) 377. A. Bartoli, J. Julve and E. Sánchez, Class. Quant. Grav. [**16**]{} (1999) 2283. M. Kaku, Nucl. Phys. [**B203**]{} (1982) 285;\
D. Boulware, G. Horowitz and A. Strominger, Phys. Rev. Lett. [**50**]{} (1983) 1726;\
E. Tomboulis, Phys. Rev. Lett. [**52**]{} (1984) 1173. T. Kugo and S. Uehara, Nucl. Phys. [**B197**]{} (1982) 378. E. Witten, Nucl. Phys. [**B311**]{} (1988/89) 46. D. Weingarten, Phys. Lett. [**B90**]{} (1980) 285, Nucl. Phys. [**B210**]{} \[FS6\] (1982) 229;\
V. Kazakov, Phys. Lett. [**150B**]{} (1985) 282;\
F. David, Nucl. Phys. [**B257**]{} (1985) 45;\
J. Ambjørn, B. Durhuus and J. Fröhlich, Nucl. Phys. [**B257**]{} (1985) 433;\
D. Boulatov, V. Kazakov, I. Kostov and A. Migdal, Nucl. Phys. [**B275**]{} (1986) 641. E. Brezin and V. Kazakov, Phys. Lett. [**B236**]{} (1990) 144;\
M. Douglas and S. Shenker, Nucl. Phys. [**B335**]{} (1990) 635;\
D. Gross and A. Migdal, Phys. Rev. Lett. [**64**]{} (1990) 127. J. Ambjørn, B. Durhuus and T. Jonsson, [*Quantum Geometry*]{}, (Cambridge Univ. Press, Cambridge, 1997). M. Agishtein and A. Migdal, Nucl. Phys. [**B385**]{} (1992) 395;\
J. Ambjørn and J. Jurkiewicz, Phys. Lett. [**B278**]{} (1992) 42;\
S. Catterall, J. Kogut and R. Renken, Phys. Lett. [**B328**]{} (1994) 277. S. Bilke, Z. Burda, A. Krzywicki, B. Petersson, J. Tabaczek and G. Thorleifsson, Phys. Lett. [**B418**]{} (1998) 266; Phys. Lett. [**B432**]{} (1998) 279. H. Egawa, A. Fujitu, S. Horata, N. Tsuda and T. Yukawa, Nucl. Phys. [**B (Proc. Suppl.) 73**]{} (1999) 795; [*Phase Transition of 4D Simplicial Quantum Gravity with $U(1)$ Gauge Field*]{}, Proc. Lattice 99, Pisa (June-July 1999)/hep-lat/9908048;\
S. Horata, H. Egawa, N. Tsuda and T. Yukawa, [*Phase Structure of Four-dimensional Simplicial Quantum Gravity with a $U(1)$ Gauge Field*]{}, hep-lat/0004021;\
S. Horata, H. Egawa and T. Yukawa, [*Geometry of 4d Simplicial Quantum Gravity with a $U(1)$ Gauge Field*]{}, hep-lat/0010050. I. Drummond and S. Hathrell, Phys. Rev. [**D21**]{} (1980) 958;\
G. Shore, Phys. Rev. [**D21**]{} (1980) 2226; Ann. Phys. [**128**]{} (1980) 376;\
S. Hathrell, Ann. Phys. [**139**]{} (1982) 136; Ann. Phys. [**142**]{} (1982) 34.
[^1]: In contrast to 2DQG, in which the classical limit is given by $N \arr -\infty$, the limit of positive large $N$ gives the correct classical limit in 4DQG [@am].
[^2]: Even in 2DQG, although we can set ${\bf \dl_B} {\cal I}=0$ if we use the flat background-metric and integrate out the $B_{\mu\nu}$ field, the nilpotency of the BRST charge at the quantum level, after all, requires condition (\[b\]). Thus, the BRST invariance in even-dimensional QG is realized dynamically.
[^3]: Some errors in the form of the effective action in section 3.3 of ref. [@h2] are corrected in this section.
[^4]: The possibility that an invariant $R^2$ term appears in the effective action is not excluded. There is a possibility that such a term appears in $V_{NS}$ at order $t^4_r$.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'Real-world clinical time series data sets exhibit a high prevalence of missing values. Hence, there is an increasing interest in missing data imputation. Traditional statistical approaches impose constraints on the data-generating process and decouple imputation from prediction. Recent works propose recurrent neural network based approaches for missing data imputation and prediction with time series data. However, they generate deterministic outputs and neglect the inherent uncertainty. In this work, we introduce a unified Bayesian recurrent framework for simultaneous imputation and prediction on time series data sets. We evaluate our approach on two real-world mortality prediction tasks using the MIMIC-III and PhysioNet benchmark datasets. We demonstrate significant performance gains over state-of-the-art methods, and provide strategies to use the resulting probability distributions to better assess reliability of the imputations and predictions.'
author:
- |
Yang Guo, Zhengyuan Liu, Pavitra Krishnswamy\*, Savitha Ramasamy\*[^1]\
Institute for Infocomm Research\
Agency for Science, Technology and Research (A\*STAR)\
Singapore\
`{pavitrak,ramasamysa}@i2r.a-star.edu.sg`
bibliography:
- 'ref.bib'
title: Bayesian Recurrent Framework for Missing Data Imputation and Prediction with Clinical Time Series
---
Introduction
============
Real-world healthcare datasets across acute and community care settings exhibit a high prevalence of missing values. Due to this challenge, predictive analysis with these data is often challenging, and insights mined from these data may not be most reliable [@liptonmodeling][@5][@1][@3][@10]. In particular, this problem is especially cumbersome for clinical time series data comprising longitudinal records of patient state. This has spurred a longstanding interest in missing data imputation methods [@5; @BayesianMissing_2003; @review_2010].
Traditional approaches for such problems have relied on statistical models and associated Bayesian inference paradigms [@multiple_2011][@9][@hastie2005elements], but these require strong constraints on the data-generating process, and treat the imputation and prediction as independent tasks [@3; @multiple_2011; @EHRD_2013]. To overcome these limitations, recent works have proposed deep learning approaches using recurrent neural networks [@5; @ML4H; @Lipton2016; @9; @1]. While these methods learn directly from data without imposing specific assumptions on the underlying processes, and show promise for accurate imputation and prediction on clinical time series, they provide deterministic outputs and neglect the uncertainty inherent to the task.
In this work, we present a unified Bayesian framework for imputation and prediction with multivariate clinical time series. We embed a Bayesian Recurrent Neural Network and a Bayesian Neural Network within a recurrent dynamical system for integrative missing value imputation and prediction. We characterize performance on mortality prediction tasks with the publicly available MIMIC-III [@6] and PhysioNet [@7] benchmark data sets, and demonstrate significant performance improvements. We further show strong correlations between the variability and accuracy of the imputations. These results suggest that our approach adapts the imputation model to the uncertainty inherent in the modelling task, and offers a principled way to assess reliability of imputations and predictions.
Methods
=======
**Problem Formulation:** We consider dataset $D$ of $N$ samples. Each sample is denoted as an input-output pair $(\mathbf{X}, Y)$, where $\mathbf{X} = \left({\mathbf{x}_{1}, \mathbf{x}_{2}, \ldots, \mathbf{x}_{M}}\right)$ denotes a multivariate time series input with $M$ features and $Y$ denotes its output. Each feature is a sequence of observations over $T$ time steps $\mathbf{x}_{i} = \left[{x^1_{i}, x^2_{i}, \ldots, x^T_{i}}\right]$. In practice, $\mathbf{X}$ may have missing values. A masking matrix $\mathbf{V}_{T \times M}$ represents the presence of missing values: $V_{i,j} = 0$ when $X_{i,j}$ is missing, otherwise $V_{i,j} = 1$. The objectives are: (a) impute the missing values in $X$ and (b) predict $Y$ given $\mathbf{X}$.
**Bayesian Recurrent Framework:** To address the above objectives, prior works leverage recurrent dynamical systems for deterministic models [@5; @Lipton2016]. Here, we augment such recurrent dynamical systems with Bayesian approaches to model the uncertainty in the imputation and prediction tasks. Fig. \[structure\] illustrates our proposed Bayesian recurrent framework for imputation and prediction.
At each time step, the input is fed through the masking layer to a Bayesian recurrent neural network. The RNN hidden state dynamics are specified as: $$\mathbf{h}^{t+1} = f(\mathbf{W}_{r}, \mathbf{h}^{t}, \mathbf{x}^t),$$ where $\mathbf{h}^{t}$ is the RNN hidden state at time $t$ and $\mathbf{W}_{r}$ denotes the weights for the recurrent layer. As the inputs may not always be regularly sampled over time, we incorporate a temporal decay factor in the hidden state dynamics [@5]. To model the uncertainty in $f$, the Bayesian RNN considers a probabilistic distribution $p(\mathbf{W}_{r}|D)$ for the weights, instead of fixed values $\mathbf{W}_{r}$ [@8].
If $\mathbf{x}^t$ has missing values for any feature, we impute the missing values and replace $\mathbf{x}^t$ with: $$\mathbf{\hat{x}}^{t} = g(\mathbf{U}_{x}, \mathbf{h}^{t-1}, \mathbf{x}^t),$$ The imputation function $g$ implicitly models the correlation amongst features [@1]. To model the uncertainty in $g$, we use a Bayesian multilayer perceptron that considers a probabilistic distribution $p(\mathbf{U}_{x}|D)$ for the weights $\mathbf{U}_{x}$ [@8].
We apply the imputations $\mathbf{\hat{X}}$ when the masking matrix $\mathbf{V}$ indicates missing values. Thus, across time steps, the updated input $\mathbf{\tilde{X}}$ is: $$\mathbf{\tilde{X}} = (\mathbf{I}-\mathbf{V})\odot\mathbf{\hat{X}} + \mathbf{V}\odot\mathbf{X}$$
With $\mathbf{\tilde{X}}$ and the RNN hidden states $\mathbf{H} = [\mathbf{h}^1, \mathbf{h}^2, \ldots, \mathbf{h}^T]$, the predicted output $\hat{Y}$ is: $$\hat{Y} = f_{\rm{out}}(\mathbf{\tilde{X}}, \mathbf{H})$$ We use a linear form for $f_{out}$. The above equations specify the overall recurrent dynamical system. Then, to obtain the imputations and predictions given input $\mathbf{X}$, we need to compute the probabilistic distribution of the overall weights $W = \left[{W_{r}, U_{x}}\right]$, given data $D$. As the true posterior distribution $p(W|D)$ is intractable in general, we have to approximate it using Bayes by Backprop [@8]. Conceptually, Bayes by Backprop minimizes the Kullback-Leibler divergence between the approximate distribution $q(W|\Theta)$ and the true posterior $p(W|D)$. As such, the loss function for our estimation not only comprises the imputation and prediction errors, but also the KL-divergence loss, as below: $$\begin{aligned}
L_{\rm{total}} &=& L_{\rm{KL}}(Y,\mathbf{\tilde{X}}, p(\mathbf{W})) + L_{\rm{imput}}(\mathbf{X}, \mathbf{\hat{X}}) + L_{\rm{pred}}(Y, \hat{Y}),\\
L_{\rm{KL}} &=& -\mathbb{E}_{q(\mathbf{W})}[\log p(Y|\mathbf{W}, \mathbf{\tilde{X}})] + KL[q(\mathbf{W}|\Theta) || p(\mathbf{W})],\\
L_{\rm{imput}} &=& (\mathbf{X} - \mathbf{\hat{X}})\odot\mathbf{V}, \\
L_{\rm{pred}} &=& L_{\rm{cross-entropy}}(\hat{Y},Y),\end{aligned}$$ where $p(\mathbf{W})$ is the prior distribution for the weights, set as a mixed Gaussian. We highlight that the imputation loss only considers performance for sample values that are *not* missing. The loss function controls the imputation, prediction and posterior distribution of the weights simultaneously. Minimizing the loss function, we obtain the posterior of the weights as well as the imputed values and output predictions. Our proposed framework adapts jointly to the uncertainty in the imputation and prediction process, functions as a regularizer to improve robustness, and also provides distributions of the resulting estimates for further study.
Experiments and Results
=======================
**Data:** To evaluate our approach, we perform experiments on mortality prediction tasks using benchmark data sets from the PhysioNet 2012 Challenge [@7] and the MIMIC-III collection [@6; @3]. These data sets comprise of multivariate time series clinical features recorded from patients in the intensive care unit (ICU). For PhysioNet, the input comprises 35 numerical features in time series samples from 4000 admissions. For MIMIC-III, the input comprises 12 numerical features in time series samples from 14,681 admissions. In both cases, each input sample comprises 48 hourly time steps and the output is in-hospital mortality. We note that both data sets are sparse, with 78% and 48% of the values missing for PhysioNet and MIMIC-III, respectively.
**Performance Metrics**: To enable evaluation of imputation performance, we simulate missingness at random (MAR) by eliminating 10% of the known input values [@1]. For this subset of data with simulated MAR, we evaluate imputation performance by computing the mean absolute error (MAE) and mean relative error (MRE). Further, we pick a test set with random subset of 20% from all data samples to evaluate the prediction performance by computing the areas under the receiver operating characteristics and the precision recall curves (AUROC, AUPRC).
**Baselines:** We compare our method against three state-of-the-art RNN based imputation methods: (a) Gated Recurrent Units (GRU-D) [@5]; (b) Recurrent Imputation in Time Series (RITS) [@1]; (c) Bayesian Recurrent Neural Networks (BRNN) that refers to a Bayesian RNN [@2] with all missing values imputed with zero. Unlike our method wherein the temporal decay factor only affects hidden states, the GRU-D baseline considers the decay factors both for input and hidden state dynamics. The RITS baseline considers a vanilla RNN without a Bayesian framework. Finally, the BRNN baseline helps evaluate the impact of a Bayesian approach for prediction, independently of data-driven imputation.
**Performance Results:** Table \[Imputation\_Performance\] provides the performance results. Where previously reported results exist, we include them with citations. We also repeat the experiments for fair comparison, as required. We observe that our method outperforms the state-of-the-art techniques, beating the closest SOTA method by upto 2% in imputation MRE and upto 3.3% in prediction AUPRC. Our imputation performance improvement is less prominent in the MIMIC-III dataset, possibly as this dataset is less sparse than PhysioNet. We also characterized the imputation performance with increasing prevalence of simulated MAR for PhysioNet. With increasing MAR rates, our method offered increasing performance improvement over the closest SOTA method (RITS). Even with a 15% increase in simulated MAR, our approach outperformed RITS by 3.5% MRE.
**Impact of Distributions:** One advantage of our Bayesian framework is that it provides the ability to get a distribution of imputations and predictions. First, for each sample with simulated MAR, we used Monte Carlo iterations to obtain a distribution of the imputed values (Fig. \[figuncertainty\]A). By visualizing the ground truth value atop the distribution (blue point), we can assess how far removed the distribution is from the ground truth. Second, we studied the variability implicit in the imputation process. For each of the simulated MAR, we obtain the variance of the distribution of imputed values $\sigma^2$ (e.g., red line). We then sort the variances of all imputations in ascending order (Fig. \[figuncertainty\]B), eliminate missing values with the highest variances, and assess the impact on imputation MAE. We note, for example, that removing missing values with $\sigma^2$ values in the top 40 percentile of variances leads to lower MAE than when all missing values are considered. This suggests that there is a monotonic relation between the accuracy and variability of imputation. Third, we breakdown this accuracy vs variability relation to the individual feature level (Fig. \[figuncertainty\]C). The same trend carries through, wherein features with higher variability in imputed values (e.g., weight, height and inspired oxygen) tend to have higher MAE. These results suggest that, in real-world scenarios, when there is no ground truth, the variance $\sigma^2$ can serve as a means to assess reliability of the imputed values.
**Dataset** **Methods** **MAE** **MRE** **AUROC** **AUPRC**
------------- ------------- ------------------ ------------------ ------------------ ------------------
PhysioNet GRU-D [@1] $0.559$ $0.776$ $0.828$ $-$
RITS [@1] $0.300$ $0.419$ $0.840$ $-$
BRNN $0.708$ $1.000$ $0.811$ $0.515$
RITS $0.296$ $0.418$ $0.834$ $0.532$
Ours $\textbf{0.282}$ $\textbf{0.398}$ $\textbf{0.866}$ $\textbf{0.553}$
MIMIC-III GRU-D $0.390$ $0.779$ $0.790$ $0.421$
BRNN $0.503$ $1.000$ $0.775$ $0.422$
RITS $0.151$ $0.300$ $0.805$ $0.432$
Ours $\textbf{0.148}$ $\textbf{0.294}$ $\textbf{0.815}$ $\textbf{0.465}$
: \[Imputation\_Performance\] Performance Comparison for Imputations and Predictions
Discussion and Future Work
==========================
We have developed a Bayesian recurrent framework to enable missing data imputation and prediction on clinical time series data sets. Our approach improves both imputation and prediction performance and is robust to increasing MAR. Further, by providing explicit probability distributions of imputed values and output predictions, it enables assessment of variability and reliability of the imputed values. This has important implications in real-world scenarios, where ground truth is lacking. Future work will consider expansions to include categorical features, varying length time series and different missingness patterns such as NMAR (Missing Not At Random) and MCAR (Missing Completely At Random) and develop more rigorous theoretical grounding.
Acknowledgement {#acknowledgement .unnumbered}
===============
The authors would like to acknowledge grant funding for Digital Health from the Science and Engineering Research Council, A\*STAR, Singapore (Project No. A1818g0044).
[^1]: denotes equal contribution.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: |
Using data recorded by the CLEO-II detector at CESR, we report new measurements of the branching fractions for the decays of the charmed baryon $\Lambda_c^+$ into $pK^-\pi^+\pi^0$, $p\overline{K}^0$, $p\overline{K}^0\pi^+\pi^-$, and $p\overline{K}^0\pi^0$, all measured relative to $pK^-\pi^+$. The relative branching fractions are $0.67\pm0.04\pm0.11,
0.46\pm0.02\pm0.04,0.52\pm0.04\pm0.05$, and $0.66\pm0.05\pm0.07$ respectively.
author:
- CLEO Collaboration
title: 'Measurement of the Branching Fractions of $\Lambda_c^+\to p\overline{K}n(\pi)$'
---
6.5 in 9.0 in -0.50in 0.00in 0.00in
=0.1cm
Since the first observation of the lowest lying charmed baryon, the $\Lambda_c^+$, there have been many measurements made of its exclusive decay channels. As it is difficult to measure the production cross-section of the $\Lambda_c^+$ baryons, decay rates are typically presented as branching ratios relative to $\Lambda_c^+
\to pK^-\pi^+$, the most easily observed decay channel. However, fewer than half of the $\Lambda_c^+$ hadronic decays are presently accounted for. Measurement of these modes is of practical as well as theoretical interest. Here, we present measurements of the branching fractions of $\Lambda_c^+$ into $pK^-\pi^+\pi^0$, $p\overline{K}^0$, $p\overline{K}^0\pi^+\pi^-$, and $p\overline{K}^0\pi^0$, all relative to $pK^-\pi^+$. The last of these is the first measurement of this mode. The other modes have been previously measured but with considerably larger uncertainties than in the present study.
The data presented here were taken by the CLEO II detector[@KUB] operating at the Cornell Electron Storage Ring. The sample used in this analysis corresponds to an integrated luminosity of 4.8 $fb^{-1}$ from data taken on the $\Upsilon(4S)$ resonance and in the continuum at energies just above and below the $\Upsilon(4S)$. We detected charged tracks with a cylindrical drift chamber system inside a solenoidal magnet. Photons were detected using an electromagnetic calorimeter consisting of 7800 cesium iodide crystals.
Particle identification of $p,K^-$, and $\pi^+$ candidates was performed using specific ionization measurements in the drift chamber, and when present, time-of-flight measurements. For each mass hypothesis, a combined $\chi^2$ probability $P_i$ was formed ($i=\pi,K,p$). Using these probablilities, a normalized probability ratio $L_i$ was evaluated, where $L_i=P_i/(P_{\pi}+P_K+P_p)$. Well identified protons peaked near $P_p=1.0$ while tracks that were identified to not be protons peak near $P_p=0.0$. For a track to be used as a proton in this study, we required it to have $L_p>0.8$, which eliminated much of the background, though with considerable diminution of efficiency. For kaons we applied a looser and more efficient cut of $L_K>0.1$. We have chosen these cuts using a Monte Carlo simulation program to maximize the significance of the signals. The proton identification requirement resulted in an efficiency that is strongly momentum dependent, whereas the kaon identification is rather efficient at all momenta. In order to reduce the large combinatoric background, we required $x_p>0.5$, where $x_p=P_{\Lambda_c}/\sqrt{E^2_{beam} - m^2_{\Lambda_c}}$ is the scaled momentum of the $\Lambda_c^+$ candidate. Approximately 60% of $\Lambda_c^+$ baryons from $c\overline{c}$ continuum events passed this requirement.
The $\overline{K}^0$ candidates were identified in their decay $K^0_s \to \pi^+\pi^-$, by reconstructing a secondary vertex from the intersection of two oppositely charged tracks in the $r-\phi$ plane. The invariant mass of the $\overline{K^0}$ candidate must lie within 9 ${\rm MeV/c^2}$ (around 3 standard deviations ($ \sigma$) of its nominal value.
The $\pi^0$ candidates were selected through their decay $\pi^0 \to \gamma\gamma$ from pairs of well-defined showers in the CsI calorimeter with a reconstructed invariant mass within $3\sigma$ of the $\pi^0$ mass. In order to reduce the combinatorial background, each $\gamma$ was required to have an energy of at least 50 MeV, and the $\pi^0$ was required to have a momentum of at least $\rm 300\ MeV/c$.
The resulting mass distributions for the 5 modes are shown in Figure 1. Each peak was fit to the sum of a Gaussian signal distribution with width fixed to that obtained from CLEO’s GEANT based Monte Carlo simulation program, and a second order polynomial background distribution. The signal widths used and the resulting signal yields are tabulated in Table 1.
----------------------------- ---------------- ---------------
Mode MC Width (MeV) Signal
$pK^-\pi^+$ 16 $10109\pm191$
$pK^-\pi^+\pi^0$ 22 $2606\pm165$
$p\overline{K}^0$ 19 $1025\pm40$
$p\overline{K}^0\pi^+\pi^-$ 15 $985\pm65$
$p\overline{K}^0\pi^0$ 27 $774\pm52$
----------------------------- ---------------- ---------------
: The number of $\Lambda_c^+$’s found with $x_p(\Lambda_c)>0.5$
The efficiency for each $\Lambda_c^+$ mode was calculated using the Monte Carlo simulation program[@GEANT]. The particle identification efficiency was checked using real data from $\Lambda \to p\pi^-$ and $D^{*+} \to K^-\pi^+\pi^+$ decays that were identified topologically. The reconstruction efficiency of the $\Lambda_c^+$ decays has some dependence on the resonant substructure of these states. In the case of the $pK^-\pi^+$ mode, the Monte-Carlo generator produced a mixture of non-resonant three-body decay together with $\Delta^{++}K^-$ and $p\overline{K}^{*0}$ decays, according to their measured branching fractions[@PDB]. These three types of decays had slightly different reconstruction efficiencies, so that including the substructure changes the efficiency by $\Delta\epsilon / \epsilon=0.02$ relative to 3-body phase space. We have also investigated the dependence of the reconstruction efficiency of the other modes on possible resonant substructure. The poor signal to background ratio did not allow a detailed measurement of the substructure of these modes. The efficiency calculation took into account the $\overline{K^0}\to K^0_s$ and $K_s^0 \to \pi^=\pi^-$ branching fractions.
We have considered many possible sources of systematic error in the measurement. The main contributors to the systematic uncertainty came from the following sources: 1) Uncertainties in the fitting procedures, which were estimated by looking at the changes in the yields using different orders of polynomial background and different signal widths (15% in the case of $pK^-\pi^+\pi^0$, but much smaller for the other modes), 2) uncertainties due to the unknown mix of resonant substructure in the multi-body decays (up to 3% depending on the mode), 3) uncertainties due to $\pi^0$ finding (5%), $K^0_s$ finding (5%) and track finding (1%), and 4) uncertainties in the reconstruction efficiency due to the particle identification criteria for protons and kaons (4%). These uncertainties have been added in quadrature to obtain the total systematic uncertainty for each mode, taking into account the fact that many of these tend to cancel in a measurement of ratios of branching fractions.
There are three main types of quark decay diagrams that contribute to $\Lambda_c^+$ decays. The simplest method is the simple spectator diagram in which the virtual $W^+$ fragments independently of the spectator quark. The second method involves the quark daughters of the $W^+$ combining with the remaining quarks. The third method, W-exchange, involves the $W^+$ combining with the initial $d$ quark. Unfortunately all the decay modes under investigation here can proceed by more than one of these decay diagrams, and their decay rates are not amenable to calculation.
In conclusion, we have measured new branching fractions of the $\Lambda_c^+$ into 4 decay modes, measured relative to the normalizing mode $\Lambda_c^+\to pK^-\pi^+$. The results for three of these modes are in agreement with, and more accurate than, previous measurements. We have made the first measurement of the decay rate of $\Lambda_c^+\to p\overline{K}^0\pi^0$. These measurements help account for the total width of the $\Lambda_c^+$ and increase the understanding of charmed baryon decays.
----------------------------- --------------------- ------------------------ -------------------------------
Mode Relative Efficiency $B/B(pK^-\pi^+)$ Previous Measurements
$pK^-\pi^+$ 1.0 1.0
$pK^-\pi^+\pi^0$ 0.383 $0.67\pm 0.04\pm 0.11$ $0.72^{+0.32}_{-0.22}$[@BARL]
$p\overline{K}^0$ 0.218 $0.46\pm 0.02\pm 0.04$ $0.44\pm0.07\pm0.05$[@AVE]
$0.55\pm0.17\pm0.14$[@ANJ]
$0.62\pm0.15\pm0.03$[@ALB]
$p\overline{K}^0\pi^+\pi^-$ 0.187 $0.52\pm 0.04\pm 0.05$ $0.43\pm0.12\pm0.04$[@AVE]
$0.98\pm0.36\pm0.08$[@BARL]
$p\overline{K}^0\pi^0$ 0.115 $0.66\pm 0.05\pm 0.07$
----------------------------- --------------------- ------------------------ -------------------------------
: The measured relative branching fractions
Y. Kubota et al., Nucl. Inst. and Methods A320, 66 (1992).
R. Brun et al., CERN/DD/EE/84-11.
Review of Particle Properties, R. Barnett et al., Phys. Rev. D54 1 (1996).
P. Avery et al., Phys. Rev. D43, 3499 (1991)
S. Barlag et al., Z. Phys. C48, 29 (1990).
J. Anjos et al., Phys. Rev. D41, 801 (1990).
H. Albrecht et al., Phys. Lett. B207, 109 (1988).
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'This article is concerned with the time evolution of the oblique laminar-turbulent bands of transitional plane Couette flow under the influence of turbulent noise. Our study is focused on the amplitude of modulation of turbulence (the bands). In order to guide the numerical study of the flow, we first perform an analytical and numerical analysis of a Stochastic Ginzburg–Landau (GL) equation for a complex order parameter. The modulus of this order parameter models the amplitude of modulation of turbulence. Firstly, we compute the autocorrelation function of said modulus once the band is established. Secondly, we perform a calculation of average and fluctuations around the exponential growth of the order parameter. This type of analysis is similar to the Stochastic Structural Stability Theory (S3T). We then perform numerical simulations of the Navier–Stokes equations in order to confront these predictions with the actual behaviour of the bands. Computation of the autocorrelation function of the modulation of turbulence shows quantitative agreement with the model: in the established band regime, the amplitude of modulation follows an Ornstein–Uhlenbeck process. In order to test the S3T predictions, we perform quench experiments, sudden decreases of the Reynolds number from uniform turbulence, in which modulation appears. We compute the average evolution of the amplitude of modulation and the fluctuations around it. We find good agreement between numerics and modeling. The average trajectory grows exponentially, at a rate clearly smaller than that of the formation of laminar holes. Meanwhile, the actual time evolution remains in a flaring envelope, centered on the average, and expanding at the same rate. These results provide further validation of the stochastic modeling for the time evolution of the bands for further studies. Besides, they stress on the difference between the oblique band formation and the formation of laminar holes.'
author:
- 'Joran Rolland[^1]'
title: 'Stochastic analysis of the time evolution of Laminar-Turbulent bands of plane Couette flow'
---
Shear turbulence, 47.27.nb – Transition to turbulence, 47.27.Cn – Stochastic analysis methods, 05.10.Gg
Introduction
============
This article studies the random time evolution of the modulation of turbulence in transitional plane Couette flow (Fig. \[fig1\] (a)). Plane Couette flow is the flow between two parallel plates moving in opposite directions and separated by a constant gap (Fig. \[fig1\] (b)). This flow is linearly stable for all Reynolds numbers $R$, the control parameter. As a consequence, the transition to turbulence is discontinuous : turbulence requires finite amplitude perturbations to be triggered. Besides, turbulent flow can coexist in space and time with laminar flow, provided that $R>R_{\rm g}$, the global threshold of transition. There are several manner to define this threshold: $R_{\rm g}$ can be the Reynolds number under which turbulence is not sustained permanently [@prigent02; @PM], or more precisely, the Reynolds number under which the mean lifetime of a turbulent germ equals the mean time elapsed before that germ splits [@shi]. The most intriguing property of the flow is that up to $R_{\rm t}>R_{\rm g}$, turbulence is sustained, but does not invade the whole domain. Instead, it takes the form of regular apparently quasi steady oblique laminar-turbulent bands [@prigent02; @PM; @dsc10]. Above $R_{\rm t}$, turbulence occupies the whole domain: again, this Reynolds number has several definitions, all related to the disappearance of the laminar troughs [@prigent02; @PM; @BT11; @RM10_1]. Note that the bands can have the two orientations (termed $+$ and $-$) with equal probability. These bands correspond to a sinusoidal modulation of the amplitude of turbulence [@BT11; @RM10_1]. Said modulation disappears in a very intermittent manner near $R_{\rm t}$ [@RM10_1; @BT05], which may in fact be very similar to a critical phenomenon [@RM10_1; @RNL]. Understanding the type of noise felt by the band is fundamental in explaining to what extent this similarity goes.
**(a)**![(a) : Colour levels of the norm of the velocity field in the band regime of plane Couette flow (numerical result), in a horizontal plane. (b) : Sketch of the plane Couette flow configuration. The wall normal direction $\vec e_y$ sets the vertical.[]{data-label="fig1"}](normeym062_.eps "fig:"){width="4cm"}**(b)**![(a) : Colour levels of the norm of the velocity field in the band regime of plane Couette flow (numerical result), in a horizontal plane. (b) : Sketch of the plane Couette flow configuration. The wall normal direction $\vec e_y$ sets the vertical.[]{data-label="fig1"}](couette_geometry.eps "fig:"){width="6cm"}
There have been many attempts at understanding the hydrodynamic mechanisms sustaining the bands. Advection of small scale structures (size of the gap) by a large scale flow (size of the bands) is central in each of these discussions [@PRL; @ispspot; @BT07]. The topic is still open, since the effect of these small scale structures on the large scale flow has not been precisely explained. Comparatively, the modeling of the laminar-turbulent coexistence is more advanced. Model equations have been proposed that describe the amplitude and localisation of turbulence [@bp] or the amplitude of modulation of turbulence $m$ (the bands) [@prigent02; @RM10_1]. Said amplitude of modulation is usually computed from a Fourier transform of the velocity field [@RM10_1; @BT11]. In both cases, it is very clear that fluctuating quantities with well defined PDFs and moments are the observables of interest [@BT11; @RM10_1].
The time average of the modulation of turbulence (or the most probable value [@BT11]) shares many properties of pattern formation or of the average of an order parameter in a critical phenomenon [@prigent02; @RM10_1]. This led to the proposal of Ginzburg–Landau types of models, for an order parameter $A$, in order to describe the deterministic part of the evolution of the amplitude of modulation of turbulence $m$. The modeling of its fluctuations, and the noise term that should be added to these evolution equations, are not that simple.
Indeed, in the case of the amplitude of turbulence, it is clear that the noise level felt is a growing function of said amplitude, and that both go to zero together. Several types of power law dependencies have been proposed [@bp; @cras]. This fact is consistent with the assertion that transition to turbulence in shear flows is extremely similar to direct percolation (DP) [@shi]. Indeed, in the Langevin description of reaction diffusion models such as DP, the correlations of the noise are proportional to the fields amplitudes [@car1]. These Langevin equations are derived from the field theories constructed to study the phase transitions in such models [@car2]. Meanwhile, some authors proposed additive gaussian stochastic forcing for the modeling of the small scale rolls and streaks [@S3T]. Note, however, that the covariance of their noise is directly dependent on the mean kinetic energy.
In the case of the amplitude of modulation of turbulence, additive white noise has been included in the first model equation for the band on qualitative grounds [@prigent02; @phD]. The comparison with experimental results is satisfactory. The physical idea motivating this model is that the noise amplitude is still related to the amplitude of turbulence, which can be thought of as a constant when considering the amplitude of modulation of turbulence. This assumes of course that $R$ is not too close to $R_{\rm g}$. This type of modeling has been tested more quantitatively when considering the random walk of the relative position of the band [@RM10_1], or the orientation fluctuations [@RM10_2], in numerical simulations. However, only the later study revealed the the fluctuations of the amplitude of modulation. The range of Reynolds numbers over which the lifetime in one orientation could be compared to the Arhenius type law predicted by the model was not large enough for the noise modeling to be entirely validated.
The precise description of the type of noise felt by the amplitude of modulation is fundamental to the understanding of the disappearance of the modulation at $R_{\rm t}$. In particular, the assertion that this phenomenon could be described by a critical phenomenon at equilibrium depends on the noise being additive and not multiplicative [@RNL]. Therefore, in order to fully justify the use of stochastic amplitude equation models for the laminar turbulent coexistence, each of the prediction of said model must be compared to numerical or experimental data.
In this article, we will endeavour to test said predictions on the temporal evolution of the modulation of turbulence under its own noise. We will take two points of views: that of the fluctuations of well established bands around their average, and the noisy growth of bands from uniform turbulence. Indeed, if one excludes the case of switching of orientation, the bands only evolve in these two situations. In order to do so, we proceed in the following manner. In section \[sgl\], we remind the stochastic Ginzburg–Landau model, which we analyse analytically and numerically in the two cases of interest. In section \[num\], we present our numerical procedure, and the processing of numerical data. We then analyse the numerical simulations in section \[res\] and compare the results to theory. We eventually discuss the results altogether in the conclusion (§ \[conc\]).
The Stochastic Ginzburg–Landau model \[sgl\]
============================================
Principle of the model\[ssp\]
-----------------------------
### The equation
We introduce the stochastic Landau limit of the Ginzburg–Landau–Langevin models used to describe the oblique bands in Taylor–Couette flow [@prigent02; @phD] and plane Couette flow [@RM10_1; @RM10_2]. This model describes the evolution of two complex fields $A_\pm=A^r_\pm+\imath A^i_\pm$ and reads: $$\tau_0 \partial_t A_\pm=\left(\epsilon-g_1|A_\pm|^2-g_2|A_\mp|^2\right)A_\pm+\xi_x^2\partial_x^2A_\pm+\xi_z^2\partial_z^2A_\pm+(\bar{\zeta}_\pm^{\rm r}(x,z,t)+\imath \zeta_\pm^{\rm i}(x,z,t)) \label{GLL}\,,$$ where $\tau_0$ is the time scale, $\xi_{x,z}$ are the correlation lengths, $g_{1,2}$ are positive inverse saturation amplitudes and $\bar{\zeta}_{\pm}^{\rm i,r}$ is a white noise in time and space of variance $\bar{\alpha}^2$ [@gar]: $$\langle \bar{\zeta}_i^j(x,z,t)\bar{\zeta}_k^l(x',z',t')\rangle=\bar{\alpha}^2\delta(x-x')\delta(t-t')\delta(z-z')\delta_{ik}\delta_{jl}\,.$$ We use brackets $\langle .\rangle$ for the ensemble averages. The moduli $|A_\pm|$ describe the amplitude of modulation of turbulence at the corresponding orientation $m_\pm$. The phases describe their relative positions in the flow.
This type of equation, with $\bar{\alpha}=0$, arises in weakly non-linear instabilities, particularly in pattern formation, when one performs a projection on the central manifold, that is, an elimination of the fast evolving space and time scales [@CH]. This type of equation arises in phase transitions as well, with $\bar{\alpha}=\sqrt{2k_BT}$, $T$ being the temperature and $k_B$ the Boltzmann constant, where it is often termed Model A (gradient system without a conservation law for $A_\pm$) [@HA].
The actual framework in which the band falls is unclear, although it contains elements from these two fields of physics. Indeed, even though it is an hydrodynamical phenomenon, the mechanisms of sustainment appear to be rather different from classical pattern formation instabilities [@PRL; @ispspot; @BT07]. No derivation from first hydrodynamical principles has been proposed. The type of coexistence, quasi steady bands or generally intermittent, depends a lot on the size of the domain [@PM]. The use of the Ginzburg–Landau model to describe the bands is therefore in the same line of thought as the use of the Landau theory to describe phase transitions. The equation is obtained through an experience of thought: a “coarse graining”: this is an average over the fast spatial scales of the actual hydrodynamics. The equation is then used to describe and understand the main physical properties of the system. Indeed, the formal resemblance between amplitude of modulation $m_\pm$ and $A_\pm$: slow spatial variation, symmetries, noisy signal, phase invariance, no superposition of $+/-$ fields, continuous apparition at $\epsilon=0$, justified the proposition of such a model to match $m_\pm$, with $\epsilon\equiv1-R/R_{\rm t}$ [@prigent02; @phD]. Many experimental and numerical data were perfectly matched by the model, and gave values for constants $g_1$, $g_2$ and $\xi_z$, ranges for $\tau_0$ and estimates of $\alpha$.
The strength of such stochastic models is that they are easily analysed for large scale systems [@bp]. We can make predictions that can then be verified in laboratory or numerical experiments. In the case of plane Couette flow, the timescales of the competitions between domains of $+$ and $-$ orientation (see [@DSH]) may very well have time scales deriving from coarsening dynamics (as noted in [@EPJB]). Said dynamics are typical of competitions between domains of $+$ and $-$ in models like the GL equation (Eq. (\[GLL\])) [@bray]. More surprisingly, the very intermittent behaviour at the disappearance of modulation may be very similar to a critical phenomenon [@RNL].
### Evolution equation of the modulus
Since our DNS are performed in a domain containing a few wavelengths of the band, we move to the Landau version of the Ginzburg–Landau–Langevin model [@RM10_1]: $$\tau_0 \partial_t A_\pm
=\underbrace{\left(\epsilon-\xi_x^2\delta k_x^2-\xi_z^2\delta
k_z^2\right)}_{\tilde{\epsilon}} A_\pm
-g_1|A_\pm|^2A_\pm-g_2|A_\mp|^2A_\pm+(\zeta^{\rm r}_\pm(t)+\imath \bar{\zeta}_\pm^{\rm i})\,.
\label{eq10}$$ The main difference with model (\[GLL\]) lies in the fact that $A_\pm$ is now only a function of time. The spatial derivatives are replaced by the distances to the optimal wavenumber of the band: $\delta k_{x,z}=k_{x,z}-k_{x,z}^{\rm c}$, with $k_{x,z}^{\rm c}$ both depending on $R$ [@prigent02; @RM10_1]. This modification is the same as the one performed in pattern formation, when a modulation of wavelength $\lambda_c=2\pi/k_c$ arises [@CH]. There is one subtle difference: this equation is derived from a spatial average of equation \[GLL\], the noise is changed from $\bar{\zeta}(x,t)$ with variance $\bar{\alpha}^2$ to $\zeta(t)$ with variance $\alpha^2=\bar{\alpha}^2/(L_xL_z)$. This type of noise forcing should be expected for spatially averaged quantities. For instance, in the case of the average kinetic energy, this is consistent with the finding that their probability density functions $\rho_E(E)$ have Large Deviations in the limit of large size [@EPJB] : $$-\frac{1}{L_xL_z}\ln(\rho_E)=I(E)\,.\label{LDP}$$ The function $I(E)$ is termed a large deviation function [@ht]. It appears to be a parabola in a very large range of Reynolds numbers, making $E$ a gaussian random variable.
In a statistically steady state, one finds one solution of equation (\[eq10\]) of spatial and ensemble average $\langle A_\pm\rangle =0+O(\alpha)$ if $\tilde{\epsilon} <0$. This solution describes uniform turbulence. One finds two solutions $\langle|A_\pm|\rangle = \sqrt{\frac{\epsilon}{g_1}}+O(\alpha)$, $A_\mp =0+O(\alpha)$ provided $g_2>g_1$ if $\tilde{\epsilon} >0$. These two solutions describe a flow where one orientation of the band is dominant and small traces of the other orientation are found, due to turbulent noise [@RM10_1].
This model can be written using the potential $\bar{V}$: $$\bar{V}=-\frac{\tilde{\epsilon}}{2}\left(|A_+|^2+|A_-|^2
\right)+\frac{g_1}{4}\left(|A_+|^4+|A_+|^4
\right)+\frac{g_2}{2}|A_+|^2|A_-|^2 \label{eq12}\,.$$ The interest of a potential approach lies in the mean field treatment of the problem: *i.e.* the assimilation of the average to the most likely value when relative fluctuations are small [@RM10_1; @LL].
If one is concerned by the dynamics of the modulus $A$ of only one orientation, because the other is negligible or the interactions are weak, the last non-linear coupling term can be neglected. In order to obtain the equation describing the time evolution of the modulus of $A$, one must make a change of variable $(A_r,A_i)\rightarrow A,\phi$ (dropping the $\pm$). It is non-linear and implies some subtleties. Indeed, a stochastic process of the form: $$\partial_t x=F(x)+G(x)\zeta(t)\,,$$ with $\zeta$ a white noise, can have several meaning which are explicit only when time is discretised. The most common definition is termed the Itô rule, where the discretised equation reads: $$x(t+dt)-x(t)=F(x(t))+G(x(t))(\zeta(t+dt)-\zeta(t))\,.$$ One can demonstrate that performing a change of variables in the same manner as for an ordinary differential equation yields a different physical process (see for instance [@gar] for a detailed discussion).
In order to perform the change of variables: a first strategy to change variables can be deriving the Fokker–Plank equation for the modulus and phase, and then deducing the Langevin equation for the modulus with a Itô rule. One can alternatively use the Itô formula directly in the Langevin equation (see [@gar] § 4.5.5 for a similar derivation). Both yields an over damped Langevin Equation: $$\tau_0 \partial_t A=-\partial_A V+\frac{\alpha^2}{2A}+\zeta_2(t) \,, \,
V=-\frac{\tilde{\epsilon}}{2}A^2+\frac{g_1}{4}A^4\,,\label{eqchvar}$$ with $\zeta_2$ a Gaussian white noise of variance $\alpha^2$. The main consequence of the non-linear change of variable is the introduction of the drift $\frac{\alpha^2}{2A}$, which prevents the modulus from reaching $0$, a property of the modulus of a complex random variable. This is the dynamical equivalent of the pdf going to zero for $A=0$ [@RM10_1; @BT11]. The approach and the results are in the same spirit if the coupling $g_2$ is taken into account.
Computation of the autocorrelation function \[Ansc\]
----------------------------------------------------
We first study the time evolution of the modulus of the order parameter $A$ around the steady state $\langle A\rangle =\sqrt{\tilde{\epsilon}/g_1}$ for $\tilde{\epsilon}>0$. For that matter, we use the time autocorrelation function : $\langle f(t) f(t')\rangle$ with $f=A-\langle A\rangle$. This function contains information that simpler estimators does not. Firstly, the amplitude of this function indicates the corresponding amplitude of the time fluctuations of $A$, and the decay time of this function indicates the correlation time of $A$. Secondly, when examined in more details, the shape of this function indicates in which regime the transitional flow is. In this section we examine the case where relaminarisation and orientation fluctuations are so rare that one can consider that they do not occur. This will yield a specific type of correlation function which goes hand in hand with the parabolic shape of the large deviation function $I$ (Eq. (\[LDP\])) [@EPJB]. Note that near $R_{\rm g}$, where relaminarisations can occur and the correlation function, along with the large deviation function, is quite different [@cras; @EPJB].
### Analytical treatment
The evolution of the order parameter around equilibrium is given by equation \[eqchvar\]. One can neglect the drift term in the evolution equation, which now reads: $$\tau_0 \partial_t f =-e f +\zeta_2(t)\,,$$ with $e=2\tilde{\epsilon}$, if the evolution is well described by the potential (Eq. (\[eqchvar\])). The expansion around equilibrium, with a decay rate $e$, is still valid, even when one enters the range of $R$ in which the deterministic part of the evolution equation is more complex than the Landau–Langevin equation [@RM10_1]. The solution of the equation is: $$f(t)=\frac{1}{\tau_0}\int_0^t {\rm d}t' \exp\left(-\frac{e}{\tau_0}(t-t')\right)\zeta_2(t')\,.\label{solsc}$$
One has the product $f(t)f(t')$: $$f(t)f(t')=\int_0^t\int_0^{t'}{\rm d}t''{\rm d}t'''\left(\frac{1}{\tau_0^2} \zeta_2(t'')\zeta_2(t''') \times\exp\left(-\frac{e}{\tau_0^2}\left(t+t'-t''-t''' \right) \right)\right)\,.$$ Taking the ensemble average, it yields: $$\langle f(t)f(t')\rangle=\int_0^t\int_0^{t'}{\rm d}t''{\rm d}t''' \left(\frac{1}{\tau_0^2} \alpha^2 \delta (t''-t''') \times\exp\left(-\frac{e}{\tau_0}\left(t+t'-t''-t''' \right) \right)\right)\,.$$ The time $t$ is taken smaller than $t'$ (the choice is of no consequence to the rest of the derivation): $$\notag\langle f(t)f(t')\rangle=\frac{\alpha^2}{\tau_0^2}\exp\left(-\frac{e}{\tau_0}(t+t') \right) \int_0^t{\rm d}t'' \exp\left(-\frac{e}{\tau_0}\left(-2t'' \right) \right)\,,$$ *i.e.* $$\langle f(t) f(t')\rangle=\frac{\alpha^2}{2e\tau_0} \exp\left( -e\frac{(t+t')}{\tau_0}\right)\left(\exp\left(2e\frac{t'}{\tau_0}\right)-1\right)
\,.\label{auto}$$ A few simplifications can be performed if $t=t'$: $$\langle f(t) f(t)\rangle=\frac{\alpha^2}{2e\tau_0} \left(1-\exp\left(-\frac{2et}{\tau_0}\right)
\right)\,,\label{limapprox}$$ and if $t$ and $t'$ are very large compared to $\tau_0/e$: $$\langle f(t) f(t')\rangle=\frac{\alpha^2}{2e\tau_0} \exp\left( -\frac{e}{\tau_0}|t-t'|
\right)\,.\label{auto_}$$ This yields classical exponentially decreasing time correlations.
### Numerical treatment of the model
The autocorrelation function can be computed from ensemble averaging of the time series, divided in $N$ times series of duration $T_0$, in the same way as the fluctuations of the phase [@RM10_1]: $$\notag \langle f(t)f(t')\rangle =\frac{1}{N}\sum_{i=1}^N\left[
\left(f(t+(i-1)T_0)-f((i-1)T_0)\right) \left(f(t'+(i-1)T_0)-f((i-1)T_0)\right)\right]\,.\label{flucmod}$$ The analytical result and the processing procedure can be validated by the numerical simulations of the Landau model. In a test case, for arbitrary values of the parameters ($\tau_0=1$, $\tilde{\epsilon}=0.29$, $\alpha=0.002$, $g_1=1$, $g_2=2$), we compute the autocorrelation in the $t,t'$ plane (Fig. \[fig38\] (a)). In figure \[fig38\] (b) we display the logarithm of the autocorrelation function as a function of $|t-t'|$ by varying $t$ (resp. $t'$) and keeping $t'$ constant (resp. $t$). The logarithm of the autocorrelation function along a $t$ or $t'$ constant line confirms that the prediction of an exponential decrease is good (Fig. \[fig38\] (b)). A fit of the logarithm by $-a|t-t'|+b$, yields $a=0.19$ and $b=-12$, to be compared respectively to $2\tilde{\epsilon}/\tau_0= 0.58$ and $ \ln(\alpha^2/(4\tilde{\epsilon}\tau_0))=-12.2$. The comparison to the value of $\alpha^2/{4\tilde{\epsilon}}\tau_0$ shows an error of only $15$%. However, the error is much larger in the case of $2\tilde{\epsilon}/\tau_0$: although the laws are correctly predicted by analytics, there is an overestimate of the constant $\tau_0$. We will be wary of that fact in the analysis of the numerical simulation of the Navier–Stokes equation.
![Result of the time integration of the Landau–Langevin model : (a) autocorrelation function $\ln\langle f(t)f(t')\rangle$ in the plane $t$, $t'$, (b) : autocorrelation as a function of $|t-t'|$.[]{data-label="fig38"}](autocor_xp7_1.eps "fig:"){width="6cm"} ![Result of the time integration of the Landau–Langevin model : (a) autocorrelation function $\ln\langle f(t)f(t')\rangle$ in the plane $t$, $t'$, (b) : autocorrelation as a function of $|t-t'|$.[]{data-label="fig38"}](autocorr_tranche_xp7_1.eps "fig:"){width="6cm"}
The growth of the order parameter \[s3T\]
-----------------------------------------
### Principle
We now study the growth of the order parameter from the original noise. This models the appearance of the band from uniform turbulence. Our goal is to propose a description of the growth that goes beyond the exponential growth and to describe the non linear effect and the effect of noise on the growth. In practice this amounts to the solving of equation (\[eqchvar\]): $$\tau_0 \partial_t A=\tilde{\epsilon} A-g_1A^3+\frac{\alpha^2}{2A}+\zeta_2\,,$$ starting from a nearly zero initial condition $A_0$ . We will use two strategies, both of them being of the stochastic structural stability theory type [@S3T], and we will determine the proper “nearly zero” that describes uniform turbulence. These theories consist of computing the ensemble average trajectory $A(t)$ and the fluctuations around them, $\sigma(t)$. Most trajectories will be contained within $\pm\sqrt{\langle \sigma^2 \rangle}(t)$ of the average. This function indicates the shape that a group of independent realisations of the growth display when put together.
The first approach consists of a mean field approximation: the fluctuations, of the order of the noise $\alpha$, are assumed to be small relative to the average. The ensemble average $A$ verifies the equation where the noise $\zeta$ is removed, with the initial condition $A_0$. Meanwhile the fluctuations $\sigma(t)$ are treated as a linear perturbation of the deterministic part of the equation , of the same order as the noise. It is somewhat artificial but the qualitative content appears more clearly.
The second approach consists of solving in the limit of $\epsilon$ not too small compared to $\egb=\alpha^2g_1$ by using an expansion in $\egb/\tilde{\epsilon}$: $$A(t)=B_0(t)+\sqrt{\frac{\tilde{\epsilon}}{g_1}}\sum_{n=1}^{\infty}\left(\frac{\egb}{\tilde{\epsilon}}\right)^\frac{n}{2}B_n(t)\,,
\label{exp}$$ with the initial conditions $B_{n>0}(t=0)=0$ and $B_0(t=0)=A_0$. Due to the expansion, each $B_{n>0}$ is the solution of a linear Langevin equation, with time varying coefficients. Since all of them are real, no resonance appears and solvability conditions are unnecessary. This expansion is more rigorous and robust in terms of order of approximation but less tractable. It mainly helps to validate the mean field strategy. Both solutions introduce the dimensionless parameter $\ggb\equiv2\alpha^2g_1/\tilde{\epsilon}^2$, a Ginzburg criterion type term [@LL], which weighs the fluctuations $\alpha/\sqrt{\tilde{\epsilon}}$ (see § \[fluc\] or [@RM10_1]) over the amplitude of the average $\sqrt{\epsilon/g_1}$. The mean field approach is valid and the expansion converges rapidly only if $\ggb$ is small.
In order to understand the features of the expanded solution, one can truncate the expansion at $n=2$ and compute $B_0$, $B_1$ and $B_2$. Indeed, at the first non-zero order in noise, the average $\langle A\rangle(t)$ only retains the terms $\langle B_0 \rangle+\alpha^2\langle B_2\rangle$, while the fluctuations are $\langle (A-\langle A\rangle)^2\rangle=\langle (B_0-\langle B_0\rangle)^2\rangle+\alpha^2\langle B_1^2\rangle$. This is because the first order contains $A_0$, which is decorrelated from $\zeta$, $B_1$ contains $\zeta$ and $B_2$ contains a deterministic term and $\zeta^2$. This order of approximation is a more rigorous formulation of the first method.
We proceed in the following way. We first compute the distribution of initial conditions. We then present the ensemble averages and fluctuations solutions. We eventually summarise the analysis and discuss the specificity of the evolution of the solutions.
### Distribution of initial conditions \[fluc\]
We first compute the initial condition $A_0$. We follow the mean field approximation of earlier studies of the bands [@RM10_1]. In a former article [@RM10_1], the average of the order parameter in the Landau–Langevin limit was obtained in the mean field limit by computing the minimum of the potential $V$ (Eq. (\[eq12\])). In the spirit of the mean field, the pdf $\rho$ is approximated by a Gaussian centered around the minimum of $V$: $$\rho\propto \exp(-(A-\langle A\rangle)^2/(2\gamma^2)\,.\label{ipdf}$$ If turbulence is uniform, one finds the minimum $\langle A\rangle =\alpha_</\sqrt{2|\tilde{\epsilon}_<}|$. We use $\alpha_<$ and $\tilde{\epsilon}_< <0$ so as to differentiate them from the value they take during the growth of the band $\tilde{\epsilon}>0$. In order to obtain the fluctuations $\gamma$, one simply has to expand $V$ to the second order and identify the corresponding coefficient with $\gamma$. One then finds: the fluctuations $\gamma= \alpha_</\sqrt{2|\tilde{\epsilon}_<|}$. In order to weigh the amplitude of the initial condition relatively to that of the final value, we introduce another dimensionless parameter $r\equiv A_0 \sqrt{g_1/\tilde{\epsilon}}$.
### Mean field solution
Starting from equation: $$\tau_0 \partial_t A=\underbrace{\tilde{\epsilon} A-g_1A^3+\frac{\alpha^2}{2A}}_{F(A)}+\zeta_2\,,$$ we apply the mean field approach.
In order to obtain the average, one has to solve: $$\tau_0 \partial_t A=\tilde{\epsilon} A-g_1A^3+\frac{\alpha^2}{2A}\,,
\label{eqmn1}$$ with the initial condition $A(t=0)=A_0$. The full resolution is somewhat involved, we describe it in appendix \[Afull\]. The solution is given by equation (\[sol\_gen\]).
If we consider the noise $\alpha$, $\alpha_<$, or the amplitude of the non linearity $g_1$ to be small but non zero, we can simultaneously take the limit $r\propto \ggb\rightarrow0$ and keep the first non zero order in equation (\[solrw\]). This limit corresponds to the first stage of the growth. This gives the equation: $$\notag A(t)=A_0\exp\left(\frac{\tilde{\epsilon} t}{\tau_0}\sqrt{1+\ggb}\right) \sqrt{(1-\frac{\alpha^2}{2\tilde{\epsilon} A_0^2})+\frac{\alpha^2}{2\tilde{\epsilon} A_0^2}\exp\left(-\frac{2\tilde{\epsilon} t}{\tau_0}\sqrt{1+\ggb}\right)}\,.\label{A_temps}$$ This reveals an interesting effect: it shows an increase of the growth rate with the noise when $\tilde{\epsilon}$ is small, embodied by $\ggb$. It strengthens the observation that the noise in the numerical solution of the stochastic GL equation modifies the growth rate [@phD]. Besides, it contains the trace of the fluctuations at $t=0$, that decreases exponentially with a rate $2\tilde{\epsilon}\sqrt{1+\ggb}/\tau_0$: the second term in the square root should decrease before we have clear exponential growth.
If one takes $\ggb =0$ in equation (\[solrw\]), we find the classical solution of the Landau equation : $$A(t)=\sqrt{\frac{\tilde{\epsilon}}{g_1}}\frac{1}{\sqrt{1+\left( \frac{\tilde{\epsilon}}{g_1A_0^2}-1\right)\exp{-\frac{2\tilde{\epsilon} t}{\tau_0}}}}\,,\,
A_0 \ll \sqrt{\frac{\tilde{\epsilon}}{g_1}}\rightarrow A(t)\simeq A_0\exp\left(\frac{\tilde{\epsilon} t}{\tau_0}\right)\,,\, A\approx \sqrt{\frac{\tilde{\epsilon}}{g_1}} \rightarrow A=\sqrt{\frac{\tilde{\epsilon}}{g_1}}\left(1-\frac{1}{2}\exp(-\frac{2\tilde{\epsilon} t}{\tau_0}) \right) \,. \label{A_temps_bis}$$ The first approximation described the growth of $A$ away from $0$. The second describes the convergence toward $\sqrt{\tilde{\epsilon}/g_1}$, this contains a simplified version of the non-linear effects of the third stage, *i.e.* saturation, without the competition between orientations. The average in the expansion contains the same physics as in the mean field approximation (§ \[Aexp\], Eq. (\[dampa\])). It validates this approximation for $\ggb$ small.
The fluctuations $\sigma (t)$ are obtained by solving: $$\tau_0 d_t \sigma=\left. \frac{dF}{dA}\right|_{A(t)}\sigma+\zeta\,.\label{eqmn2}$$ One can write $\sigma=\lambda C_1$, with $\tau_0d_t C_1=dF/dA|_{A}C_1$ and $C_1(t=0)=1$ and $\lambda(t)$ verifying $C_1\tau_0d_t \lambda=\zeta$ and $\lambda(t=0)=0$. One simply finds: $$C_1(t)=\exp\left(\int_{t'=0}^t{\rm d}\,t'\frac{1}{\tau_0}\left. \frac{dF}{dA}\right|_{A(t')}\right)\,,\,\lambda(t)=\frac{1}{\tau_0}\int_{t'=0}^t{\rm d}t' \zeta(t')\exp\left(-\int_{t''=0}^{t'}{\rm d}t''\,\frac{1}{\tau_0}\left.\frac{dF}{dA}\right|_{A(t'')} \right)\,,$$ which yields $\sigma$: $$\sigma(t)=\frac{1}{\tau_0}\int_{t'=0}^t{\rm d}t' \zeta(t')\exp\left(\int_{t''=t'}^t{\rm d}\,t''\left. \frac{dF}{dA}\right|_{A(t'')} \right)\,.
\label{eqb1}$$ The fluctuations have a zero average around the mean at order $1$ in $\alpha$, even in logarithmic scale, since one has: $$\ln(A+\sigma)=\ln(A)+\sigma/A+O(\alpha^2)\label{logsig}\,.$$ We compute the explicit expression of the ensemble average of $\sigma$ and $B_1$ in Appendix \[EAS\]. The calculation is the same in both cases.
### Time evolution of fluctuation type terms\[ftt\]
The time dependence of $\sigma(t)$ (Eq. (\[eqb1\])) and its moments are not obvious. We perform a few approximations to make them more explicit. At small times, $A$ is small and $dF/dA(A(t))$ can be expanded. In the first order, it is equal to $\tilde{\epsilon}$. Inserting in equation \[flucb1\], one has: $$\langle\sigma^2\rangle=\frac{\alpha^2}{\tau_0^2}\int_{t'=0}^t{\rm d}t'\exp\left(\frac{2\tilde{\epsilon}(t-t')}{\tau_0} \right)\,,$$ which can be rewritten and integrated: $$\langle\sigma^2\rangle=\frac{\alpha^2\exp\left( \frac{2\tilde{\epsilon}t}{\tau_0}\right)}{\tau_0^2} \left[-\frac{\tau_0}{2\tilde{\epsilon}}\exp\left( \frac{-2\tilde{\epsilon}t'}{\tau_0}\right) \right]_{t'=0}^t\,.$$ One eventually finds: $$\langle\sigma^2\rangle=\frac{\alpha^2}{2\tilde{\epsilon}\tau_0}\left(\exp\left(\frac{2\tilde{\epsilon}t}{\tau_0} \right)-1 \right)\,.
\label{flucres}$$ Note that this means that in the growth phase, the fluctuations of the logarithm of the order parameter $\sigma_{\ln}$ given by equation (\[logsig\]) are constant. Indeed: $$\sigma_{\ln}\equiv \sqrt{\langle \ln(A)^2-\langle \ln(A)\rangle^2\rangle}\simeq\sqrt{\langle \sigma^2\rangle}/A\label{sln}$$ is the ratio of two functions growing exponentially at the same rate.
The fluctuations $\sigma$ and $B_{1,2}$ have two successive regimes: expansion then retraction of perturbations induced by noise. From a different point of view, the first regime is controlled by linear growth and the second is controlled by non-linear saturations. The quantity: $$\int_{t''=t'}^t {\rm d}t''\exp(dF/dA|_{A(t)})\,,$$ and more particularly the sign of $dF/dA|_{A(t)}$, controls the regime in which the band is. This is exact in the case of the expansion and in good approximation in the mean-field case. Indeed, when considering equation (\[eqb1\]), one can see that the amplitude of the noise will be multiplied or divided, depending on the sign of the quantity in the exponential. The value of $t$ (and the corresponding $A(t)$ at which it changes signs can be easily estimated in order zero in noise (Eq. (\[A\_temps\_bis\])). Indeed, one has: $$\left.\frac{dF}{dA}\right|_{A(t)}=\tilde{\epsilon}-2g_1A^2(t)=\tilde{\epsilon}\frac{\left( \frac{1}{r}-1\right)\exp\left(-\frac{2\tilde{\epsilon} t}{\tau_0} \right)-1}{1+\left(\frac{1}{r}-1\right)\exp\left(-\frac{2\tilde{\epsilon} t}{\tau_0} \right)}\,.$$ The numerator cancels out provided the initial condition is small enough ($r>1$) and does so at: $$t_{0}=\frac{\tau_0}{2\tilde{\epsilon}}\ln\left(\frac{1}{r}-1 \right) \label{tps}\,.$$ One then has $\langle A(t)\rangle=\sqrt{\tilde{\epsilon}/(2g_1)}$ at the first non-zero order in noise amplitude.
At long times $t>t_{0}$, one can have a qualitative discussion about the behaviour of fluctuations (§ \[ftt\]). The rate at which they grow decreases in time, the expansion becomes slower and slower until time reaches $t_0$ (Eq. (\[tps\])), at which time non-linear effects are felt and the analytics depart from the behaviour of the flow. Indeed, they do not contain the competition between orientations $A_\pm$. The analytical description of the switching is limited to the transition from one well to another (*via* residency times [@RM10_2]). After said competition, both analytics and the flow agree again. In that last regime, one finds that the fluctuations decrease in time, and connect with the autocorrelation studied in the next section.
At longer times, the same type of expansion that has been done at small times can be done, and yields the results of the former section, that is, the autocorrelation function.
Again, the fluctuations in the expansion contain the same content as that of the mean field solution (Eq. (\[flucb1\])), and recover the exponential decay of the initial fluctuations (Eq. (\[dampa\])). Note that the time evolution of this term is deterministic (exponential decay), only its amplitude and its sign is random.
Numerical Procedure \[num\]
===========================
Numerical Simulation of Navier–Stokes equations \[chan\]
--------------------------------------------------------
We simulate the incompressible Navier–Stokes equations in the plane Couette configuration (Fig. \[fig1\] (b)). We use periodical boundary conditions in the streamwise ($\mathbf{e}_x$) and spanwise ($\mathbf{e}_z$) directions. The velocities are made dimensionless by $U$ and the sizes by $h$ and durations by $h/U$, which leaves the Reynolds number $R=hU/\nu$, with $\nu$ the kinematic viscosity as the control parameter of the flow. The whole behaviour of the flow is set by $R$ together with the sizes $L_x$ and $L_z$ [@PM].
The numerical integration is performed using the code [Channelflow]{} written by J. Gibson [@gibs]. We follow two different procedures. On the one hand, we perform classical Direct Numerical Simulations (DNS), with a resolution of $8/3$ dealiased Fourier modes per unit length in the streamwise and spanwise directions and $27$ Chebychev modes in the wall normal direction. This number of modes has proven sufficient to obtain well resolved DNS [@MR10; @dsc10]. On the other hand, we use the code to run a reduced order model. For that matter, we use $1$ dealiased Fourier mode per unit length in the streamwise direction, $2$ dealised Fourier modes per unit length in the spanwise direction and $15$ Chebychev modes in the wall normal direction. This approach gives reliable representations of the bands and does not alter the physics of the flow [@MR10; @RM10_1]. The price to pay is a decrease of the transition thresholds from $[R_{\rm q}; R_{\rm t}]\approx [325 ; 415]$ to $[275 ; 355]$ [@MR10].
For both procedures, the oblique bands (Fig. \[fig1\] (a)) are obtained in the following way : we start from a smooth random initial condition. It is then integrated in time at $R=500$ (for DNS) or at $R=450$ (low order modeling) for a duration of $500$. This gives a realistic uniformly turbulent flow. It can be used as an initial condition for a quench : a sudden decrease of the Reynolds number (§ \[gr\], see [@bddm] for laboratory experiments, or [@EPJB] for numerical experiments with our settings). Further integration in time at said Reynolds numbers can be done in order to produce decorrelated initial conditions for quenches.
In order to obtain bands at $R\in [R_{\rm g}; R_{\rm t}]$, we start from one of the uniformly turbulent initial conditions, decrease the Reynolds number to $R$, then integrate for a duration of $1500$. The flow obtained can then be used as an initial condition for a simulation of the bands at $R$.
Processing of numerical data \[proc\]
-------------------------------------
In order to follow the time evolution of the modulation of turbulence in the flow, we need a quantity which indicates whether the turbulence is organised in bands and which orientation the bands take. For that matter, we take advantage of the sinusoidal modulation of turbulence of the band [@prigent02; @BT11; @RM10_1]. Once the fundamental mode of the bands $(k_x,\pm k_z)$ is identified, two similar strategies can be followed to define a so-called order parameter. The first one is adapted if the size of the system ($L_{x,z}$) is clearly larger than the wavelengths of the bands $(2\pi/k_{x,z})$. One can then compute two Hilbert transforms of the signal measured (in numerical studies, the velocity field, or in the experimental studies the light intensity) [@prigent02; @phD]. This yields two complex functions $a_+(x,z,t)$ and $a_-(x,z,t)$ which vary more slowly in space than to the wavelength of the band. They are filtered. Their moduli give the amplitude of the modulation of the respective orientation at position $(x,z)$ at time $t$ and their phase give the relative shift of each patch of band.
The second strategy is adapted if the size of the system is comparable to the wavelengths of the bands. Following Tuckerman & Barkley [@BT11], we proposed [@RM10_1] to use the Fourier transform of the $x$ component of the velocity field to define the order parameters $a_\pm=m_\pm(t)e^{\imath\phi_\pm(t)}$ by computing: $$m_\pm^2=\frac{1}{2}\int_{y=-1}^{y=1}|\hat{u}_x|^2(k_x,y,\pm k_z){\rm d}y \,,\,
\phi_\pm=\arg\left(\hat{u}_x(k_x,0,\pm k_z\right)\,.$$ The phases $\phi_\pm$ of the order parameters give the relative position of the respective orientation of the band in the domain, while the moduli of the order parameters give the amplitude of the modulation of the respective orientation. In this article, we follow this approach. Since we do not study the behaviour of the phases, we only use $m_\pm$, termed the order parameters by an abuse of language. This quantity is to be compared to the modulus of the order parameter $|A_\pm|$ studied theoretically in the former section. The order parameters are the ideal quantities to study the behaviour of the band itself: unlike the turbulent fraction, they capture the spatial organisation of turbulence.
Numerical results \[res\]
=========================
In this section, we present the results of the numerical simulations and compare them with theory. We follow the same approach as in the former theoretical section, firstly considering the evolution around the equilibrium position (§ \[eq\]) and secondly detailing the growth of the order parameter (§ \[gr\]).
Around equilibrium \[eq\]
-------------------------
We first examine the fluctuation of the band around its equilibrium, using the time autocorrelation function. For that matter, we produce bands using our procedure. Since the time series required are extremely long, we use the reduce order procedure. We sample a time series of the order parameter $m$ of duration $400000h/U$ at $R=290$, far from $\rt=355$. This Reynolds number corresponds to $\tilde{\epsilon}\simeq0.18$. This ensures that the orientation will not change. The size of the domain is $L_x\times L_z=110\times 64$, in order to accommodate one band.
![Processing of the time series of the order parameter extracted from numerical simulations of the Navier–Stokes equations at $R=290$, $N_y=15$ : (a) autocorrelation function in the $t,t'$ plane, (b) : logarithm of the autocorrelation function as a function of $|t-t'|$, (c) : autocorrelation function at $t=t'$.[]{data-label="fig38_"}](autocorr_contour_0.eps "fig:"){width="6cm"}![Processing of the time series of the order parameter extracted from numerical simulations of the Navier–Stokes equations at $R=290$, $N_y=15$ : (a) autocorrelation function in the $t,t'$ plane, (b) : logarithm of the autocorrelation function as a function of $|t-t'|$, (c) : autocorrelation function at $t=t'$.[]{data-label="fig38_"}](loglin.eps "fig:"){width="6cm"}![Processing of the time series of the order parameter extracted from numerical simulations of the Navier–Stokes equations at $R=290$, $N_y=15$ : (a) autocorrelation function in the $t,t'$ plane, (b) : logarithm of the autocorrelation function as a function of $|t-t'|$, (c) : autocorrelation function at $t=t'$.[]{data-label="fig38_"}](autocorr_tt_R290.eps "fig:"){width="6cm"}
The autocorrelation function is computed from the time series using equation \[flucmod\]. The results are presented in figure \[fig38\_\]. The whole autocorrelation function is displayed in figure \[fig38\_\] (a). We display the logarithm of the autocorrelation function as a function of $|t-t'|$ in figure \[fig38\_\] (b) by varying $t$ (resp. $t'$) and keeping $t'$ (resp. $t$) constant (in the same manner as Fig. \[fig38\] (b)). Moreover, we display the the correlation function as a function of $t=t'$ in figure \[fig38\_\] (c). The results are well fitted by the logarithm of equation (\[auto\_\]) (linear decrease, Fig. \[fig38\_\] (b)), and equation (\[limapprox\]) (exponential convergence, Fig. \[fig38\_\] (c)). This indicates that the time evolution of the amplitude of modulation is very well described by the Ornstein-Uhlenbeck process derived from the GL model (Eq. \[GLL\]).
The fit yields the values $\tau_0\simeq 68$ (if $t=t'$) and $\tau_0\simeq 257$ (function of $|t-t'|$) for the characteristic time and $\alpha^2/(4\tilde{\epsilon}\tau_0)\simeq 2\cdot10^{-5}$, directly related to the amplitude of the noise felt by the bands. One finds discrepancies for $\tau_0$, which calls for an alternate estimation. Indeed, the study of the random walk of the phase yields $\alpha/\tau_0=4\cdot10^{-4}$ [@RM10_2]. In order to estimate the amplitude of the noise, a value of $\tau_0=30$ (found for counter rotating Taylor–Couette flow) was used, yielding $\alpha\simeq 0.012$. The results we have here allow for an estimation of $\alpha$ independently of $\tau_0$, in a situation where the validation of the procedure and the analytics by numerics showed no incertitude. This yields $\alpha\simeq0.036$. Knowing $\alpha$ allows for estimation of $\tau_0\simeq90$ independently of the uncertain fits.
Growth \[gr\]
-------------
In order to study the appearance and organisation of the bands, we monitor the growth of the order parameter $m_\pm$ in DNS of quenches in a domain of size $L_x\times L_z=110\times 32$, from $R=500$ to $R=370$, with the high resolution. The uniformly turbulent initial condition is integrated up to $T=275$ for sixteen decorrelated initial conditions. Time series of the logarithm of the square of the order parameter are displayed together and discussed (Fig. \[trempe5\]). There are thirty two curves, since both orientations are concerned by the growth.
The amplitude of modulation $m_\pm$ starts from a small but non-zero value. In a first stage, it fluctuates around a constant value, as expected from the analysis of section \[fluc\]. This stage corresponds to the adjustment of the coherent structures to the new Reynolds number and the formation of laminar holes [@EPJB]. The amplitude of the modulation $m^2$ then grows randomly inside a flaring envelope in the second stage. However, in logarithmic scale, the time series of $m_\pm$ taken together reveal a well defined bundle of curves with a constant thickness and a definite slope as noted from equations (\[A\_temps\_bis\]), (\[logsig\]) and (\[flucres\]) (Fig. \[trempe5\]). Eventually, the order parameters saturate, and a competition between each orientation takes place. This happens at a time $t$ which is approximately the half of the duration of the establishment of the bands, as noted from equation (\[tps\]). The competition is ultimately won by one of the two orientations, both of them having the same probability to arise. We compute the average of all the experiments $\langle m(t) \rangle$ as well as the standard deviation $\sigma_m(t)$ at each time. This gives us an estimate of the average evolution. In that case, the average is $\langle m(t) \rangle$ and the fluctuations around it are $\langle m(t) \rangle \pm \sigma_m$. These are indicated in logarithmic scale by the red lines in figure \[trempe5\]. Note that most time evolutions are contained within $\langle m\rangle \pm \sigma_m(t)$. These DNS results already show qualitative agreement with the S3T analysis of the growth (§ \[s3T\]). Indeed, the initial condition is non zero and distributed with a given variance. The order parameter first grows around an exponential trend with fluctuations of exponential amplitude around it. The logarithm of $m$ displays a linear tendency with constant fluctuations around it. The competition between orientation is not well predicted by analytics. Note that the growth is relatively fast, since $R=370$ is quite far away from $R_{\rm t}$ at this resolution.
![Logarithm of the square of the order parameter. The bright green lines corresponds to the time series. The bright red lines indicate the average. The bright red lines with upward and downward trianglesindicate the average plus/minus the fluctuations. The dashed red line indicates the linear fit of the data in the relevant range.[]{data-label="trempe5"}](m2_log_32_moy_fluc_fit.eps){width="7.5cm"}
In order to perform a full quantitative comparison between the theory and DNS, we perform systematic quenches for a large range of $R$ with the low order procedure, since we need a very large number of repetitions. The simulations are performed in a domain of size $128\times 48$, which easily accommodate a band for all $R\in[R_{\rm g}; R_{\rm t}]$. For each Reynolds number, ten decorrelated initial conditions are quenched.
In order to extract the average behaviour of $m(t)$, the average $\langle m(t) \rangle$ is fitted in logarithmic scale in a window of duration $T$. This is motivated by the fact that the duration of the first exponential stage is not *a priori* known. It should be bounded below, as noted from equation (\[A\_temps\]). We found that starting the fit $100h/U$ after the beginning of the runs was a good compromise. It should be bounded above by $t_0$ (Eq. \[tps\]). In our study the fit do not use data past $700h/U$ after the beginning of the run. This assumes that the fluctuations $\sigma(t)$ around the mean will be averaged out, as predicted by the model. The growth rate is displayed as a function of the Reynolds number, for a range of fitting window duration $T$ in figure \[taur\] (a). Note that the characteristic time of organisation of the laminar-turbulent coexistence into bands measured here is several times larger than the characteristic time of formation of laminar holes measured from earlier numerical experiments [@EPJB]. The former are of order $O(10^2)$ while the later are of order $O(10^3)$.
For each value of $T$, $1/\tau$ is linear provided $R\gtrsim 320$. However, at constant $R$, the inverse decay shows large fluctuations around an average behaviour. A large incertitude is found on the growth rate, as was the case of Taylor–Couette flow [@phD]. Following the processing of the model, we perform a fit of all the $1/\tau$ lines in the range $320 \le R \le 340$ by $(1/\tau_0)(1-R_1/\rt)$ so as to minimise the incertitude on the fitting parameters. This gives $\rt \simeq345$ and $\tau_0\simeq 37$ (black dashed lines). One can alternatively fit the results of figure \[taur\] (a) line by line, for each fitting window. The results for $R_{\rm t}$, as a function of the fitting window duration, are displayed in figure \[taur\] (b).
There is a large incertitude on the characteristic time $\tau_0$. Indeed, the fitting parameter is $1/\tau_0$, as a consequence, the incertitude on $\tau_0$ is $\pm\tau_0^2 \Delta (1/\tau_0)\simeq 13$. The time scales $\tau_0$ computed here are consistent with the most precise estimation of the former section, given the error bars. Most fitting approach will only give a range for $\tau_0$, hence a large incertitude on the value of that parameter.
![(a) : growth rate of the the order parameter, as a function of the Reynolds number ($N_y=15$) for several duration of fitting window. (b) : Threshold Reynolds number $R_{\rm t}$ as a function of the duration of the fitting window.[]{data-label="taur"}](1stau_R_f.eps "fig:"){width="6cm"}![(a) : growth rate of the the order parameter, as a function of the Reynolds number ($N_y=15$) for several duration of fitting window. (b) : Threshold Reynolds number $R_{\rm t}$ as a function of the duration of the fitting window.[]{data-label="taur"}](Rt_T_moy.eps "fig:"){width="6cm"}
We then compute systematically the fluctuations of the logarithm of $m$, $\sigma_{\ln}$ (Eq. (\[sln\])) at small time, in order to quantify the impression of figure \[trempe5\] (a,b). We use the same set of numerical experiments and compute the fluctuations in the same manner. The fluctuations as a function of time are displayed in figure \[figsig\] (a). In agreement with the analytical results, we find that $\sqrt{\langle\sigma^2\rangle}$ is a nearly constant quantity. However, it does not seem to depend on the Reynolds number. This assertion is verified in the case of the time average (Fig. \[figsig\] (b)) which shows no clear tendency. Overall, we find a value of approximately $0.5$, which should be compared to $\alpha\sqrt{g_1}/(\tilde{\epsilon}\sqrt{2\tau_0})$. Once known quantities $\tilde{\epsilon}\simeq 0.1$ and $g_1\simeq 100$ are removed, this gives us $\alpha^2/(\tau_0)=O(10^{-5})$. This shows reasonable quantitative agreement with the former section.
![(a) : Time evolution of the fluctuations of the logarithm $\sigma_{\ln}$ around the average for $R\in [310 ; 340]$, in low order simulations. (b) time average of the fluctuations of the logarithm $\sigma_{\ln}$ around the average as a function of the Reynolds number.[]{data-label="figsig"}](evol_sigma.eps "fig:"){width="8cm"}![(a) : Time evolution of the fluctuations of the logarithm $\sigma_{\ln}$ around the average for $R\in [310 ; 340]$, in low order simulations. (b) time average of the fluctuations of the logarithm $\sigma_{\ln}$ around the average as a function of the Reynolds number.[]{data-label="figsig"}](sigma_moy.eps "fig:"){width="6.5cm"}
Conclusion \[conc\]
===================
This article examined the stochastic evolution of the oblique modulation of turbulence in plane Couette flow. We performed analytical and numerical treatment of the Stochastic Landau model that describes the bands in order to predict the behaviour of the flow around its equilibrium as well as the growth of bands from uniform turbulence. We then validated this model against Direct Numerical Simulations. These results also explained observations on previous numerical simulations of the model [@phD], such as the shift of the growth rate of the order parameter by noise.
The consequences of these results are twofold. On the one hand, we can consider quenches and the formation of laminar holes and then laminar turbulent bands from uniform wall turbulence. The comparison between the current results and those of a former article shows the clear timescale separation between the formation of laminar holes and the organisation into oblique bands [@EPJB]. The former phenomenon is most likely a local process, which is caused by viscous decay and failure of sustainment of turbulence. It should be present in all shear flows, including those which do not have a spanwise extension like Hagen–Poiseuille pipe flow. Meanwhile, the second phenomenon is most likely a combination of global processes, which involves the advection of small scale structure by large scale flows and feed back on turbulence (see [@PRL; @ispspot; @BT07] and references within).
On the other hand, these results further strengthen the assertion that the amplitude of modulation of turbulence is very well described by a Ginzburg–Landau equation with an additive noise. This means that some interesting properties of this type of equation should be expected for the modulation of turbulence. In particular, this allows us to easily predict facts that should manifest themselves in very large size systems. For instance, quenches in very large systems first lead to a spatial coexistence of domains of both orientations. It is very likely that the dynamics of the domain sizes is predicted by the GL equation, which means that we would find coarsening dynamics and specific exponents for the domain sizes as function of time (see [@EPJB; @bray]). Moreover, it is also very likely that the behaviour predicted near $R_{\rm t}$: a critical phenomenon and the divergence of the response function of $m_\pm$ with size as $R\rightarrow R_{\rm t}$ should arise from the very intermittent regime of $R\simeq R_{\rm t}$. Early simulations indicate that this is very likely the case [@RNL].
All things considered, the results of this article confirm earlier assumptions on both the sustainment mechanisms of the oblique bands and their large scale behaviour. This motivates further studies in both directions.
acknowledgement {#acknowledgement .unnumbered}
===============
The author thanks the hospitality of the INLN (Université Nice Sophia Antipolis, France), where parts of this work where performed. He also thanks the GRADE language service of Goethe Univeristät, Frankfurt, for helpful comments and corrections on the manuscript. This work was granted access to the HPC and visualization resources of “Centre de Calcul Interactif” hosted by “Université Nice Sophia Antipolis”.
Full solution for the mean field approach \[Afull\]
===================================================
In order to solve equation (\[eqmn1\]), we use a separation of variables, which gives: $$\int_{A_0}^{A(t)}\frac{AdA}{\left(\frac{\alpha^2}{2}+\tilde{\epsilon} A^2-g_1A^4 \right)}=\frac{t}{\tau_0}$$ The fraction in the integral can be rewritten as: $$\frac{1}{2\tilde{\epsilon}\sqrt{1+\frac{2\alpha^2g_1}{\tilde{\epsilon}^2}}}\left(\frac{1}{A+\sqrt{\frac{\frac{\tilde{\epsilon}}{g_1}+\sqrt{\frac{\tilde{\epsilon}^2}{g_1^2}+2\frac{\alpha^2}{g_1}}}{2}}} +\frac{1}{A-\sqrt{\frac{\frac{\tilde{\epsilon}}{g_1}+\sqrt{\frac{\tilde{\epsilon}^2}{g_1^2}+2\frac{\alpha^2}{g_1}}}{2}}} -\frac{1}{A+\sqrt{\frac{\frac{\tilde{\epsilon}}{g_1}-\sqrt{\frac{\tilde{\epsilon}^2}{g_1^2}+2\frac{\alpha^2}{g_1}}}{2}}}
-\frac{1}{A-\sqrt{\frac{\frac{\tilde{\epsilon}}{g_1}-\sqrt{\frac{\tilde{\epsilon}^2}{g_1^2}+2\frac{\alpha^2}{g_1}}}{2}}}
\right)$$ The integral is then straightforward and leads to: $$\ln\left(\left(\frac{A^2(t)-\frac{\frac{\tilde{\epsilon}}{g_1}+\sqrt{\frac{\tilde{\epsilon}^2}{g_1^2}+2\frac{\alpha^2}{g_1}}}{2}}{A^2_0-\frac{\frac{\tilde{\epsilon}}{g_1}+\sqrt{\frac{\tilde{\epsilon}^2}{g_1^2}+2\frac{\alpha^2}{g_1}}}{2}}\right) \left(\frac{A^2_0-\frac{\frac{\tilde{\epsilon}}{g_1}-\sqrt{\frac{\tilde{\epsilon}^2}{g_1^2}+2\frac{\alpha^2}{g_1}}}{2}}{A^2(t)-\frac{\frac{\tilde{\epsilon}}{g_1}-\sqrt{\frac{\tilde{\epsilon}^2}{g_1^2}+2\frac{\alpha^2}{g_1}}}{2}} \right)\right)=\frac{2t\tilde{\epsilon}}{\tau_0}\sqrt{1+\frac{2\alpha^2g_1}{\tilde{\epsilon}^2}}$$ Note that $r$ it is of the same order of $\ggb$, $r\simeq \alpha_<\sqrt{g_1 \tilde{\epsilon}_</\tilde{\epsilon}}$. A reorganisation then leads to the solution $A(t)$. The solution then reads: $$A(t)=\sqrt{\frac{\tilde{\epsilon}}{g_1}}\sqrt{\frac{\frac{1+\sqrt{1+\ggb}}{2}\frac{r^2-\frac{1-\sqrt{1+\ggb}}{2}}{r^2-\frac{1+\sqrt{1+\ggb}}{2}}-\exp\left(-\frac{2t\tilde{\epsilon}}{\tau_0}\sqrt{1+\ggb}\right)\left( \frac{1-\sqrt{1+\ggb}}{2}\right)}{\frac{r^2-\frac{1-\sqrt{1+\ggb}}{2}}{r^2-\frac{1+\sqrt{1+\ggb}}{2}}-\exp\left(-\frac{2t\tilde{\epsilon}}{\tau_0}\sqrt{1+\ggb}\right)}}\,.\label{sol_gen}$$
In order to examine the first stages of the growth of the order parameter, one can first rewrite this solution: $$A(t)=A_0\exp\left(\frac{t\tilde{\epsilon}}{\tau_0}\sqrt{1+\ggb}\right)
\sqrt{\frac{\left(\frac{1+\sqrt{1+\ggb}}{2}-\frac{\ggb}{4r^2} \right)-\exp\left(-\frac{2t\tilde{\epsilon}}{\tau_0}\sqrt{1+\ggb}\right)\left(\frac{1-\sqrt{1+\ggb}}{2}-\frac{\ggb}{4r^2} \right)}{\left( r^2-\frac{1-\sqrt{1+\ggb}}{2} \right)\exp\left(\frac{2t\tilde{\epsilon}}{\tau_0}\sqrt{1+\ggb}\right)-\left(r^2-\frac{1+\sqrt{1+\ggb}}{2} \right)}}
\label{solrw}$$
Expansion up to order 2 \[Aexp\]
================================
In this appendix, we present the expansion of the solution up to order, in order to compare it to the mean field solution. The purpose is to validate the mean field approach.
Expanded solution
-----------------
In order to obtain the $0$th order of the expansion (\[exp\]), one has to solve: $$\tau_0\partial_t B_0=\tilde{\epsilon} B_0-g_1B_0^3=\tilde{F}(B_0)\,,$$ with the initial condition $B_0(t=0)=A_0$. The solution is simply (\[A\_temps\_bis\]).
The first order of the expansion is then solution of: $$\tau_0 d_t B_1=\left. \frac{d\tilde{F}}{dA}\right|_{B_0(t)}B_1+\zeta^1\,,$$ with $\zeta^1$ of variance $1$ (rescaled by $\alpha$), and it is obtained in the same manner as the fluctuations in the mean field approach. One has: $$B_1(t)=\int_{t'=0}^t{\rm d}t'\frac{1}{\tau_0} \zeta^1(t')\exp\left(\int_{t''=t'}^t{\rm d}\,t''\frac{1}{\tau_0}\left. \frac{d\tilde{F}}{dA}\right|_{B_0(t'')} \right)\,.
\label{eqb1_}$$ Note that it has the same structure as $\sigma$.
One can then obtain $B_2$ in the same fashion. It is solution of: $$\tau_0d_tB_2=\left.\frac{d\tilde{F}}{dA}\right|_{B_0}B_2+\underbrace{\frac{2\sqrt{\tilde{\epsilon}}}{B_0\sqrt{g_1}}+\frac{1}{2}\sqrt{\frac{\tilde{\epsilon}}{g_1}}\left.\frac{d^2\tilde{F}}{dA^2}\right|_{B_0}B_1^2}_{\mu}\,.$$ The term $2\sqrt{\tilde{\epsilon}}/(B_0\sqrt{g_1})$ comes from the drift term. The function $\mu$ is random and contains $\frac{2\alpha^2}{B_0}$ and $\zeta^1(t')\zeta^1(t'')$. It has a nonzero average. One therefore has: $$B_2(t)=\frac{1}{\tau_0}\int_{t'=0}^t{\rm d}t' \mu(t')\exp\left(\int_{t''=t'}^t{\rm d}\,t''\frac{1}{\tau_0}\left. \frac{d\tilde{F}}{dA}\right|_{B_0(t'')} \right)\,.
\label{eqb2}$$
Ensemble averages of solutions \[EAS\]
--------------------------------------
Taking the averages, one finds $\langle B_1\rangle=0$. Besides: $$\langle B_2 \rangle= \frac{1}{\tau_0}\int_{t'=0}^t{\rm d}t'\exp\left(\frac{1}{\tau_0}\int_{t''=t'}^t{\rm d}\,t''\left. \frac{d\tilde{F}}{dA}\right|_{B_0(t'')} \right) \left( \frac{2\sqrt{\tilde{\epsilon}}}{B_0\sqrt{g_1}} +\frac{1}{2}\sqrt{\frac{\tilde{\epsilon}}{g_1}}\left.\frac{d^2\tilde{F}}{dA^2}\right|_{B_0}\int_{t''=0}^{t'}{\rm d}t''\exp\left(\frac{2}{\tau_0}\int_{t'''=t''}^{t'}{\rm d}t''' \left.\frac{d\tilde{F}}{dA}\right|_{B_0(t'')} \right)\right)\label{flucb2}$$
Moving to the fluctuations $\alpha^2\langle B_1^2 \rangle\simeq\langle \sigma^2\rangle$: $$\langle \sigma^2 \rangle=\frac{\alpha^2}{\tau_0^2}\int_{t'=0}^{t}{\rm d}t'\exp\left(\frac{2}{\tau_0}\int_{t''=t'}^{t}{\rm d}t'' \left.\frac{dF}{dA}\right|_{A(t'')} \right)
\label{flucb1}$$ We can neglect the difference between $\tilde{F}$ and $F$. On can analyse the behaviour of this solution. It is $0$ at $t=0$ and has an exponential growth led by $dF/dA$ initially positive. When $dF/dA$ decreases, this growth is more limited, and if crosses $0$, the fluctuations then retract.
The effect of the distribution of initial conditions on the $0^{th}$ and in $A$ in the limit $r\rightarrow 0$ order of the expansion can be shown by writing $A_0=\langle A_0\rangle+\delta A_0$. The departure $\delta A_0$ is distributed according to the initial pdf (Eq. \[ipdf\]). This yields after expansion of $B_0$: $$\notag B_0=\underbrace{\sqrt{\frac{\tilde{\epsilon}}{g_1}}\frac{1}{\sqrt{1+\left( \frac{\tilde{\epsilon}}{g_1\langle A_0\rangle^2}-1\right)\exp{-\frac{2\tilde{\epsilon} t}{\tau_0}}}}}_{\langle B_0\rangle} \left( 1+\underbrace{\frac{\left(\frac{\tilde{\epsilon}}{g_1} \right)}{1+\left( \frac{\tilde{\epsilon}}{g_1\langle A_0\rangle^2}-1\right)\exp{-\frac{2\tilde{\epsilon} t}{\tau_0}}}\delta A_0\exp\left(-\frac{2\tilde{\epsilon} t}{\tau} \right)+\ldots}_{\langle .^2\rangle\rightarrow\sigma_{B_0}} \right)\label{dampa}$$
That correction term stays small: it is well approximated by $\langle A_0\rangle^2g_1/\tilde{\epsilon}$ at $t=0$. One can note that the additional corrections all decrease exponentially with time. The ensemble average of the expansion in $\sqrt{\egb/\epsilon}$ (Eq. \[dampa\] and particularly Eq. \[flucb2\]) has most of the content of the mean field solution. It contains the limit $r\rightarrow0$ only through $\langle B_2\rangle$. This term grows until non linearity $g_1$ is felt (Eq. (\[tps\])) and then decays. The whole expansion of $\langle B_{2p}\rangle$ terms may contain the modification of the growth rate, again of order $\ggb$.
[9]{} A. Prigent, G. Grégoire, H. Chaté, O. Dauchot, W. van Saarlos, *Large-scale finite wavelength modulation within turbulent shear slow*, Phys. Rev. Lett. **89**, 014501 (2002). J. Philip, P. Manneville, *From temporal to spatiotemporal dynamics in transitional plane Couette flow*, Phys. Rev. E, **83**, 036308 (2011). L. Shi, M. Avila, B. Hof, *Scale invariance at the onset of turbulence in Couette flow*, Phys. Rev. Let. **110**, 204502 (2013). Y. Duguet, P. Schlatter, D.S. Henningson, *Formation of turbulent patterns near the onset of transition in plane Couette flow*, J. Fluid Mech. **650**, 119–129 (2010). L. Tuckerman, D. Barkley, *Patterns and dynamics in transitional plane Couette flow*, Phys. Fluids **23**, 041301 (2011). J. Rolland, P. Manneville, *Ginzburg–Landau description of laminar-turbulent oblique bands in transitional plane Couette flow*, Eur. Phys. J. B **80**, 529–544 (2011). D. Barkley, L. Tuckerman, *Computational Study of Turbulent Laminar Patterns in Couette flow*, Phys. Rev. Let. **94**, 014502 (2005). J. Rolland *Analyse taille finie d’une crise de fluctuations dans l’écoulement de Couette plan transitionel*, Proceedings of the eighteenth Rencontres du non-linéaire, (2015). Y. Duguet, P. Schlatter, *Oblique Laminar-turbulent interfaces in plane shear flows*, Phys. Rev. Let. **110**, 034502 (2013). J. rolland, *Formation of spanwise vorticity in oblique turbulent bands of transitional plane Couette flow, part 1: numerical experiments*, Eur. J. Mech. B Flu. **50**C, 52-59 (2015). D. Barkley, L. S. Tuckerman, *Mean flow of turbulent-laminar patterns in plane Couette flow*, J. Fluid Mech. **576**, 109–137 (2007). D. Barkley, *Modeling the transition to turbulence in shear flows*, J. Phys.: Conf. Ser. **318**, 032001 (2011). Y. Pomeau, *The transition to turbulence in parallel flows: A personal view*, C. R. Mécanique, **343**, 210–218 (2014). P.-A. Rey, J. L. Cardy, *Asymptotic form of the approach to equilibrium in reversible recombination reactions* J. Phys. A: Math. Gen. **32**, 1585–1603 (1999). J. L. Cardy, U. C. Täuber, *Field Theory of branching and annihilating random walks*, J. Stat. Phys. **90**, 1–56 (1998). B. F. Farrel, P. Ioannou, *Dynamics of streamwise rolls and streaks in turbulent wall-bounded shear flow*, J. Fluid Mech. **708**, 149–196 (2012). A. prigent, G. gregoire, H. Chaté, O. Dauchot, *Long-Wavelength modulation of turbulent shear flows*, Physica D **174**, 100–113 (2002). J. Rolland, P. Manneville, *Temporal fluctuations of laminar turbulent oblique bands in transitional plane Couette flow*, J. Stat Phys. **142**, 577–591 (2011). C.W. Gardiner, *Handbook of stochastic methods for physics, chemistry and the natural sciences*, Springer (2003). M.C. Cross, P.C. Hohenberg, *Pattern formation outside of equilibrium*, Rev. mod. phys. **65**, 851–1123 (1993). P.C. Hohenberg, B.I. Halperin, *Theory of dynamic critical phenomena*, Rev. Mod. Phys. **49**, 435–479 (1977). L.D. Landau, E. Lifshitz, *Statistical physics*, Ellipses (1994). Y. Duguet, P. Schlatter, D.S. Henningson, *Formation of turbulent patterns near the onset of transition in plane Couette flow*, J. Fluid Mech. **650**, 119–129 (2010). J. Rolland, *Mechanical and statistical study of the laminar hole formation in transitional plane Couette flow*, Eur. Phys. J. B **88**:66 (2015). A. J. Bray, *Theory of phase ordering kinetics*, Adv. Phys. **51**, 481–587 (2002). H. Touchette, *The large deviation approach to statisical mechanics* Phys. Rep. **478**, 1–69, (2009). S. Bottin, O. Dauchot, F. Daviaud, P. Manneville, *Experimental evidence of streamwise vortices as finite amplitude solution in transitional plane Couette flow*, Phys. Fluids **10**, 2597 (1998). J. Gibson, J. Halcrow, P. Cvitanović, *Visualizing the geometry of state space in plane Couette flow*, J. Fluid Mech. **611**, 107–130 (2008). P. Manneville, J. Rolland, *On modelling transitional turbulent flows using under-resolved direct numerical simulations*, Theor. Comput. Fluid Dyn. **25**, 407–420 (2011). A. Prigent, *La Spirale Turbulente : motif de grande longueur d’onde dans les écoulements cisaillés turbulents*, PhD manuscript (2001).
[**Graphical Abstract accompanying the article on the EPJE webside**]{}
(21,12) (10,11.5)[Amplitude of modulation for 16 trajectories]{} (10,9)[{width="6cm"}]{} (7.85,7)(2,5.68) (8.6,7)(6,5.68) (9.45,7)(10,5.68) (10.3,7)(14,5.68) (11.15,7)(18,5.68) (2,5)[{width="4cm"}]{}(6,5)[{width="4cm"}]{} (10,5)[{width="4cm"}]{}(14,5)[{width="4cm"}]{} (18,5)[{width="4cm"}]{} (2,3.5)[{width="4cm"}]{}(6,3.5)[{width="4cm"}]{} (10,3.5)[{width="4cm"}]{}(14,3.5)[{width="4cm"}]{} (18,3.5)[{width="4cm"}]{} (2,2)[{width="4cm"}]{}(6,2)[{width="4cm"}]{} (10,2)[{width="4cm"}]{}(14,2)[{width="4cm"}]{} (18,2)[{width="4cm"}]{} (10,0.5)[Snapshot of three trajectories]{}
[^1]: LadHyX, UMR 7646 CNRS, Palaiseau 91128 France, *current address:* Insitut für Atmosphäre und Umwelt, Goethe Universität, 60438 Frankfurt, Germany [[email protected]]{}
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'This is a survey on an analogue of tropical convexity developed over the max-min semiring, starting with the descriptions of max-min segments, semispaces, hyperplanes and an account of separation and non-separation results based on semispaces. There are some new results. In particular, we give new “colorful” extensions of the max-min Carathéodory theorem. In the end of the paper, we list some consequences of the topological Radon and Tverberg theorems (like Helly and Centerpoint theorems), valid over a more general class of max-T semirings, where multiplication is a triangular norm.'
address:
- 'Department of Mathematics, West Chester University, PA 19383, USA, and Institute of Mathematics, P.O. Box 1-764, Bucharest, Romania'
- 'University of Birmingham, School of Mathematics, Watson Building, Edgbaston B15 2TT, UK'
author:
- Viorel Nitica
- Sergeĭ Sergeev
title: 'Tropical convexity over max-min semiring'
---
[^1]
Introduction\[sec1\]
====================
The max-min semiring is defined as the unit interval $\mmset=[0,1]$ with the operations $a\oplus b:=\max(a,b)$, as addition, and $a\otimes b:=\min(a,b)$, as multiplication. The operations are idempotent, $\max(a,a)=a=\min(a,a)$, and related to the order: $$\label{first-eq-662}
\max(a,b)=b\Leftrightarrow a\leq b\Leftrightarrow \min(a,b)=a.$$ One can naturally extended them to matrices and vectors leading to the max-min (fuzzy) linear algebra of [@BCS-87; @Gav-01; @Gav:04]. We denote by $\mmset(d,m)$ the set of $d\times m$ matrices with entries in $\mmset$ and by $\mmset^d$ the set of $d$-dimensional vectors with entries in $\mmset$. Both $\mmset(d,m)$ and $\mmset^d$ have a natural structure of semimodule over the semiring $\mmset$.
The [**max-min segment**]{} between $x=(x_i)_i,y=(y_i)_i\in\mmset^d$ is defined as $$\label{segm0}
\begin{aligned}
\ [x,y]_{\oplus} &= \{\alpha\otimes x\oplus \beta\otimes y\mid \,\alpha \oplus \beta =1\}.\ %\text{or}\\
%[x,y]_{\oplus}& =\{z\mid z_i=\max(\min (\alpha ,x_i),\min (\beta ,y_i))\ \forall i,\ \max \,(\alpha ,\beta )=1 \}.
\end{aligned}$$
A set $C\subseteq\mmset^d$ is called [**max-min convex**]{}, if it contains, with any two points $x,y,$ the segment $[x,y]_{\oplus}$ between them. For a general subset $X\subseteq\mmset^d$, define its [**convex hull**]{} $\conv(X)$ as the smallest max-min convex set containing $X$, i.e., the smallest set containing $X$ and stable under taking segments . As in the ordinary convexity, $\conv(X)$ is the set of all [*max-min convex combinations*]{} $$\label{convX}
\bigoplus_{i=1}^m \lambda_i\otimes x^i\colon m\geq 1,\ \bigoplus_{i=1}^m \lambda_i=1,$$ of all $m$-tuples of elements $x^1,\ldots,x^m\in X$. The max-min convex hull of a finite set of points is also called a [*max-min convex polytope*]{}.
A [**(max-min) semispace**]{} at $x\in\mmset^d$ is defined as a maximal max-min convex set not containing $x$. A straightforward application of Zorn’s Lemma shows that if $C\subseteq \mmset^d$ is convex and $x\notin C$, then $x$ can be separated from $C$ by a semispace. It follows that the semispaces constitute the smallest intersectional basis of max-min convex sets. This fact is true more generally in abstract convexity. Some new phenomena appear in max-min convexity, which further emphasize the importance of semispaces in any convexity theory. For example, separation of a point and a convex set by hyperplanes it is not always possible in max-min convexity [@Nit-09], [@N-Ser1].
The max-min segments and semispaces were described, respectively, in [@NS-08I; @Ser-03] and in [@NS-08II]. In the present paper, the max-min segments are introduced in Section \[s:segments\]. We recall the structure of max-min semispaces in Section \[s:semispaces\] together with some immediate consequences from abstract convexity. In [@N-Ser1; @N-Ser2] further progress is made in the study of max-min convexity focusing on the role of semispaces. Being motivated by the Hahn-Banach separation theorems in the tropical (max-plus) convexity [@Zim-77] and extensions to functional and abstract idempotent semimodules [@CGQS-05; @LMS-01; @Zim-81], we compared semispaces to max-min hyperplanes in [@N-Ser1], and developed an interval extension of separation by semispaces in [@N-Ser2]. These results are summarized in Section \[s:separation\]. Another principal goal of this paper is to investigate classical convexity results such as the theorems of Caratheódory, Helly and Radon in the realm of max-min convexity. These results are presented in Sections \[s:carath\], \[s:intsep\] and \[s:radon-helly\] and are inspired by a paper of Gaubert and Meunier [@G-Meu], in which similar statements can be found for the case of max-plus convexity. The max-min Carathéodory theorem with some “colorful” extensions is presented in Section \[s:carath\]. The strongest extension relies on what we call the internal separation theorem, which is proved in Section \[s:intsep\]. In the last section, motivated by the fuzzy algebra of [@pap], we consider a more general class of max-T semirings, where the role of multiplication is played by a triangular norm. We show how the topological Radon and Tverberg theorems can be applied to obtain, in particular, the max-min analogues of Radon, Helly, Centerpoint and (in part) Tverberg theorems.
Description of segments {#s:segments}
=======================
In this section we describe general segments in $\mmset^{d},$ following [@NS-08I; @Ser-03], where complete proofs can be found. Note that the description of the segments in [@NS-08I; @Ser-03] is done for the equivalent case where $\mmset=[-\infty, +\infty]$.
Let $x=(x_{1},...,x_{d}),$ $y=(y_{1},...,y_{d})\in\mmset^{d},$ and assume that we are in the *case of comparable endpoints*, say $x\leq y$ in the natural order of $\mmset^{d}.$ Sorting the set of all coordinates $\{x_{i},y_{i},i=1,...,d\}$ we obtain a non-decreasing sequence, denoted by $t_1,t_2,\ldots, t_{2d}$. This sequence divides the set $\mmset$ into $2d+1$ subintervals $\sigma_0=[0,t_{1}],\,\sigma_1=[t_1,t_2],...,\sigma_{2d}=[t_{2d},1]$, with consecutive subintervals having one common endpoint.
Every point $z\in [x,y]_{\oplus}$ is represented as $z=\alpha\otimes x\oplus\beta\otimes y$, where $\alpha=1$ or $\beta=1$. However, case $\beta=1$ yields only $z=y$, so we can assume $\alpha=1$. Thus $z$ can be regarded as a function of one parameter $\beta$, that is, $z(\beta)=(z_{1}(\beta),...,z_{d}(\beta))$ with $\beta\in\mmset$. Observe that for $\beta\in\sigma_0$ we have $z(\beta)=x$ and for $\beta\in\sigma_{2d}$ we have $z(\beta)=y$. Vectors $z(\beta)$ with $\beta$ in any other subinterval form a conventional [*elementary segment*]{}. Let us proceed with a formal account of all this.
\[tcomp\] Let $x,y\in\mmset^d$ and $x\leq y$.
- We have $$\label{e:chain}
[x,y]_{\oplus}=\bigcup_{l=1}^{2d-1} \{z(\beta)\mid\beta\in\sigma_l\},$$ where $z(\beta)=x\oplus(\beta\otimes y)$ and $\sigma_{\ell}=[t_l,t_{l+1}]$ for $\ell=1,\ldots,2d-1$, and $t_1,\ldots,t_{2d}$ is the nondecreasing sequence whose elements are the coordinates $x_i,y_i$ for $i=1,\ldots,d$.
- For each $\beta\in\mmset$ and $i$, let $M(\beta)=\{i\colon x_i\leq\beta\leq y_i\}$, $H(\beta)=\{i\mid\beta\geq y_i\}$ and $L(\beta)=\{i\colon\beta\leq x_i\}$. Then $$\label{zibeta}
z_i(\beta)=
\begin{cases}
\beta, &\text{if $i\in M(\beta)$},\\
x_i, &\text{if $i\in L(\beta)$},\\
y_i, &\text{if $i\in H(\beta)$},
\end{cases}$$ and $M(\beta),L(\beta), H(\beta)$ do not change in the interior of each interval $\sigma_{\ell}$.
- The sets $\{z(\beta)\mid\beta\in\sigma_{\ell}\}$ in are conventional closed segments in $\mmset^d$ (possibly reduced to a point), described by where $\beta\in\sigma_{\ell}$.
For *incomparable endpoints* $x\not\leq y,\,y\not\leq x,$ the description can be reduced to that of segments with comparable endpoints, by means of the following observation.
\[tincomp\]Let $x,y\in
\mmset^d$. Then $[x,y]_{\oplus}$ is the concatenation of two segments with comparable endpoints, namely $\lbrack x,y]_{\oplus}=[x,x\oplus y]_{\oplus}\cup [x\oplus y,y]_{\oplus}.$
All types of segments for $d=2$ are shown in the right side of Figure 1.
The left side of Figure 1 shows a diagram, where for $x=(x_1,x_2,x_3)$ and $y=(y_1, y_2, y_3)$, the segments $[x_1,y_1],
[x_2,y_2],$ and $[x_3,y_3]$ are placed over one another, and their arrangement induces a tiling of the horizontal axis, which shows the possible values of the parameter $\beta$. The partition of the real line induced by this tiling is associated with the intervals $\sigma_l$, and the sets of [*active indices*]{} $i$ with $z_i(\beta)=\beta$ associated with each $\sigma_l$ are also shown.
at (4,3) [Segments in $\mmset^2$, comparable endpoints]{}; (0,0)–(2,0)–(2,2)–(0,2)–(0,0)–(2,2); (.1, .6)–(.45,.6)–(.45,1.7); (3,0)–(5,0)–(5,2)–(3,2)–(3,0)–(5,2); (3.1,.6)–(3.6,.6)–(4.5,1.5)–(4.5,1.9); (6,0)–(8,0)–(8,2)–(6,2)–(6,0)–(8,2); (6.1,.6)–(6.6,.6)–(7.5,1.5)–(7.9,1.5); (0,-3)–(2,-3)–(2,-1)–(0,-1)–(0,-3)–(2,-1); (.5,-2.9)–(.5,-2.6)–(1.5,-2.6); (3,-3)–(5,-3)–(5,-1)–(3,-1)–(3,-3)–(5,-1); (3.5,-2.9)–(3.5,-2.5)–(4.5,-1.5)–(4.5,-1.1); (6,-3)–(8,-3)–(8,-1)–(6,-1)–(6,-3)–(8,-1); (6.5,-2.9)–(6.5,-2.5)–(7.5,-1.5)–(7.8,-1.5); at (4,-4) [Segment in $\mmset^2$, incomparable endpoints]{}; (3,-7)–(5,-7)–(5,-5)–(3,-5)–(3,-7)–(5,-5); (3.4,-5.5)–(4.6,-5.5)–(4.6,-6.7);
(-10,-6)–(-2,-6); (-9,-2)–(-5,-2); (-8,-1)–(-3,-1); (-7,0)–(-4,0); (-9,-6)–(-9,1); (-8,-6)–(-8,1); (-7,-6)–(-7,1); (-5,-6)–(-5,1); (-4,-6)–(-4,1); (-3,-6)–(-3,1); at (-9,-6.5) [$t_1$]{}; at (-8,-6.5) [$t_2$]{}; at (-7,-6.5) [$t_3$]{}; at (-5,-6.5) [$t_4$]{}; at (-4,-6.5) [$t_5$]{}; at (-3,-6.5) [$t_6$]{}; at (-1.8,-6.5) [$\beta$]{};
at (-9.5,-5.5) ; at (-8.5,-5.5) ; at (-7.5,-5.5) ; at (-6,-5.5) ; at (-4.5,-5.5) ; at (-3.5,-5.5) ; at (-2.5,-5.5) ; at (-1.5,-5.5) ;
at (-9.5,-4.5) [$\sigma_0$]{}; at (-8.5,-4.5) [$\sigma_1$]{}; at (-7.5,-4.5) [$\sigma_2$]{}; at (-6,-4.5) [$\sigma_3$]{}; at (-4.5,-4.5) [$\sigma_4$]{}; at (-3.5,-4.5) [$\sigma_5$]{}; at (-2.5,-4.5) [$\sigma_6$]{};
at (-9.5,-2) [$x_1$]{}; at (-4.5,-2) [$y_1$]{}; at (-8.5,-1) [$x_2$]{}; at (-2.5,-1) [$y_2$]{}; at (-7.5,0) [$x_3$]{}; at (-3.5,0) [$y_3$]{};
at (-5.5,3) [Diagram showing intervals $\sigma_{\ell}$ and sets]{}; at (-5.5,2.3) [of coordinates moving together $M(\beta)$]{};
\[rsamerole\] We observe that, similarly to the max-plus case (see [@NS-1], Remark 4.3) in $\mmset^d$ there are elementary segments in only $2^d-1$ directions. Elementary segments are the “building blocks” for the max-min segments in $\mmset^d,$ in the sense that every segment $[x,y]\subset \mmset^d$ is the concatenation of a finite number of elementary subsegments (at most) $2d-1$, respectively $2d-2$, in the case of comparable, respectively incomparable, endpoints.
Max-min segments allow to introduce a natural metric on $\mmset^d$ ([@EJN]). More precisely, one defines the distance between two points to be the Euclidean length of the max-min segment joining them.
Description of semispaces {#s:semispaces}
=========================
For any point $x^0=(x_{1}^{0}, \dots, x_{d}^{0})\in \mmset^d$ we define a finite family of subsets $S_0(x^0),\dots,S_d(x^0)$ in $\mmset^d$. These subsets were shown to be semispaces in [@NS-08II Proposition 4.1]. A point $x^0$ is called [*finite*]{} if it has all coordinates different from zeros and ones. This definition is motivated by the isomorphic version of max-min algebra where the least element (and zero of the semiring) is $-\infty$, and the greatest element (and unity of the semiring) is $+\infty$.
Without loss of generality we may assume that $x^0$ is [**non-increasing**]{}: $ x_{1}^{0}\geq \dots \geq x_{d}^{0}. $ Writing this more precisely we have $$\label{permut5}
\begin{gathered}
x_{1}^{0}
=\dots =x_{k_{1}}^{0}>\dots >x_{k_{1}+l_{1}+1}^{0}=\dots =x_{k_{1}+l_{1}+k_{2}}^{0}>\dots \\
>x_{k_{1}+l_{1}+
k_{2}+l_{2}+1}^{0}=\dots =x_{k_{1}+l_{1}+k_{2}+l_{2}+k_{3}}^{0}>\dots\\
>x_{k_{1}+l_{1}+\dots +k_{p-1}+l_{p-1}+1}^{0}=\dots =x_{k_{1}+l_{1}+\dots +k_{p-1}+l_{p-1}+k_{p}}^{0}\\
>\dots >x^0_{k_{1}+l_{1}+\dots +k_{p}+l_{p}}(=x^0_d),
\end{gathered}$$ where $\sum_{j=1}^p(k_j+l_j)=d$, $k_1=0$ if the sequence starts with strict inequalities and $l_p=0$ if the sequence ends with equalities.
Let us introduce the following notations: $$\begin{split}
L_{0} &= 0,K_{1}=k_{1},L_{1}=K_{1}+l_{1}=k_{1}+l_{1},\\
K_{j} &= L_{j-1}+k_{j}=k_{1}+l_{1}+...+k_{j-1}+l_{j-1}+k_{j}\quad
(j=2,...,p), \\
L_{j} &= K_{j}+l_{j}=k_{1}+l_{1}+...+k_{j}+l_{j}\quad (j=2,...,p);
\end{split}$$ we observe that $l_{j}=0$ if and only if $K_{j}=L_{j}.$
We are ready to define the subsets. We need to distinguish the cases when the sequence ends with zeros or begin with ones, since some subsets $S_i$ become empty in that case.
\[def:semi\] Let $x^0\in\mmset^d$ be a non-increasing vector\
a) If $x^{0}$ has $0<x_i^0<1$ for all $1\leq i\leq d$, then: $$\begin{split}
S_{0}(x^0)=&\{x\in \mmset^{d}|x_{i}>x_{i}^{0}\text{ for some }1\leq i\leq
d\}, \\
S_{K_{j}+q}(x^0)=&\{x\in \mmset^{d}|x_{K_{j}+q}<x_{K_{j}+q}^{0},\text{ or }%
x_{i}>x_{i}^{0}\\
& \text{ for some }K_{j}+q+1\leq i\leq d\} (q=1,...,l_{j};j=1,...,p \text{ if }l_{j}\neq 0),\\
S_{L_{j-1}+q}(x^0)=&\{x\in \mmset^{d}|x_{L_{j-1}+q}<x_{L_{j-1}+q}^{0},\text{
or }x_{i}>x_{i}^{0}\\
& \text{ for some }K_{j}+1\leq i\leq d\} \\
& (q=1,...,k_{j};j=1,...,p\text{ if }k_{1}\neq 0,\text{ or }j=2,...,p\text{ if
}k_{1}=0).
\end{split}$$
b\) If there exists an index $i\in \{1,...,d\}$ such that $x_{i}^{0}=1,$ but no index $j$ such that $x_{j}^{0}=0,$ then the subsets are $S_{1},...,S_{d}$ of part a).
c\) If there exists an index $j\in \{1,...,d\}$ such that $x_{j}^{0}=0,$ but no index $i$ such that $x_{i}^{0}=1,$ then the subsets are $S_{0},S_{1},...,S_{\beta -1}$ of part *a)*, where $
\beta :=\min \{1\leq j\leq n|\;x_{j}^{0}=0 \}.
$
d\) If there exist an index $i\in \{1,...,d\}$ such that $x_{i}^{0}=1,$ and an index $j$ such that $x_{j}^{0}=0,$ then the subsets are $S_{1},...,S_{\beta -1}$.
Let now $x^0\in\mmset^d$ have arbitrary order of coordinates, and let us formally extend Definition \[def:semi\]. For this, consider a permutation $\pi$ of the index set $\{1,\ldots,d\}$ such that the vector $(x_{\pi(1)}, x_{\pi(2)},\ldots,x_{\pi(d)})$ is non-increasing. Let $\overline{\pi}:\mmset^d\to\mmset^d$ be the invertible map of $\mmset^d$ induced by the permutation $\pi$. Then we can define $S_i(x^0)=\overline{\pi}^{-1} (S_j(\overline{\pi}(x^0)))$, where $j=\pi(i)$.
Further, for any $x^0\in \mmset^d$ we denote by $I(x^0)$ the set of indices $i$ such that $S_{\pi(i)} (\overline{\pi}(x^0))$ is present in Definition \[def:semi\]. Observe that $I(x^0)$ consists of the components $i$ such that $x^0_i>0$ and, possibly, $0$.
Pictures of all semispaces at a finite point for $d=2$ are shown in Figure 2.
at (4,2.5) [Semispaces at a point with equal coordinates]{};
(.7,0)–(2,0)–(2,2)–(0,2)–(0,.7)–(.7,.7)–(.7,0); (.7,0)–(2,0)–(2,2)–(0,2)–(0,.7); (0,.7)–(0,0)–(.7,0)–(.7,.7)–(0,.7); (.7,.7) circle (.05);
(3.7,0)–(3,0)–(3,2)–(3.7,2)–(3.7,0); (3.7,0)–(3,0)–(3,2)–(3.7,2); (3.7,0)–(5,0)–(5,2)–(3.7,2)–(3.7,0); (3.7,.7) circle (.05);
(6,.7)–(6,0)–(8,0)–(8,.7)–(6,.7); (6,.7)–(6,0)–(8,0)–(8,.7); (6,.7)–(6,2)–(8,2)–(8,.7)–(6,.7); (6.7,.7) circle (.05);
at (4,-.5) [Semispaces at a point with unequal coordinates]{};
(1.2,-3)–(1.2,-2.5)–(0,-2.5)–(0,-1)–(2,-1)–(2,-3)–(1.2,-3); (1.2,-3)–(2,-3)–(2,-1)–(0,-1)–(0,-2.5); (0,-3)–(1.2,-3)–(1.2,-2.5)–(0,-2.5)–(0,-3); (1.2,-2.5) circle (.05);
(4.2,-3)–(4.2,-2.5)–(5,-2.5)–(5,-1)–(3,-1)–(3,-3)–(4.2,-3); (3,-3)–(4.2,-3); (5,-2.5)–(5,-1)–(3,-1)–(3,-3); (4.2,-3)–(5,-3)–(5,-2.5)–(4.2,-2.5)–(4.2,-3); (4.2,-2.5) circle (.05);
(6,-2.5)–(6,-3)–(8,-3)–(8,-2.5)–(6,-2.5); (6,-2.5)–(6,-3)–(8,-3)–(8,-2.5); (6,-2.5)–(8,-2.5)–(8,-1)–(6,-1)–(6,-2.5); (7.2,-2.5) circle (.05);
The following theorem is the main result in [@NS-08II]. See also [@N-Ser2].
\[p:semisp-conv\] For any $p\in \mmset^d$ the sets $S_i(p), i\in I(p),$ are maximal max-min convex avoiding the point $p$. Thus for any $p\in \mmset^d$, there exists at least one and at most $d+1$ semispaces $S_i(p), 0\le i\le d,$ at $p$.
For all $C\subseteq \mmset^d$ max-min convex and any $p\in \mmset^d
\setminus C$, there exists a semispace $S_i(p)$ such that $C\subseteq S_i(p)$ and $p\not \in S_i(p)$.
The complement of a semispace $S_i(p)$ is denoted by $\complement
S_i(p)$. These complements are also called [*sectors*]{}, in analogy with the max-plus convexity.
The lemma below follows from the abstract definition of the semispaces and it is our main tool in extending Caratheódory theorem and its colorful versions to the max-min setup. As only a finite number of semispaces at a given point exist, the max-min convexity can be regarded as a multiorder convexity [@NS-08I; @NS-08II].
\[basic-tool\] Let $X\subseteq \mmset^d$ and $p\in \mmset^d$. Then the following statements are equivalent:
1. $p\in \conv(X)$;
2. for all $i\in I(p),$ there exists $x^i\in X$ such that $x^i\in \complement S_i(p)$.
\(i) $\rightarrow$ (ii) By contradiction. Assume there is $i_0\in I(p)$ such that $X\cap \complement S_{i_0}(p)=\emptyset$. Then $p\in\conv(X)\subseteq S_{i_0}(p)$, in contradiction to $p\not \in S_{i_0}(p)$.
\(ii) $\rightarrow$ (i) By contradiction. Assume that $p\not \in
\conv(X)$. As $\conv(X)$ is a convex set, it follows from Theorem \[p:semisp-conv\] that there exists $i_0\in I(p)$ such that $\conv(X)\subseteq S_{i_0}(p)$, which implies $\complement
S_{i_0}(p)\subseteq \complement \conv(X)$. But from (ii), there exists $x_{i^0}\in \complement S_{i_0}(p)\cap \conv(X)$, which gives a contradiction.
Separation and non-separation {#s:separation}
=============================
In what follows $\mmset^d$ has the usual Euclidean topology. If $A\subseteq \mmset^d$, we denote by $\overline{A}$ the closure of $A$, by $\text{int}(A)$ the interior of $A$ and by $\complement A$ the complement of $A$.
In the tropical convexity, all semispaces are open tropical halfspaces expressed as solution sets to a strict two-sided max-linear inequality. See e.g. [@NS-1]. Thus the closures of semispaces are hyperplanes.
In the case of max-min convexity, hyperplane in $\mmset^d$ can be defined as the solution set to a max-min linear equation [$$\label{e:maxmin-hyper}
\max(\min (a_1,x_1),\ldots,\min(a_d,x_d),a_{d+1})=
\max(\min (b_1,x_1),\ldots,\min(b_d,x_d),b_{d+1}).$$ ]{}
The structure of a max-min hyperplane is presented in [@Nit-09]. One investigates the distribution of values for the left and right hand side of , and then identifies the regions in $\mmset^d$ where the values of the sides coincide. We illustrate this procedure in Figure 3, which shows the structure of a max-min hyperplane (line) in $\mmset^2$. The left side pictures show the distribution of values for both sides of : for the white regions the distribution is uniform and the value is equal to the coordinate of the finite point on the main diagonal that belongs to their boundary; the regions labeled $x_1$ are tiled by vertical lines each of value equal to its $x_1$ coordinate, and the regions labeled $x_2$ are tiled by horizontal lines each of value equal to its $x_2$ coordinate. The right side picture shows the graph of the line.
 \[figure1\]
In [@N-Ser1] we investigated the relation between the max-min hyperplanes and the closures of semispaces ${S}_i(x)$. We recall that the *diagonal* of $\mmset^d$ is the set $\cD_d=\{(a,\ldots,a)\in \mmset^d\mid a\in\mmset\}.$
\[t:NitSer1\] A closure of semispace is a hyperplane if and only if it can be represented as $\overline{S}_i(y)$ for some $y$ belonging to the diagonal.
Theorem \[t:NitSer1\] shows exactly when classical separation by hyperplanes is possible.
Let $x\in\mmset^d$, then any closed max-min convex set $C\subseteq\cB^d$ not containing $x$ can be separated from $x$ by a hyperplane if and only if $x$ lies on the diagonal.
In [@N-Ser2], we found a way to enhance separation by semispaces showing that a point can be replaced by a box, i.e., a Cartesian product of closed intervals. Namely, we investigated the separation of a box $B=[\podx_1,\nadx_1]\times\ldots\times[\podx_d,\nadx_d]\subseteq \mmset^d$ from a max-min convex set $C\subseteq\mmset^d$, by which we mean that there exists a set $S$ described in Definition \[def:semi\], which contains $C$ and avoids $B$.
Assume that $\nadx_1\geq\ldots\geq\nadx_d$ and suppose that $t(B)$ is the greatest integer such that $\nadx_{t(B)}\geq \podx_i$ for all $1\leq i\leq t(B)$. We will need the following condition: $$\label{sep-cond}
\begin{split}
&\text{ If }(\nadx_1=1 )\ \&\ (y_l\geq\podx_l,1\le l\le d)\ \&\\
&(\nadx_l<y_l\ \text{for some $l\leq t(B)$}),\ \text{ then } y\notin C.
\end{split}$$ Note that if the box is reduced to a point and if $\nadx_1=1$, then $\nadx_l=1$ for all $l\leq t(B)$ so that $\nadx_l<y_l$ is impossible. So always holds in the case of a point.
\[interval-sep\] Let $B=[\podx_1,\nadx_1]\times\ldots\times[\podx_d,\nadx_d]\subseteq \mmset^d$, and let $C\subseteq\mmset^d$ be a max-min convex set avoiding $B$. Suppose that $B$ and $C$ satisfy . Then there is a semispace that contains $C$ and avoids $B$.
The box $B$ can be a point and in this case condition always holds. Therefore, some results on max-min semispaces [@NS-08II] can be deduced from Theorem \[interval-sep\]. The following is an immediate corollary of Theorem \[interval-sep\] and Proposition \[p:semisp-conv\].
\[ns-08ii\] Let $x\in\mmset^d$ be non-increasing and $C\subseteq\mmset^d$ be a max-min convex set avoiding $x$. Then $C$ is contained in one $S_i(x), i\in I(p),$ as in Definition \[def:semi\]. Consequently these sets are indeed the family of semispaces at $x$.
However, separation by semispaces is impossible when $B, C$ do not satisfy .
\[BC:nonsep\] Suppose that $B=[\podx_1,\nadx_1]\times\ldots\times[\podx_d,\nadx_d]\subseteq \mmset^d$ and the max-min convex set $C\subseteq\mmset^d$ are such that $B\cap C=\emptyset$ but the condition does not hold. Then there is no semispace that contains $C$ and avoids $B$.
In [@N-Ser2] we also investigate the separation of max-min convex sets by a box, and by a box and a semispace. We show that both kinds of separation are always possible if $n=2$, but they are not valid in higher dimensions.
Caratheódory theorems {#s:carath}
=====================
In this section we investigate classical convexity results in max-min setup.
Consider $X=\{x^1,x^2,\dots,x^m\}\subseteq \mmset^d,$ $m\ge d+1.$ Assume that $p\in \conv(X)$. Then there exists $X'=\{x'^i\vert i\in I\}\subseteq X,$ $1\le \vert I \vert \le d+1,$ such that $p\in \conv(X').$
By Lemma \[basic-tool\], implication (i) $\rightarrow$ (ii), $p\in \conv(X)$ shows that for any $i\in I(p)$ there exists $x'^i\in X\cap \complement S_i(p)$. Define $X'=\{x'^i\vert i\in I(p)\}\subseteq X$. Then again by Lemma \[basic-tool\], now implication (ii) $\rightarrow$ (i), it follows that $p\in \conv(X').$
\[cara-color-weak\] Let $X^0, X^1, \dots,$ $X^{d}$ be subsets in $\mmset^d$ and $p\in \mmset^d$. Assume that $p\in \conv(X^i)$ for all $0\le i\le d$. Then, up to a permutation of indices, there exist $x^i\in X^i,i\in I(p),$ such that $p\in \conv(\{x^i\vert i\in I(p)\}).$
From Lemma \[basic-tool\], implication (i) $\rightarrow$ (ii), it follows that there exist $x^i_j\in X^i, 1\le i \le d+1, j\in I(p),$ such that $x^i_j\in \complement S_j(p),j\in I(p)$. Then again from Lemma \[basic-tool\], implication (ii) $\rightarrow$ (i), and from $x^i:=x_i^i\in \complement S_i(p), i\in I(p),$ it follows that $p\in \conv(\{x^i\vert i\in I(p)\}).$
\[l:sepincl\] Let $p,q\in \mmset^d$. Then for all $i\in I(q)$ there exists $j\in I(p)$ such that $\complement S_j(p)\subseteq \complement S_i(q)$.
The statement is equivalent to $S_i(q)\subseteq S_j(p)$. This follows from the fact that the convex set $S_i(q)$ has to be included in a semispace at $p$.
We now explain the concept of internal separation property, in the max-min setting. The proof of internal separation property is deferred to the next section.
Given $X=\{x^0,\ldots,x^{d}\}\subseteq\mmset^d$, we say that a finite point $p\in \conv(X)$ internally separates $x_0,\ldots,x_p$ if up to a permutation, each semispace $S_i(p), 0\le i \le d,$ corresponds to $x^i\in\complement S_i(p)$.
\[t:intsep\] For any subset $X=\{x^0,\ldots,x^{d}\}\subseteq\mmset^d$, consisting of finite points, $\conv(X)$ contains a point $p$ with internal separation property.
We will need yet another simple observation, to obtain the colorful Carathéodory theorem in most general form. Let $\mmseto$ be a closed interval on the real line strictly containing $\mmset=[0,1]$, and denote by $\underline{0}$, resp. $\overline{1}$ the least, resp. the greatest element of $\mmseto$. We have $\underline{0}<0<1<\overline{1}$, and we can define the max-min semiring over $\mmseto$ with zero $\underline{0}$ and unity $\overline{1}$. For $X\subseteq\mmset^d$, denote by $\convo(X)$ the max-min convex hull of $X$ in $\mmseto^d$.
\[l:extend\] For any $X\subseteq\mmset^d$, we have $\convo(X)=\conv(X)$.
The “new” convex hull $\convo(X)$ is the set of combinations $$\label{convoX}
\bigoplus_{i=1}^m \lambda_i\otimes x^i\colon m\geq 1,\ \lambda_i\in\mmseto,\ \bigoplus_{i=1}^m \lambda_i=\overline{1},$$ taken for all $m$-tuples of points $x^i$ from $X$.
To obtain $\conv(X)\subseteq\convo(X)$, observe that when $\lambda_i=1$ in is changed to $\lambda_i=\overline{1}$ the “product” $\lambda_i\otimes x^i$ is unaffected (since all components of $x^i$ are $\leq 1$). To show $\convo(X)\subseteq\conv(X)$, use the same observation to change $\lambda_i=\overline{1}$ to $\lambda_i=1$ in . Next, no combination (now with $1$ instead of $\overline{1}$) has any negative components since all $x^i$ are nonnegative and there is a point with coefficient $\lambda_i=1$. Hence all $\lambda_i\colon \underline{0}\leq\lambda_i<0$ can be changed to $0$ without affecting . This completes the proof.
\[c:extend\] A max-min convex set $C\subseteq\mmset^d$ remains max-min convex in $\mmseto^d$.
\[t:cara-color-strong\] Let $X^0, X^1, \dots,$ $X^{d}\subseteq\mmset^d$, and $C\subseteq
\mmset^d$ be a max-min convex set. Assume that $C\cap \conv(X^i)\not
=\emptyset$ for all $0\le i\le d$. Then there exist $x^i\in X^i, 0\le i\le d,$ such that $C\cap \conv(\{x^0,x^1,\dots,x^{d}\})\not =\emptyset.$
Assume first that all points in $X^0,X^1,\ldots, X^d$ are finite. Take $p^i\in C\cap \conv(X^i), 0\le i\le d.$ By Theorem \[t:intsep\] we can select a point $q$ which separates $p^0,p^1,\dots,p^d$ internally, thus $p^i\in\complement S_i(q)$ for all $i$. As $p^i\in C, 0\le i\le d,$ by Lemma \[basic-tool\] one has also $q\in C$. It remains to show that $q\in \conv(\{x^0,x^1,
\dots, x^{d}\})$, with some $x^i\in X^i, 0\le i\le d.$
By Lemma \[l:sepincl\], for any $0\le i\le d$, there exists $0\le
j\le d$ such that $\complement S_j(p^i)\subseteq \complement
S_i(q)$. As $p^i\in \conv(X^i)$, by Lemma \[basic-tool\] there exists $x^i\in X_i\cap \complement S_j(p_i)$. Hence $x^i\in
\complement S_i(q)$. Hence again by Lemma \[basic-tool\] one has $q\in \conv(\{x^0,x^1, \dots, x^{d}\})$. This proves the claim under assumption that $X^0,X^1,\ldots, X^d$ have only finite points.
Without that assumption, regard $X^0, X^1, \dots,$ $X^{d},C\in\mmset^d$ as subsets of $\mmseto^d$ where $\mmseto$ is a closed interval strictly containing $\mmset$. By Corollary \[c:extend\], $C$ remains max-min convex in $\mmseto^d$, and by Lemma \[l:extend\] none of the convex hulls in the claim change when they are considered in $\mmseto^d$. This extension makes all points in $X^0,X^1,\ldots, X^d$ finite, and the previous argument works in $\mmseto^d$ (with sectors in $\mmseto^d$).
We conclude the section with the proof of internal separation property in the cases when 1) $\conv(X)$ has a non-empty interior, 2) all vectors $p^{\ell}$ are non-increasing. These proofs can be skipped by the reader, who can proceed to a general proof of Theorem \[t:intsep\] written in the next section. Let us introduce the notion of interior of a max-min convex set.
Interior of a max-min convex set $C\in\mmset^d$, denoted by $\inter(C)$ is the subset of $C$ consisting of points $y$ such that there is an open $d$-dimensional box $(y_1-\epsilon,y_1+\epsilon)\times\cdots\times(y_d-\epsilon,y_d+\epsilon)$ for some $\epsilon>0$
\[p:inter\] Assume $X=\{x^0,x^1,\dots, x^{d}\}\subseteq \mmset^d$ generates a max-min polytope $S=\conv(X)$ with non-empty interior. Then for any point $p\in \inter(S)$ with all coordinates different, up to a permutation of indices, one has $x^i\in \complement S_i(p), i\in I(p)$.
We proceed by contradiction. As $p$ has all coordinates different and it is away from the boundary, the interiors of $\complement S_i(p), 0\le i\le d,$ are disjoint. If $p$ does not internally separate the points of $X$, then there exists $i\colon 0\le i\le d$ such that $\inter(\complement
S_i(p))\cap X=\emptyset$. However, as the complement $\complement(\inter(\complement S_i(p)))$ is the topological closure of $S_i(p)$, it is a max-min convex set, and hence $\conv(X)\cap
\inter(\complement S_i(p))=\emptyset$. But then $p$ is not in the interior of $\conv(X)$.
The notion of interior and, more generally, of dimension in max-min convexity will be investigated in another publication. We now treat the other special case.
\[propo-same-order\] Assume that $x^{\ell}\in \mmset^d, 0\le \ell\le d,$ are non-increasing, i.e., $$\label{eq-ineeq-34}
x^{\ell}_1\geq x^{\ell}_2\geq\ldots\geq x^{\ell}_d,\quad 0\le {\ell}\le d,$$ and finite. Then there exists $p\in\mmset^d$ such that $x^\ell\in\complement S_{\ell}(p)$ for all $\ell\in\{0,1,\ldots,
d\}$.
Let $y_d:=\max_{\ell=0}^d x_d^{\ell}$, and $\ell'_1$ be an index where this maximum is attained. Reordering the points, we can assume $\ell'_1=d$. Let $y_{d-1}:=\max_{\ell=0}^{d-1} x_{d-1}^{\ell}$ and $\ell'_2$ be an index where this maximum is attained. Reordering the points $x^0,\ldots, x^{d-1}$ we can assume $\ell'_2=d-1$. On a general step of this procedure, we have obtained the partial maxima $y_d, y_{d-1},\ldots, y_{d-t+1}$ equal to $x_d^{d}$, $x_{d-1}^{d-1},\ldots, x_{d-t+1}^{d-t+1}$ (having reorganized the given points $x$), and we define $y_{d-t}:=\max_{\ell=0}^{d-t}
x_{d-t}^{\ell}$, requiring that $y_{d-t}=x_{d-t}^{d-t}$. On the last step, we have $y_1=\max(x_1^0,x_1^1)$ and swap $x^0$ with $x^1$ (if necessary) to obtain $y_1=x_1^1$.
This process defines the vector $y=(y_1,\ldots,y_d)$ and rearranges the given points $x^{0},\ldots,x^{d}$ in such a way that $$y_t={\max}_{\ell=0}^t x_t^\ell=x_t^t,\quad\forall t\in\{1,\ldots,d\}.$$
Now define $p$ to be the largest non-increasing vector satisfying $p\leq y$. We will show that $p$ is a point that we need. Before the main argument we observe that $$\label{e:zprop1}
p_t\leq y_t=x_t^t\quad \forall t\in\{1,\ldots,d\},$$ and $$\label{e:zprop2}
\left\{
\begin{split}
p_1&=y_1,\\
p_t&=y_t={\max}_{\ell=0}^t x_t^\ell=x_t^t\quad \text{if $p_t<p_{t-1}$}.
\end{split}
\right .$$ Only has to be shown. Indeed, if $p_1<y_1$, then $(y_1,p_2,\ldots,p_d)$ is a non-increasing vector bounded by $y$ from above and contradicting the maximality of $z$, so $p_1=y_1$ holds. If $p_t<p_{t-1}$ and $p_t<y_t$ then defining $p'_t:=\min(p_{t-1},y_t)$ we have $p_{t-1}\geq p'_t\geq p_{t+1}$ and $p'_t\leq y_t$, so again, $(p_1,\ldots,p_{t-1},p'_t,p_{t+1},\ldots,p_d)$ is a non-increasing vector bounded by $y$ from above and contradicting the maximality of $p$.
For what follows, we refer the reader to Definition \[def:semi\], that describes the structure of the semispaces.
We now show that $x^\ell\in\complement S_\ell(p)$ for all $\ell\in\{0,1,\ldots,d\}$, starting with $\ell=0$. In this case we need to argue that $x_t^0\leq p_t$ for all $t$. Indeed, when $p_{t-1}>p_t$, the inequality $x_t^0\leq p_t$ follows from (second part). If $p_{t-1}=p_t$, then either $p_1=\ldots=p_t$, or $p_{t-i-1}>p_{t-i}=\ldots=p_{t-1}=p_t$. In the first case we have $x_s^0\leq p_s$ for $s=1$, and in the second case for $s=t-i$, and in both cases the required inequality $x_t^0\leq
p_t$ follows since $x^0$ is a non-increasing vector.
When $\ell>0$ and $p_{\ell-1}>p_\ell$, we have $x_\ell^\ell=p_\ell$ by , so $x_\ell^\ell\geq p_\ell$. When $p_{t-1}>p_t$, the inequalities $x_t^\ell\leq p_t$ for $t>\ell$ follow from , and when $p_{t-1}=p_t$, we have $p_{t-i-1}>p_{t-i}=\ldots=p_t$ for some $i$, where $t-i\geq \ell$. In this case $x_{t-i}^{\ell}\leq p_{t-i}$ follows from , and we use that $x^\ell$ is non-increasing to obtain $x_t^\ell\leq p_t=p_{t-i}$.
If $p_{\ell-1}=p_\ell$, then either $p_\ell=p_{\ell+1}=\ldots=p_d$, or there exists $i$ such that $p_\ell=\ldots=p_{\ell+i}>p_{\ell+i+1}$. In this case $x_\ell^\ell\geq p_\ell$ follows from , and the inequalities $x_t^\ell\leq p_t$ for $t>\ell+i$ are shown as in the previous case.
The proof is complete.
Internal separation property {#s:intsep}
============================
This section is devoted to the proof of Theorem \[t:intsep\] (the internal separation property). Let $u^{(i)}$, for $i=1,\ldots,d+1$ be the given points in $\mmset^d$, and let $A\in\mmset^{(d+1)\times d}$ be the matrix where these vectors are rows. For such a matrix, denote by $A^{(h)}$ the Boolean matrix with entries $$a_{ij}^{(h)}=
\begin{cases}
1, &\text{if $a_{ij}\geq h$},\\
0, &\text{if $a_{ij}<h$}.
\end{cases}$$ Following the literature on max-min algebra, we may call it the [*threshold matrix*]{} of level $h$. Let $t$ be the greatest $h$ for which $A^{(h)}$ contains a $d\times d$ submatrix with a nonzero permanent (in other words, a permutation with nonzero weight).
For every $h>t$, every $d\times d$ submatrix of $A^{(h)}$ has zero permanent. Take $h>t$ to be smaller than any entry of $A$ that is greater than $t$, and consider the bipartite graph corresponding to $A^{(h)}$[^2]. As $A^{(h)}$ has zero permanent, the size of maximal matching in that graph is less than $d$. By the König theorem, the size of maximal matching is equal to the size of the minimal vertex cover. In particular, there exists a subset of rows $M_2$ and a subset of columns $N_2$ with number of elements $m_2$ and $n_2$ respectively, such that $m_2+n_2<d$ and such that all $1$’s of $A^{(h)}$ are in these columns and rows. Let $M_1$, resp. $N_1$, be the complements of $M_2$, resp. $N_2$ in $\{1,\ldots,d+1\}$, resp. $\{1,\ldots,d\}$. Then all entries of the submatrix $A_{M_1N_1}^{(h)}$ are zero, and hence all entries of $A_{M_1N_1}$ are less than or equal to $t$, and we have $m_1+n_1>d+1$, where $m_1$, resp. $n_1$ are the number of elements in $M_1$, resp. $N_1$.
Thus $A$ contains an $m_1\times n_1$ submatrix $B^{\leq
t}:=A_{M_1N_1}$ where all entries do not exceed $t$ and we have $m_1+n_1>d+1$. At the same time, there is a row index $f$ which we call the [*free index*]{}, and a permutation $\pi\colon
\{1,\ldots,d+1\}\bez\{f\}\mapsto\{1,\ldots,d\}$ such that $a_{i\pi(i)}\geq t$ for all $i\neq f$. The pair $(B^{\leq t},\pi)$ will be called a [*(König) diagram*]{}. Denote the number of intersections of $\pi$ with $A_{M_1N_1}$ by $r$ and with $A_{M_2N_2}$ by $s$. Then we obtain, having $d$ as the sum of the number of intersections of $\pi$ with $A_{M_1N_1}$, $A_{M_1N_2}$, $A_{M_2N_1}$ and $A_{M_2N_2}$ that $$\label{nequalsto}
d=
\begin{cases}
r+(m_1-r-1)+(n_1-r)+s, &\text{if $f\in M_1$},\\
r+(m_1-r)+(n_1-r)+s, &\text{if $f\notin M_1$}.
\end{cases}$$ Eliminating $r$ from we obtain $$\label{lequalto}
r=
\begin{cases}
m_1+n_1-(d+1)+s,&\text{if $f\in M_1$},\\
m_1+n_1-d+s, &\text{if $f\notin M_1$}.
\end{cases}$$
We see that with $m_1$, $n_1$ and $d$ fixed, the number $r$ is minimal when $f\in M_1$ and $s=0$. Such diagrams will be called [*tight*]{}. See Figure 4 for an illustration of a tight diagram. The entries in $\pi$ are represented by \*. In general, the [*tightness*]{} of a diagram is defined as the non-positive integer $m_1+n_1-d-1-r$.
(0,3)–(5,3)–(5,9)–(0,9)–(0,3);
(0,0)–(8,0)–(8,9)–(0,9)–(0,0); (0,8)–(8,8); (0,3)–(8,3); (5,0)–(5,9);
at (2.5,6) [$B^{\leq t}$]{}; at (6.5,1.5) [VOID]{}; at (4.1,8.5) [FREE ROW]{}; at (-.8,6) [$M_1$]{}; at (-.8,2) [$M_2$]{}; at (2.5,9.6) [$N_1$]{}; at (6.5,9.6) [$N_2$]{};
at (.5,.5) [\*]{}; at (1.5,1.5) [\*]{}; at (2.5,2.5) [\*]{}; at (3.5,3.5) [\*]{}; at (4.5,4.5) [\*]{};
(5,5) – (8,5);
(8,5) – (8,9) node \[black,midway,xshift=0.8cm\] [$\Tilde{M}_1$]{};
at (5.5,5.5) [\*]{}; at (6.5,6.5) [\*]{}; at (7.5,7.5) [\*]{};
Let us indicate some sufficient conditions for $(B^{\leq t},\pi)$ to be tight (the proof is omitted).
\[l:simpletight\] The diagram $(B^{\leq t},\pi)$ is tight if $m_1+n_1=d+2$, $f\in M_1$ and $\pi$ intersects with $B^{\leq t}$ only once. In particular, if $B^{\leq t}$ is a column, then $(B^{\leq t},\pi)$ is tight.
Substituting $m_1+n_1=d+2$ and $r=1$ in the first line of we have $s=0$.
Our next aim is to show that there always exists at least one tight diagram, and let us start with a pair of auxiliary lemmas.
\[l:sinking\] Let $(B^{\leq t},\pi)$ be not tight, and let $(k_0,\pi(k_0))\in M_1\times
N_1$. Then we have one of the following alternatives:
- There exists a sequence $k_0,\ldots, k_l$ such that $(k_i,\pi(k_i))\in M_2\times N_1$ for $i=1,\ldots,l-1$, $(k_l,\pi(k_l))\in M_2\times N_2$ or $k_l$ is free, and $a_{k_i\pi(k_{i-1})}>t$ for all $i=1,\ldots,l$;
- There is a tight diagram $(\Tilde{B}^{\leq t},\pi)$.
If we have $a_{i\pi(k_0)}\leq t$ for all $i$, then the entire column with index $\pi(k_0)$ can be taken for $\Tilde{B}^{\leq t}$, that is $M_1=\{1,\dots,d+1\}, N_1=\pi(k_0)$ and the diagram $(\Tilde{B}^{\leq t},\pi)$ is tight (by Lemma \[l:simpletight\]). If this is not the case, select $k'_1\in
M_2$ with $a_{k'_1\pi(k_0)}>t$. Then we proceed as in the following general description (with the sequence $k_0, k'_1$).
In general, suppose that we have found a sequence of rows $k_0,
k'_1,\ldots, k'_l$ where $k_0\in M_1$, $k'_1,\ldots, k'_l\in M_2$ and $\pi(k_0),\pi(k'_1),\ldots,\pi(k'_{l-1})\in N_1$ with the following property:
(\*) For each $s\colon 1\leq s\leq l$ there is a subsequence $k_0,k_1\ldots k_r$ of $k_0,k'_1,\ldots, k'_s$ such that $k_r=k'_s$ and $a_{k_i\pi(k_{i-1})}>t$ for all $i=1,\ldots,r$.
If $\pi(k'_l)$ is in $N_2$ or $k'_l$ is free then we are done. Otherwise consider the submatrix extracted from the columns $\pi(k_0),\pi(k'_1),\ldots,\pi(k'_l)$ and all rows except for $k'_1,\ldots, k'_l$. If this submatrix does not contain any entries greater than $t$ then it can be taken for $\Tilde{B}^{\leq t}$ and the diagram $(\Tilde{B}^{\leq t},\pi)$ is tight by Lemma \[l:simpletight\]. Otherwise we choose $k'_{l+1}\notin\{k_0,k'_1,\ldots,k'_l\}$ in $M_2$ in such a way that $a_{k'_{l+1}\pi(i)}>t$ for some $i$ in $\{k_0,k'_1,\ldots,k'_l\}$. Then $k_0,k'_1,\ldots, k'_l,k'_{l+1}$ satisfies the property (\*), and the process is continued until the intersection of $\pi$ with $M_2\times N_1$ is exhausted and we end up either with a free $k_l$, or such that $(k_l,\pi(k_l))\in M_2\times N_2$.
(0,5)–(5,5)–(5,9)–(0,9)–(0,5);
(0,0)–(8,0)–(8,9)–(0,9)–(0,0); (0,1)–(8,1); (0,5)–(8,5); (5,0)–(5,9);
(5.5,1.7)–(5.5,5.3); at (5.5,5.5)[$\bullet$]{};
(4.5,2.4)–(4.5,1.7); at (4.5,1.5) [$\bullet$]{};
(2.5,4.4)–(2.5,2.7); at (2.5,2.5) [$\bullet$]{};
(1.5,5.4)–(1.5,4.7); at (1.5,4.5) [$\bullet$]{};
at (2.5,7) [$B^{\leq t}$]{}; at (4.1,.5) [FREE ROW]{}; at (-.8,7) [$M_1$]{}; at (-.8,3) [$M_2$]{}; at (2.5,9.6) [$N_1$]{}; at (6.5,9.6) [$N_2$]{};
at (.5,6.5) [\*]{}; at (1.5,5.5) [\*]{}; at (2.5,4.5) [\*]{}; at (3.5,3.5) [\*]{}; at (4.5,2.5) [\*]{}; at (5.5,1.5) [\*]{};
(10,5)–(15,5)–(15,9)–(10,9)–(10,5);
(10,0)–(18,0)–(18,9)–(10,9)–(10,0); (10,1)–(18,1); (10,5)–(18,5); (15,0)–(15,9);
(10.5,6.4)–(10.5,3.7); at (10.5,3.5) [$\bullet$]{};
(13.5,3.6)–(13.5,.7); at (13.5,.5) [$\bullet$]{};
at (12.5,7) [$B^{\leq t}$]{}; at (13.7,.5) [FREE ROW]{}; at (9.4,7) [$M_1$]{}; at (9.4,3) [$M_2$]{}; at (12.5,9.5) [$N_1$]{}; at (16.5,9.5) [$N_2$]{};
at (10.5,6.5) [\*]{}; at (11.5,5.5) [\*]{}; at (12.5,4.5) [\*]{}; at (13.5,3.5) [\*]{}; at (14.5,2.5) [\*]{}; at (15.5,1.5) [\*]{};
(0,5)–(5,5)–(5,9)–(0,9)–(0,5);
(0,0)–(8,0)–(8,9)–(0,9)–(0,0); (0,1)–(8,1); (0,5)–(8,5); (5,0)–(5,9);
(1,9)–(1,5)–(4,5)–(4,9);
(1,5)–(4,8); (1,6)–(4,9); (1,7)–(3,9); (1,8)–(2,9); (2,5)–(4,7); (3,5)–(4,6);
(1,0)–(1,3)–(4,3)–(4,0); (1,0)–(4,3); (2,0)–(4,2); (3,0)–(4,1); (1,1)–(3,3); (1,2)–(2,3);
at (4.5,7) [$B^{\leq t}$]{}; at (2.5,7) [${\tilde B}^{\leq
t}$]{}; at (4.1,.5) [FREE ROW]{}; at (-.8,7) [$M_1$]{}; at (-.8,3) [$M_2$]{}; at (2.5,9.6) [$N_1$]{}; at (6.5,9.6) [$N_2$]{};
at (.5,6.5) [\*]{}; at (1.5,5.5) [\*]{}; at (2.5,4.5) [\*]{}; at (3.5,3.5) [\*]{}; at (4.5,2.5) [\*]{}; at (5.5,1.5) [\*]{};
Now we consider a reverse process.
\[l:lifting\] Let $(B^{\leq t},\pi)$ be not tight, and let $(k_0,\pi(k_0))\in M_2\times N_2$. Then we have one of the following alternatives:
- There exists a sequence $k_0,\ldots, k_l$ such that $(k_i,\pi(k_i))\in M_1\times N_2$ for $i=1,\ldots,l-1$, $(k_l,\pi(k_l))\in M_2\times N_2$ or $k_l$ is free, and $a_{k_i\pi(k_{i-1})}>t$ for all $i=1,\ldots,l$;
- There is a tighter diagram $(\Tilde{B}^{\leq t},\pi)$.
If we have $a_{i\pi(k_0)}\leq t$ for all $i$, then the column index $\pi(k_0)$ can be added to $M_1$ and the resulting diagram $(\Tilde{B}^{\leq t},\pi)$ is tighter (i.e., has a greater tightness) than $(B^{\leq t},\pi)$, since the size of $\Tilde{B}^{\leq t}$ increased while the number of intersections with $\pi$ is the same. Otherwise we can select $k'_1\in M_1$ with $a_{k'_1\pi(k_0)}>t$ and proceed as in the following general description (with the sequence $k_0, k'_1$).
In general, suppose that we have found a sequence of rows $k_0,
k'_1,\ldots, k'_l$ where $k_0\in M_2$, $k'_1,\ldots, k'_l\in M_1$ and $\pi(k_0),\pi(k'_1),\ldots,\pi(k'_l)\in N_2$ with the property (\*) in the proof of Lemma \[l:sinking\].
If $\pi(k'_l)$ is in $N_1$ or is free then we are done. Otherwise consider the submatrix extracted from the columns of $N_1$ and $\pi(k_0),\pi(k'_1),\ldots,\pi(k'_l)$, and all rows of $M_1$ except for $k'_1,\ldots, k'_l$. If this submatrix does not contain any entries greater than $t$ then it can be taken for $\Tilde{B}^{\leq
t}$ and the diagram $(\Tilde{B}^{\leq t},\pi)$ is tighter than $(B^{\leq t},\pi)$ since the sum of dimensions increases by one but the number of intersections of $\pi$ with $\Tilde{B}^{\leq t}$ is the same. Otherwise we choose $k'_{l+1}\notin\{k_0,k'_1,\ldots,k'_l\}$ in $M_2$ in such a way that $a_{k'_{l+1}\pi(i)}>t$ for some $i$ in $\{k_0,k'_1,\ldots,k'_l\}$. Then the sequence $k_0,k'_1,\ldots, k'_l,k'_{l+1}$ satisfies the property (\*), and the process is continued until the intersection of $\pi$ with $M_1\times N_2$ is exhausted and we end up either with a free $k_l$, or such that $(k_l,\pi(k_l))\in M_1\times N_1$.
(0,4)–(4,4)–(4,9)–(0,9)–(0,5);
(0,0)–(8,0)–(8,9)–(0,9)–(0,0); (0,8)–(8,8); (0,4)–(8,4); (4,0)–(4,9);
at (2,7) [$B^{\leq t}$]{}; at (4,8.5) [FREE ROW]{}; at (-.8,7) [$M_1$]{}; at (-.8,3) [$M_2$]{}; at (2.5,9.6) [$N_1$]{}; at (6.5,9.6) [$N_2$]{};
at (1.5,4.5) [\*]{};
(1.5, 4.3)–(1.5,2.6); at (1.5,2.5) [$\bullet$]{};
at (5.5,7.5) [\*]{};
(5.5, 7.3)–(5.5,6.7); at (5.5,6.5) [$\bullet$]{};
at (6.5,6.5) [\*]{};
(6.5,6.3)–(6.5,5.7); at (6.5,5.5) [$\bullet$]{};
at (7.5,5.5) [\*]{};
(7.5,5.3)–(7.5,4.7); at (7.5,4.5) [$\bullet$]{};
at (4.5,2.5) [\*]{};
(4.5, 2.7)–(4.5,7.3); at (4.5,7.5) [$\bullet$]{};
(10,5)–(15,5)–(15,9)–(10,9)–(10,5);
(10,0)–(18,0)–(18,9)–(10,9)–(10,0); (10,8)–(18,8); (10,5)–(18,5); (15,0)–(15,9);
at (12.5,7) [$B^{\leq t}$]{}; at (13.7,8.5) [FREE ROW]{}; at (9.2,7) [$M_1$]{}; at (9.2,3) [$M_2$]{}; at (12.5,9.6) [$N_1$]{}; at (16.5,9.6) [$N_2$]{};
at (16.5,6.5) [\*]{};
(16.5,6.7)–(16.5,8.3); at (16.5,8.5) [$\bullet$]{};
at (17.5,5.5) [\*]{};
(17.5,5.7)–(17.5,6.3); at (17.5,6.5) [$\bullet$]{};
at (15.5,1.5) [\*]{};
(15.5,1.7)–(15.5,5.3); at (15.5,5.5) [$\bullet$]{};
(0,4)–(4,4)–(4,9)–(0,9)–(0,4);
(0,0)–(8,0)–(8,9)–(0,9)–(0,0); (0,4)–(8,4); (4,0)–(4,9);
(0,6)–(5,6)–(5,9); (7,9)–(7,6)–(8,6);
(0,7)–(2,9); (0,6)–(3,9); (1,6)–(4,9); (2,6)–(5,9); (3,6)–(5,8); (0,8)–(1,9); (7,8)–(8,9); (7,7)–(8,8); (7,6)–(8,7); (4,6)–(5,7);
(0,4)–(5,4)–(5,5)–(0,5); (7,4)–(7,5)–(8,5); (0,4)–(1,5); (1,4)–(2,5); (2,4)–(3,5); (3,4)–(4,5); (4,4)–(5,5); (7,4)–(8,5);
at (3.2,5.5) [$B^{\leq t}$]{}; at (2.5,7.5) [${\tilde B}^{\leq t}$]{}; at (-.8,7) [$M_1$]{}; at (-.8,3) [$M_2$]{}; at (2.5,9.6) [$N_1$]{}; at (6.5,9.6) [$N_2$]{};
at (6.5,6.5) [\*]{}; at (7.5,5.5) [\*]{};
at (4.5,2.5) [\*]{}; at (5.5,7.5) [\*]{};
We mainly need to show the following.
\[l:improve\] If $(B^{\leq t},\pi)$ is not tight, then there is a tighter diagram $(\Tilde{B}^{\leq t},\Tilde{\pi})$.
By contradiction, suppose that a tighter diagram does not exist. Then, Lemma \[l:sinking\] yields a sequence $k_{l_0},k_{l_0+1}\ldots, k_{m_0}$, where $l_0=0$, $(k_{l_0},\pi(k_{l_0}))\in
M_1\times N_1$, $(k_{m_0},\pi(k_{m_0}))\in M_2\times N_2$ or $k_{m_0}$ is free, $(k_s,\pi(k_s))\in M_2\times N_1$ for all $s=l_0,\ldots,m_0-1$, and $a_{k_s\pi(k_{s-1})}>t$ for all $s=l_0+1,\ldots,m_0$.
If $k_{m_0}$ is free then we define $\Tilde{\pi}$ by $\Tilde{\pi}(k_s):=\pi(k_{s-1})$ for $s=l_0+1,\ldots,m_0$. The row $k_{l_0}$ becomes free, and for all the remaining indices $i$ we define $\Tilde{\pi}(i):=\pi(i)$. We see that the number of intersections of $\Tilde{\pi}$ with $B^{\leq t}$ is one less than that of $\pi$ with $B^{\leq t}$, hence $(B^{\leq t},\Tilde{\pi})$ is tighter.
Otherwise, Lemma \[l:lifting\] yields a sequence $k_{m_0},k_{m_0+1}\ldots, k_{l_1}$, where $(k_{m_0},\pi(k_{m_0}))\in
M_2\times N_2$, $(k_{l_1},\pi(k_{l_1})\in M_1\times N_1$ or $k_{l_1}$ is free, $(k_s,\pi(k_s))\in M_1\times N_2$ for all $s=m_0,\ldots,l_1-1$, and $a_{k_s\pi(k_{s-1})}>t$ for all $s=m_0+1,\ldots,l_1$.
If $k_{l_1}$ is free, then the diagram can be improved as above, replacing $m_0$ with $l_1$ in the definition of $\Tilde{\pi}$.
The composition of sinking and lifting, or if any of these procedures end up with a free row index, will be called a [*(full) turn*]{} of the [*trajectory*]{}.
The sinking and lifting procedures are then applied again and again, until either one of the following holds.
a\) On some turn, let it be turn number $(s+1)$, we encounter a row index $k_{l_s+t}, t\ge 1,$ which is already in the trajectory, written as $k_{l_r+t'}$ (with $r<s$ or $t'=0$ and $r\leq s$). In this case we make a cyclic trajectory $k_{l_s},k_{l_s+1},\ldots, k_{l_s+t},
k_{l_r+t'+1},\ldots, k_{l_s}$ where no two intermediate indices are repeated.
b\) There are no repetitions but we meet a free row index in the end.
In both cases, let $p$ be the length of the trajectory, and rename the indices of the resulting cyclic trajectory without repetitions, or the resulting acyclic trajectory ending with the free row index, to $l_0,l_1,\ldots,l_p$. Clearly, for any two adjacent indices $l_s$ and $l_{s+1}$ of this trajectory, we have $a_{l_{s+1}\pi(l_s)}>t$, and either $(l_{s+1},\pi(l_s))\in M_1\times N_2$, or $(l_{s+1},\pi(l_s))\in M_2\times N_1$. This shows that defining $\Tilde{\pi}$ by $\Tilde{\pi}(l_s)=\pi(l_{s-1})$ for $s=1,\ldots,p$, setting $l_p$ as the new free row in case b), and defining $\Tilde{\pi}(i):=\pi(i)$ for all the remaining row indices, we obtain a tighter diagram $(B^{\leq t},\Tilde{\pi})$, since the number of intersections of $\Tilde{\pi}$ with $B^{\leq t}$ strictly decreases, by the number of full turns made by the trajectory. Thus the diagram $(B^{\leq t},\pi)$ can be improved in any case.
\[t:konig-ext\] Let $A\in\mmset^{(d+1)\times d}$ and let $t$ be the greatest number $h$ such that $A^{(h)}$ has a $d\times d$ submatrix with nonzero permanent. Then for this value $t$ there is a tight diagram $(B^{\leq t},\pi)$, such that all entries of $B^{\leq
t}$ are not greater than $t$, and all entries of $\pi$ are not smaller than $t$.
The König theorem (by the discussion in the beginning of this section) yields a diagram $(B^{\leq t},\pi)$ which is not necessarily tight. However, a tight diagram can be obtained from it by repeated application of Lemma \[l:improve\].
We will prove the following claim [**by induction**]{}:\
If $A\in\mmset^{(d+1)\times d}$ (with finite entries) contains a permutation $\pi$ such that $a_{i\pi(i)}\geq t$ for all $i$ (except for $i$ being the free row $f$), then there is a point $z$ with all coordinates not less than $t$, which internally separates the rows of $A$.
The case $d=1$ is the [**basis**]{} of induction. In this case $A$ consists of just two numbers, say $x$ and $y$, and we can take $z=\max(x,y)$ as the “separating point”. Then one of the numbers belongs to the sector $\{s\mid s\leq z\}$, and the remaining one to $\{s\mid s\geq z\}$.
We now assume that the claim holds for all $d<n$, and let $A\in\mmset^{(n+1)\times n}$ have only finite entries. By Theorem \[t:konig-ext\], there is a permutation $\pi$, a free index $f$ such that $a_{i\pi(i)}\geq t$ for all $i\neq f$, and a submatrix $B^{\leq t}:=A_{M_1N_1}$ with $a_{ij}\leq t$ for $i\in
M_1,j\in N_1$ such that the diagram $(B^{\leq t},\pi)$ is tight. Let $M_2$ and $N_2$ be the complements of $M_1$ in $\{1,\ldots,n+1\}$ and of $N_1$ in $\{1,\ldots,n\}$, respectively. As the diagram is tight, for each column with an index in $N_2$ the corresponding entry of $\pi$ is in $A_{M_1N_2}$. Let $\Tilde{M}_1$ be the set of rows consisting of the free row (which belongs to $M_1$ since the diagram is tight), and the rows of $M_1$ such that $\pi(i)\in N_2$, see Figure \[f:tight\]. Then the number of elements in $N_2$ is one less than that of $\Tilde{M}_1$, and the matrix $A_{\Tilde{M}_1N_2}$ contains a permutation $\pi'$ induced by $\pi$, with all entries not smaller than $t$. Let $n'$ be the number of elements in $N_2$, so $n'<n$. By the induction hypothesis there exists an $n'$-component vector $z$ internally separating the rows of $A_{\Tilde{M}_1N_2}$.
Define $x$ by $x_i=z_i$ for $i\in N_2$ and $x_i=t$ for $i\in N_1$. We claim that $x$ is the separating point. Since the diagram is tight, we have $\pi(i)\in N_1$ for all $i\in M_2$, and we also have $\pi(i)\in N_1$ for all $i\in M_1\bez\Tilde{M}_1$ by the definition of $\Tilde{M_1}$. This implies that $x$ satisfies $a_{i\pi(i)}\geq
t$ for all $i\notin\Tilde{M_1}$, determining the sectors in which the rows with these indices lie. The sectors for the rows with indices in $\Tilde{M}_1$ are determined by $z$ (i.e., by induction), also using that $a_{ij}\leq t$ for all $i\in\Tilde{M}_1$ and $j\in
N_1$.
An application of topological Radon theorem {#s:radon-helly}
===========================================
In this section we go beyond the max-min semiring considering what we call the [*max-T*]{} semiring $\Tmax$: this is the unit interval $\mmset=[0,1]$ equipped with the tropical addition $a\oplus
b:=\max(a,b)$ and multiplication $\otimes_T$ played by a $T$-norm $T\colon \mmset\times\mmset\to\mmset$. These operations were introduced in [@sklar] and a standard reference is the monograph [@pap].
A triangular norm (briefly $T$-norm) is a binary operation $T$ on the unit interval $[0, 1]$ which is associative, monotone and has $1$ as neutral element, i.e., it is a function $T : [0,1]^2\to [0,1]$ such that for all $x,y,z\in [0,1]$:
1. $T(x,T(y,z))=T(T(x,y),z)$,
2. $T(x,y)\le T(x,z)$ and $T(y,x)\le T(z,x)$ whenever $y\le z$,
3. $T(x,1)=T(1,x)=x$.
A $T$-norm is continuous if for all convergent sequences $(x_n)_n, (y_n)_n\in [0,1]^{\mathbb{N}}$ we have $$\lim_{n\to \infty}T(x_n,y_n)=T(\lim_{n\to \infty} x_n, \lim_{n\to \infty} y_n).$$
The axioms of semiring also require $0$ to be absorbing with respect to multiplication, that is, $T(x,0)=T(0,x)=0$. Note that this law follows from (T2,T3) and since $1$ is the greatest element.
The multiplication $\otimes_T$ can be any of the continuous T-norms known in the fuzzy sets theory, including the usual multiplication, $\otimes=\min$ which we studied above, and the [Ł]{}ukasiewicz T-norm $a\otimes_{\text{\L}} b:=\max(0,a+b-1)$.
Note that the case of usual multiplication yields a part of the max-times semiring, isomorphic to the non-positive part of the tropical/max-plus semiring.
Below we consider $\mmset^d$, the set of $d$-vectors with components in $\mmset$, equipped with the componentwise tropical addition and T-multiplication by scalars. A set $C\subseteq\mmset^d$ is called [*max-T convex*]{} if, together with any $x,y\in C$, it contains all combinations $\lambda\otimes_T x\oplus\mu\otimes_T y$ where $\lambda\oplus\mu=1$.
For any set $X\subseteq\mmset^d$, the [*max-T convex hull*]{} of $X$ is defined as the smallest max-T convex set containing $X$. Using the axioms of semiring, or 1)-4) above, it can be shown that the max-T convex hull of $X$ is the set of all [*max-T convex combinations*]{} $$\bigoplus_{i=1}^m \lambda_i\otimes_T x^i\colon m\geq 1,\ \bigoplus_{i=1}^m \lambda_i=1,$$ of all $m$-tuples of elements $x^1,\ldots,x^m\in X$. The max-T convex hull of a finite set of points is also called a [*max-T convex polytope*]{}.
We further make use of the following theorem of general topology that can be found in [@barany]. By the unit simplex of dimension $d$ we mean the set $$\Delta_d=\left \{(\mu_0,\mu_1,\dots,\mu_d)\in \R^{d+1}\vert \sum_{i=0}^d \mu_i=1, 0\le \mu_i\le 1\right \},$$ in the usual real space $\R^{d+1}$ and with the usual arithmetics.
If $f$ is any continuous function from $\Delta_0^{d+1}$ to a $d$-dimensional linear space, then $\Delta_0^{d+1}$ has two disjoint faces whose images under $f$ are not disjoint.
Let $X$ be a set of $d + 2$ points in $\mmset^d$. Then there are two pairwise disjoint subsets $X^1$ and $X^2$ of $X$ whose max-T convex hulls have a common point.
Let $X=\{x^0,x^1,\dots,x^{d+1}\}\subseteq \Tmax^d$. We construct a continuous map $f$ from $\Delta_d$ to the max-T convex hull of $X$ that maps the faces of $\Delta_d$ into max-T convex hulls of subsets of $X$ and apply topological Radon’s theorem to $f$. Define $$\Delta_d^{\max}=\left \{ (\mu_0,\mu_1,\dots,\mu_{d+1})\in [0,1]^{d+1}\vert \max\{\mu_i, 0\le i\le d+1\}=1\right\}.$$
Using ordinary arithmetics, consider the map $\phi_1:\Delta_d^{\max}\to \Delta_0^d$ given by: $$\phi_1(\mu_0,\mu_1,\dots,\mu_{d+1})=\left ( \frac{\mu_0}{\sum_{i=0}^{d+1}\mu_i}, \frac{\mu_1}{\sum_{i=0}^{d+1}\mu_i}, \dots, \frac{\mu_{d+1}}{\sum_{i=0}^{d+1}\mu_i} \right ),$$ which is clearly a homeomorphism, and thus has a continuous inverse. Moreover, for any subset of indices $I=\{i_1,i_2, \dots, i_k\}\subseteq \{1,2,\dots,d+1 \}$, $\phi_1$ maps the max-T convex hull of the standard vectors $e^{i_1}, \dots, e^{i_k}$ into the face of the simplex $\Delta_0^d$ determined by the vertices $e^{i_1}, \dots, e^{i_k}$.
Consider also the map $\phi_2$ defined on $\Delta_d^{\max}$ with values in $\mmset^d$ given by: $$\phi_2(\mu_0,\mu_1,\dots,\mu_{d+1})=\max(\mu_0\otimes x^0,\mu_1\otimes x^1,\dots,\mu_{d+1}\otimes x^{d+1}),$$ which for any subset of indices $I$ as above takes the max-T convex hull of the standard vectors $e^{i_1}, \dots, e^{i_k}$ into the max-T convex hull of the vectors $x^{i_1}, \dots, x^{i_k}$.
Define now $f=\phi_2\circ \phi_1^{-1}$ on $\Delta_0^d$ with values in $\mmset^d$. Applying to it the topological Radon theorem we get the claim.
It is of interest to find a purely combinatorial proof of max-min Radon’s theorem, or in the case of other known T-norms.
The following theorem is known more generally in abstract convexity, as a consequence of Radon’s theorem.
Let $F$ be a finite collection of max-T convex sets in $\mmset^d$. If every $d + 1$ members of $F$ have a nonempty intersection, then the whole collection have a nonempty intersection.
Let $C^1, \dots, C^n$ be max-T convex sets in $\mmset^d$ and suppose that whenever $d+1$ sets among them are selected, they have a nonempty intersection. We proceed by induction on $n$. First assume that $n = d + 2$. Define $x^i$ to be a point in the set $\cap^{d+2}_{j=1; j\not =i}C_j$. We have then $d + 2$ points $x^1, \dots, x^{d+2}$. If two of them are equal, then this point is in the whole intersection. Hence, we can assume that all the $x^i$ are different. By the Radon theorem, we have two disjoint subsets $S$ and $T$ partitioning $\{1,\dots,d + 2\}$ such that there is a point $x$ in $\conv(\cup_{i\in S}x^i)\cap \conv(\cup_{i\in T}x^i)$. This point $x$ belongs to every $C^i$. Indeed, take $j\in\{1,\dots,d + 2\}$, which is either in $S$ or in $T$. Suppose without loss of generality that $j\in S$. Then, $\conv(\cup_{i\in T}x^i)$ is included in $C^j$ , and so $x \in C^j$ . The case $n = d + 2$ is proved.
Suppose now that $n > d+2$ and that the theorem is proved up to $n-1$. Define $C'^{n-1} := C^{n-1}\cap C^n$. When $d + 2$ convex sets $C^i$ are selected, they have a nonempty intersection, according to what we have just proved. Hence, every $d+1$ members of the collection $C^1, \dots, C^{n-2}, C^{n-1}$ have a nonempty intersection.
By induction, the whole collection has a nonempty intersection.
The following two theorems are also known more generally in abstract convexity, as a consequences of Helly’s theorem.
Let $P$ be a collection of $n$ points in $\mmset^d$. Then there exists a point $p\in \mmset^d$ (the *centerpoint*) such that every max-T convex set containing more than $\frac{dn}{d+1}$ points of $P$ also contains $p$.
First construct all max-T convex polytopes containing more then $\frac{dn}{d+1}$ points in $P$. Any point lying in all such polytopes is the required point. Consider a $(d + 1)$-tuple of such polytopes. The complement of each polytope in the tuple contains less then $\frac{n}{d+1}$ points from $P$. The union of all $(d+1)$ complements of the polytopes in the tuple contains less then $n$ points from $P$. Thus the complement of the union, which is the intersection of all polytopes, is nonempty. We only have to prove that given a set of convex polytopes such that every $(d + 1)$-tuple has a non-empty intersection, all of them have a non-empty intersection. But this is Helly’s theorem.
As $\mmset^d=[0,1]^d$ is endowed with the usual Euclidean topology we observe that a max-T convex set is compact if and only if it is closed.
\[helly-infinite\] Suppose $F$ is an infinite, possibly uncountable family of max-T convex and compact sets in $\mmset^d$. Suppose that every $d + 1$ of them have a nonempty intersection. Then the whole family has a non-empty intersection.
Let $F=\{B_i\}_{i\in I}$. According to Helly’s theorem, every finite collection of $B_i$’s has a nonempty intersection. Fix a member $K$ of $F$ and define $G_i=\complement B_i$. Assume that no point of $K$ belongs to all $B_i$. Then the family $\{G_i\}$ form an open cover for the the compact set $K$. One can find a finite subcover $G_{i_1},\dots,G_{i_l}$ such that $K\subseteq G_{i_1}\cup \dots \cup G_{i_l}$. But this means $K\cap B_{i_1}\cap \dots \cup B_{i_l}=\emptyset$, a contradiction.
Let us conclude this section with Tverberg’s theorem for max-T, which can be derived from the more general topological version.
If $f$ is any continuous function from $\Delta_{(d+1)(r-1)}$ to a $d$-dimensional linear space, then $\Delta_{((d+1)(r-1)}$ has $r$ disjoint faces whose images under $f$ contain a common point.
Let $X$ be a set of $(d+1)(r-1) + 1$ points in $\mmset^d$. Then there are $r$ disjoint subsets $X^1,\ldots,X^r$ of $X$ whose max-T convex hulls have a common point.
It is known that the topological Tverberg’s theorem is true for $d\geq 1$ and $r$ equal to a prime number [@BSS-81], and moreover for $d\geq 1$ and $r$ equal to a power of a prime [@Vol-96]. By the above argument, it also shows Tverberg’s theorem in max-T for these cases.
[10]{}
E. G. Bajmóczy and I. B' ar' any. A common generalization of Borsuk’s and Radon’s theorem. Acta Mathematica Hungarica **34** (1979) 347–350.
I. B' ar' any, S. B. Shlosman and A. Szüks. On a topological generalization of a theorem of Tverberg. J. Lond. Math. Soc. **23** (1981) 158–161.
P, Butkovič, K. Cechlárová and P. Szabo, Strong linear independence in bottleneck algebra. Linear Algebra Appl., **94** (1987) 133-155.
G. Cohen, S. Gaubert, J.P. Quadrat, and I. Singer. Max-plus convex sets and functions. In G. Litvinov and V. Maslov, editors, [*Idempotent Mathematics and Mathematical Physics*]{}, volume 377 of [*Contemporary Mathematics*]{}, pages 105–129. AMS, Providence, 2005. E-print arXiv:math/0308166.
M. Develin and B. Sturmfels, Tropical convexity, Documenta Math., **9** (2004) 1–27. M. Gavalec. Solvability and unique solvability of max-min fuzzy equations. Fuzzy Sets and Systems **124** (2001) 385-393.
M. Gavalec. Periodicity in Extremal Algebra. Gaudeamus, Hradec Králové, 2004.
S. Gaubert and F. Meunier. Caratheodory, Helly and the Others in the Max-Plus World. Discrete and Computational Geometry, **43**, (2010) 648–662.
J. Eskeldson, M.Jaffe, V. Nitica. A metric on max-min algebra, Comporary Mathematics, this volume, AMS.
E.P. Klement, R. Mesiar, E. Pap, Triangular Norms, Kluwer Academic Publishers, Dordrecht, 2000.
G. L. Litvinov, V. P. Maslov, G. B. Shpiz, Idempotent functional analysis: An Algebraic Approach, Math Notes **69** (2001), 758–797.
V. Nitica. The structure of max-min hyperplanes. Linear Algebra Appl. **432** (2010), 402–-429.
V. Nitica and S. Sergeev. On semispaces and hyperplanes in max-min convex geometry. Kybernetika **46** (2010), 548–557.
V. Nitica and S. Sergeev. An interval version of separation by semispaces in max-min convexity. Linear Algebra Appl. **435** (2011), 1637–-1648.
V. Nitica and I. Singer. Max-plus convex sets and max-plus semispaces I. Optimization **56** (2007) 171–205.
V. Nitica and I. Singer. Contributions to max-min convex geometry. I. Segments. Linear Algebra Appl. **428** (2008), 1439–1459.
V. Nitica and I. Singer. Contributions to max-min convex geometry. II. Semispaces and convex sets. Linear Algebra Appl. **428** (2008), 2085–2115.
B. Schweizer and A. Sklar, Probabilistic Metric Spaces, North-Holland, New York, 1983
S. N. Sergeev. Algorithmic complexity of a problem of idempotent convex geometry. Math. Notes (Moscow), **74** (2003), 848–852.
A. Yu. Volovikov. On a topological generalization of the Tverberg theorem. Math. Notes (Moscow), **59** (1996), 324-326.
K. Zimmermann. A general separation theorem in extremal algebras. Ekonom.-Mat. Obzor (Prague), **13** (1977), 179–201.
K. Zimmermann. Convexity in semimodules. Ekonom.-Mat. Obzor (Prague), **17** (1981), 199–213.
[^1]: Viorel Nitica was partially supported by a grant from Simons Foundation 208729. Sergeĭ Sergeev is supported by EPSRC grant RRAH15735, RFBR-CNRS grant 11-0193106 and RFBR grant 12-01-00886.
[^2]: One part of the vertices represents the rows, and the other represents the columns. The graph contains an edge between the row vertex $i$ and the column vertex $j$ if and only if $a_{ij}^{(h)}=1$, that is, $a_{ij}\geq h$.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'Hybrid silicon pixel detectors are currently used in the innermost tracking system of the Compact Muon Solenoid (CMS) experiment. Radiation tolerance up to fluences expected for a few years of running of the Large Hadron Collider (LHC) has already been proved, although some degradation of the part of the silicon detector closer to the interaction point is expected. During the LHC upgrade phases, the level of dose foreseen for the silicon pixel detector will be much higher. To face this aspect, dedicated irradiation tests with fluences above $\mathcal{O}(10^{15})$ n$_{\textrm{eq}}$/cm$^2$ have been performed on the silicon sensor and readout chip. Changes in the operation of the sensor and readout chip as a function of the fluence are presented. The charge collection efficiency has been studied: partial recovery of the detector efficiency can be achieved by operating the detectors in a controlled environment and at higher bias voltage.'
author:
- 'G. Tinti for the NSF PIRE group'
title: Study of the readout chip and silicon sensor degradation for the CMS pixel upgrade
---
Introduction
============
The LHC accelerator and the CMS experiment have performed very well in the last year, delivering an integrated luminosity of 3 [$\textrm{fb}^{-1}\xspace$]{}up to the time of these proceedings (September 2011), with maximum instantaneous luminosity of 2.86 $\times 10^{33} \textrm{cm}^{-2} \textrm{s}^{-1}$. The silicon pixel modules are exposed to the highest ionization due to their proximity to the interaction point. A future upgrade of the LHC performance, foreseen in 2016, will push towards reaching high radiation fluences sooner, as the instantaneous luminosity after the upgrade will increase to $2 \times 10^{34} \textrm{cm}^{-2} \textrm{s}^{-1}$ and the radial distance of the first layer of the pixel detector will possibly be reduced. For these reasons, the limits of the radiation hardness for both the sensor and the readout chip is important to determine the detector lifetime.
Section \[CMSpixel\] introduces the characteristics of the CMS pixel detector. Possible changes for the upgrade plan are also discussed. The effects that the radiation damage has on the sensor and the readout chip are presented in Section \[limits\]. The measurements of the performances of irradiated chips and sensors are presented in Section \[measurements\]. In particular, the variation of the collected charge induced by a minimum ionizing particle is presented. The performance of both the sensor and readout chips is examined. Section \[risetime\] is dedicated to the specific changes in the characteristic time of the rising edge of the pulse height signal. Conclusions are presented in Section \[conclusions\].
The CMS sensor and readout chip {#CMSpixel}
===============================
The CMS pixel detector [@tdr; @wolfram] are “hybrid” detectors as the silicon sensor is separated from the silicon readout chip. Each pixel, which has a size of $100~\mu \textrm{m} \times 150~\mu \textrm{m}$, has an indium bump bond connecting the sensor to the readout chip. The sensor has $n^+$ implants on the $n$ substrate and the $p$ junction on the backside. This choice allows the collection of electrons, which have a higher mobility and larger Lorentz angle (i.e., the charge sharing for tracks not perpendicular to the pixel front surface is enhanced). In addition, to extend the radiation hardness tolerance, the substrate is enriched with oxygen. Guard rings on the back side put the sensor edges at ground potential, limiting the risks of sparks at the edges. An extensive review of the radiation hardness property of $n^+$-in-$n$ sensors is available in [@book].
A total of 52$\times$80 pixels are read by the PSI46V2 readout chip (ROC), which is described in [@ROC]. For each single pixel, the signal from the sensor enters two charge sensitive stages: the preamplifier and the shaper. The zero suppression is performed by a comparator. Ideally, the minimum threshold should be as low as possible but above the noise and cross-talk levels. Lowering the threshold to the minimum improves the tracking reconstruction (as the charge sharing between pixels is exploited) and extends the lifetime of the sensor.
The CMS pixel detector system is divided into a three layer barrel and two disk end-caps. The first layer of the barrel, which has a radial distance of 4.4 cm from the interaction point, is the one most exposed to radiation. Two additional layers are placed at radii 7.3 cm and 11.2 cm. The disks are placed at a longitudinal distance $z = \pm 34.5$ cm and $z = \pm 46.5$ cm. This geometry allows very good physics performance, with coverage up to $|\eta|<2.5$ and at least two pixel hits for any track inside the coverage. A detailed description of the CMS pixel detector system and its readout is available in [@wolfram].
The pixel detector in CMS is performing well [@performances]. A Minimum Ionizing Particle (MIP) is expected to create about 22000 electrons. The average noise measured in the present CMS pixel detector is 120 electrons, much lower than the expected signal. However, the minimum threshold for a particle to be detected is on average 2500 electrons. The threshold, which is much higher than the noise, is due to the internal cross talk of the readout chip. Ongoing efforts are devoted to the understanding of different sources of internal cross talk and will be implemented in the new version of the PSI46 ROC.
An upgraded pixel detector is foreseen for the so-called “phase I upgrade” in 2016. A new pixel detector with four barrel layers and three endcap disks is being designed. The new geometry minimizes the material budget while increasing the tracking points up to four. The high rates foreseen justify the study of a fast digital readout architecture to ensure readout efficiency. From the sensor point of view, the major change, in addition to the increase of the rate of events, is the new bi-phase CO$_2$ cooling system [@CO2], which will allow pixel operation at $-20^\circ$C coolant temperature. More details of the upgrade project are given in these proceedings [@dpfupgrade]. The present $\textrm{C}_6 \textrm{F}_{14}$ cooling system has also been designed to run at $-20^\circ$C coolant temperature, but problems in operating the pixel detector due to the high humidity require to run at $+7^\circ$ C coolant temperature. Investigations are ongoing to reach a lower operating temperature keeping the humidity under control.
Radiation damage {#limits}
================
The LHC accelerator has already delivered nearly 3 [$\textrm{fb}^{-1}\xspace$]{}. For the first layer of the barrel pixel detector, 1 [$\textrm{fb}^{-1}\xspace$]{}corresponds to a measured fluence of $2.4 \times 10^{12}$ [$\textrm{n}_{\textrm{eq}}/\textrm{cm}^2$]{}. Up to 2016, when the phase I detector is installed, the predicted integrated luminosity is 70 [$\textrm{fb}^{-1}\xspace$]{}, or approximately a fluence of $2 \times 10^{14}$ [$\textrm{n}_{\textrm{eq}}/\textrm{cm}^2$]{}at the innermost layer. In an extreme scenario where the current pixel detector will be functioning up to an integrated luminosity of 250 [$\textrm{fb}^{-1}\xspace$]{}, the fluence is estimated to be $\mathcal{O}(10^{15})$ [$\textrm{n}_{\textrm{eq}}/\textrm{cm}^2$]{}.
The effects of the radiation damage on the silicon sensor have been extensively described in [@book]. We can expect three major effects:
- **Increase in the leakage current of the sensor**: defects are created in the sensor and cause the leakage current to increase. The leakage current increase can be monitored to understand the radiation damage from the detector itself. At the integrated luminosity of 1 [$\textrm{fb}^{-1}\xspace$]{}, the leakage current per pixel has been measured to be 0.75 nA at the innermost layer. The preamplifier is guaranteed to work up to at least 100 nA/pixel. The leakage current increase will cause a reduction of the pulse height estimated to be $\frac{-0.2\%}{\textrm{nA}}$ per pixel.
- **Charge trapping and partial depletion of the sensor**: there can be a loss in pulse height due to partial depletion or charge trapped in the crystal bulk before collection. The CMS pixel detector does not manifest signs of charge trapping and partial depletion at the present fluence. Partial recovery of the pulse height loss is achievable by increasing the bias voltage.
- **Readout chip damage:** the internal voltages can changed with radiation and affect the operation of the ROC. These kinds of effects are not observed yet for the CMS pixel detector.
The studies presented in this paper intend to determine at which fluence the above radiation induced effects become visible and can affect the full operation of the detectors.
Radiation hardness measurements {#measurements}
===============================
Samples identical to the present CMS silicon sensors and ROCs have been tested to determine the charge collected as a function of the fluence. The samples had been irradiated with protons of 26 MeV momentum at the Karlsruhe Institute of Technology proton cyclotron facility [@kit] in 2010. In addition to unirradiated samples for reference, the studied samples have fluences of $3 \times 10^{14}$ [$\textrm{n}_{\textrm{eq}}/\textrm{cm}^2$]{}, $6 \times 10^{14}$ [$\textrm{n}_{\textrm{eq}}/\textrm{cm}^2$]{}and $1.2 \times 10^{15}$ [$\textrm{n}_{\textrm{eq}}/\textrm{cm}^2$]{}. A sample of fluence $3 \times 10^{15}$ [$\textrm{n}_{\textrm{eq}}/\textrm{cm}^2$]{}was also tested for the measurement presented in Section \[risetime\]. To avoid annealing, the samples were kept in a freezer at $-18^\circ$C when not under test. Preliminary current/voltage curves, which are shown in Figure \[fig:iv\], confirm the goodness of the samples before testing. When under test, the samples were cooled through a Peltier cooler to $-20^\circ$C or $-25^\circ$C depending on the fluence. A stable temperature within $0.3^\circ$C was maintained along with a minimal humidity of less than 30% during the measurement. The setup and procedures implemented here closely follow the measurements presented in [@Rohe:2009; @Rohe:2010].
![Current versus bias voltage plots for the tested samples. The curves do not show any sign of breakdown or anomalous behavior of the tested detectors. The fluence and the temperature at which the samples were tested are reported in the legend. Note that the bias voltage applied and current are negative, but absolute values are graphed.[]{data-label="fig:iv"}](IV.eps){width="90mm"}
The response of the detectors was studied using MIPs produced by the decay of $^{90}$Sr, which has a product endpoint of 2.28 MeV. The trigger was provided by a scintillator positioned underneath the pixel detector. Particles from the radioactive source had to travel both through the pixels and the scintillator to fire the trigger. In addition to MIPs, lower energy beta particles are also produced by decay of $^{90}$Sr. Those particles suffer from multiple scattering and give rise to multiple hit clusters in the pixels. To limit the effect of the low energy particles, only single hit clusters are considered in the analysis. This preferentially selects particles that are traveling from the source to the scintillator perpendicular to the pixels. Due to the uncertainty introduced by the multiple scattering, no absolute sensor efficiencies are reported in this paper.
The most probable value of the charge deposited in the pixels has been determined by fitting the charge collected spectrum with a Landau function convoluted with a Gaussian. The fit has been performed by fitting only the peak region, in order to remove the bias introduced by the tail of the distribution. Figure \[fig:landau\] shows an example of a fitted spectrum.
![Example of the distribution for the collected charge for a non irradiated sample at full depletion bias voltage. The most probable value of deposited charge is obtained by fitting the spectrum in the peak region with a Landau convoluted with a Gaussian, as shown in the plot.[]{data-label="fig:landau"}](landau_fit.eps){width="90mm"}
The collected charge has been studied as a function of the bias voltage. The bias voltage is actually negative, but absolute values are reported here. As it is clearly seen in Figure \[fig: charge\_bias\], by increasing the bias voltage the full depletion of the silicon substrate is achieved and the full charge is collected. In the case of the unirradiated samples, the plateau region is achieved already at a bias voltage of 100 V. When the fluence of the sample increases, it is still possible to achieve the full depletion by increasing the bias voltage up to at least 600 V for a fluence of $1.2 \times 10^{15}$ [$\textrm{n}_{\textrm{eq}}/\textrm{cm}^2$]{}. The figure also shows that the pixel detector could be safely operated even at bias voltages as high as 1000 V.
![Most probable charge versus bias voltage for the samples at different fluences. The plateau region, which indicates that the full depletion has been achieved, is visible at all fluences, but at higher bias voltages for irradiated samples.[]{data-label="fig: charge_bias"}](charge_bias.eps){width="90mm"}
Figure \[fig:charge\_fluence\] shows the charge collected as a function of the fluence once the full depletion is achieved. Even at a fluence of $1.2 \times 10^{15}$ [$\textrm{n}_{\textrm{eq}}/\textrm{cm}^2$]{}, more than 13000 electrons are collected for a MIP. Even though the sensor efficiency is high even at high irradiation, the Lorentz angle, which enhances the charge sharing, diminishes at higher voltages. Separate studies from the one presented here are needed to determine the impact that the lowering of the Lorentz angle has on the spatial resolution.
![Charge at full detector depletion as a function of the fluence. For unirradiated samples the collected charge is approximately 22000 electrons. At a fluence of $1.2 \times 10^{15}$ [$\textrm{n}_{\textrm{eq}}/\textrm{cm}^2$]{}, the collected charge is about 13000 electrons. The measurements presented here agree with the ones reported in [@Rohe:2010].[]{data-label="fig:charge_fluence"}](charge_fluence.eps){width="90mm"}
The readout chips were functioning with minimal changes to the nominal settings. Only small changes to the preamplifier and shaper feedback circuits were applied. The measured noise level has been stable even at high irradiation (at full depletion). In the measurements the threshold was set to 3900 electrons, which is low enough compared to the charge collected. For the sample at $3 \times 10^{15}$ [$\textrm{n}_{\textrm{eq}}/\textrm{cm}^2$]{}, a full collected charge versus bias voltage scan had been collected, but the threshold of 3900 electrons was too high compared to the collected charge. A new collected charge versus bias voltage scan was planned with lower threshold, but the sample stopped functioning for unknown reasons.
Characteristic time measurement {#risetime}
===============================
The rise time of the amplifiers in the readout chip might change due to radiation. As the clock cycle of the ROC is 25 ns (equivalent to the nominal LHC bunch crossing interval), it is important that the pulse height is collected in time. Radiation damage could cause the preamplifier risetime to be slower. The analog pulse shape curve of the PSI46V2 cannot be measured directly. An indirect measurement of the rising edge of the pulse shape can be achieved by scanning the delay between the calibration signal and the leading edge of the clock cycle that is readout. This method has already been explained in [@Rohe:2010]. Here, a quantitative measurement of the characteristic time of the rising edge of the pulse height is performed. The pulse height (PH) behavior as a function of time (t) is parametrized as: $$PH(t)= p_0+ p_1 \cdot t \cdot e^{-\frac{t}{p_2}}
\label{eq:risetime}$$ and fitted to the rising edge curve obtained from the measurement to obtain the parameters $p_0$, $p_1$ and $p_2$. Figure \[fig:risetimefit\] shows the pulse height rising edge and the corresponding fit. To balance the time required for the measurement and enough statistic to cancel the pixel by pixel variation, the measurements of the rising edge presented here are on a uniform sampling of the 25% of the pixels of the ROC.
![Rising edge of the amplifier as obtained by an indirect scan, as explained in [@Rohe:2010]. The curve obtained by fitting the rising edge with the function in equation is superimposed in red.[]{data-label="fig:risetimefit"}](risetimefit.eps){width="90mm"}
The characteristic time is evaluated as the time at which the pulse height reaches 90% of the value at a time $t=p_2$. For unirradiated ROCs without sensor, the characteristic time has been measured to be $(27 \pm 1)$ ns. For unirradiated ROCs with sensor, the measured time is $(30 \pm 1)$ ns. The sensor adds an extra capacitance in series to the readout chip.
The measurement of the characteristic time has also been performed on the irradiated samples to record the variation as a function of the fluence. Table \[tabtime\] shows the measured characteristic times [^1].
Radiation ($10^{14}$ [$\textrm{n}_{\textrm{eq}}/\textrm{cm}^2$]{}) time (ns)
-------------------------------------------------------------------- ------------
0 30 $\pm$ 1
3 38 $\pm$ 2
6 39 $\pm$ 3
12 39 $\pm$ 3
30 36 $\pm$ 2
: Characteristic amplifier risetimes as measured as a function of the fluence received by the sensors and ROCs under test.
\[tabtime\]
Exposing the sample to radiation increases the characteristic time of the preamplifier, but there is not a large difference depending on the irradiation received. This effect might suggest a saturation effect most likely from the sensor after irradiation. Specific tests need to be performed on irradiated bare ROCs (without the sensor) in order to decouple the effects.
The preamplifier risetime, thus the characteristic time, depends on the choice of the analog current in the readout chip. In all the measurements presented here, the analog current has been fixed to 24 mA, using the nominal settings of the PSI46V2 ROC. An increase of the analog current should recover the irradiated risetime values to meet the value before irradiation, but specific studies have not been done yet.
Conclusions
===========
The present CMS sensor and readout chips have been tested up to fluences of $1.2 \times 10^{15}$ [$\textrm{n}_{\textrm{eq}}/\textrm{cm}^2$]{}. By increasing the bias voltage to at least 600 V, the charge collected is more than 13000 electrons, ensuring good sensor efficiency and ROC behavior. The readout chip does not manifest major changes up to the tested fluences. A fluence of $1.2 \times 10^{15}$ [$\textrm{n}_{\textrm{eq}}/\textrm{cm}^2$]{}corresponds to an integrated luminosity higher than 250 [$\textrm{fb}^{-1}\xspace$]{}on the first layer of the barrel pixel detector and assures that the CMS pixel detector can function through the planned phase I upgrade and even longer in case there is a delay. As the CMS pixel are running at somewhat warm temperatures due to humidity problems, the leakage current could become critical for the first layer. After the upgrade, the pixel detector will not suffer from leakage current so much as the operating temperature will be lowered to $-20^\circ$C [@upgrade]. Further measurements are foreseen by the PSI and NSF PIRE groups in order to test ROCs up to a fluence of $5\times 10^{15}$ [$\textrm{n}_{\textrm{eq}}/\textrm{cm}^2$]{}and to properly characterize the behavior of the amplifier risetime as a function of the fluence.
The work presented here could not be possible without the help of the PSI pixel group, in particular Dr. T. Rohe, Dr. W. Erdmann, Dr. H.-C. Kästli and J. Sibille. Prof. A. Bean (KU) has been precious to keep the measurement on track. The measurements were taken at PSI by the NSF PIRE group students, in particular by Joaquin Siado (UPRM). This work is supported by the PIRE grant OISE-0730173 of the US-NFS.
[100]{} The CMS Collaboration, “The CMS experiment at the CERN LHC”, JINS 3 S08004 (2008). W. Erdmann, “The CMS pixel detector”, Int. J. Mod. Phys. A25 (2010) 1315. L. Rossi [*et al.*]{}, “Pixel Detectors. From Fundamentals to Applications”, Spinger (2006). H. C. K[ä]{}stli [*et al.*]{}, “Design and performance of the CMS pixel detector readout chip”, Nucl. Instrum. Meth. A 565 (2006) 188. The CMS Collaboration, “CMS Tracking Performance Results from early LHC Operation”, Eur. Phys. J. C70 (2010) 1165. L. Feld [*et al.*]{}, “CO$_2$ cooling for the CMS tracker at SLHC”, JINST 6 C01091 (2011). N. Parashar, “CMS pixel detector upgrade”, these proceedings. ZAG ZYKLOTRON AG, [*http://www.zyklotron-ag.de/*]{}. T. Rohe [*et al.*]{}, “Signal height in silicon pixel detectors irradiated with pions and protons”, Nucl. Instrum. Meth. A 612 (2010) 493. T. Rohe [*et al.*]{}, “Radiation hardness of CMS pixel barrel modules”, Nucl. Instrum. Meth. A 624 (2010) 414. W. Erdmann [*et al.*]{},“Upgrade plans for the CMS pixel barrel detector”, Nucl. Instrum. Meth. A 617 (2010) 534.
[^1]: The measurement at a fluence of $3 \times 10^{15}$ [$\textrm{n}_{\textrm{eq}}/\textrm{cm}^2$]{}has actually been performed only on 10% of the readout chip pixels. This could explain the slightly lower value obtained for this fluence compared to the other fluences.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
author:
- 'A. J. Gallagher'
- 'M. Bergemann'
- 'R. Collet'
- 'B. Plez'
- 'J. Leenaarts'
- 'M. Carlsson'
- 'S. A. Yakovleva'
- 'A. K. Belyaev'
bibliography:
- '../../../master.bib'
date: 'Received ... / Accepted ...'
subtitle: 'II. 3D non-LTE formation of lines in the solar atmosphere'
title: Observational constraints on the origin of the elements
---
Introduction {#sec:intro}
============
Barium is key element that is used in heavy element studies in stars. Its abundance patterns in the halo, in field stars, and in clusters have been carefully measured over the past several decades. Barium, like most other heavy elements, mostly forms via a series of neutron captures through either the rapid (r-) process or slow (s-) process channels. These two neutron capture channels have very different sites. After the discovery and analysis of 2017gfo [the electromagnetic counterpart of GW170817 @Valenti2017], it is highly probable that the r-process mostly occurs in neutron star mergers [@Thielemann2011]. Conversely, the majority of barium in the Sun [$81\%$ @Arlandini1999] ostensibly formed via the s-process in thermally-pulsing asymptotic giant branch (TP-AGB) stars [@Smith1988]. However, other sites for the s-process and r-process do and most likely exist. Naturally, the barium isotope ratio, ${f_{\rm odd}}$[^1], of a star is a useful quantity as it provides precise information on the s- and r-process contribution, but is exceedingly difficult to measure. Therefore, this information is only measured in some thick disk and halo stars, where this parameter is most interesting [@Magain1993; @Magain1995; @Mashonkina1999; @Gallagher2010; @Gallagher2012; @Gallagher2015].
Most abundances, save those such as lithium that are measured in absolute units, are measured relative to the solar abundances. This helps to mitigate systematic errors within spectroscopic abundances and yields extra information about stellar populations, evolutionary stages and ages, that measurements in absolute units might not. As a result, the solar abundances are extremely important to stellar astrophysics. Consequently, very accurate measurements of the solar abundance are needed, which employ sophisticated model atmospheres and spectrum synthesis techniques such as 3D hydrodynamics and non-local thermodynamic equilibrium (non-LTE) physics. In recent years, with the development of faster and larger computers, it has been possible to develop and implement these methods [@Asplund2003; @Steffen2015; @Klevas2016; @Amarsi2016o; @Mott2017; @Amarsi2017si; @Nordlander2017].
One of the main aims of this paper series is to report on our development of the one dimensional (1D) and three-dimensional (3D) statistical equilibrium codes – [MULTI 2.3]{} [@Carlsson1986] and [MULTI3D]{} [@Leenaarts2009] – as we include new or better physics into their program flows. Given how important barium is to galactic chemical evolution studies because it traces the impact of neutron-capture nucleosynthesis, we present a thorough analysis of the solar barium abundance using a handful of optical lines computed using the two statistical equilibrium codes and the same barium model atom.
The statistical equilibrium of has already been a subject of several detailed studies [@Mashonkina1996; @Mashonkina1999; @Shchukina2009; @Andrievsky2009; @Korotin2015]. The first such study was conducted by @Gigas1988 in Vega. There are, however, important differences between our work and these earlier studies. First, we use the new quantum-mechanical rates for transitions caused by inelastic collisions with hydrogen atoms from @Belyaev2018. We also examine the impact dynamical gas flows have on by utilising a 3D radiative hydrodynamical model to compute full 3D non-LTE radiative transfer, as well as 3D LTE, 1D LTE, and 1D non-LTE. Ab initio collisional damping from @Barklem2000 was included in the linelist.
It has been observationally confirmed that the resonance line at 4554Å is sensitive to the chromospheric effects[^2], and so naturally a polarised spectrum of the resonance line is also sensitive to the quantum interferences [see, e.g. @Kostik2009; @Shchukina2009; @Belluzzi2013; @Smitha2013; @Kobanov2016], however, this is beyond the scope of this paper.
The paper is structured as follows. In Sect. \[sec:model\] we describe the observations, we detail the model atmospheres, model atoms and spectral synthesis codes; in Sect. \[sec:lineformation\] we discuss the impact that various model assumptions have on our results; in Sect. \[sec:results\] we describe the analysis and results from our line analysis; and in Sect. \[sec:conclusions\] we summarise the study.
Models and Observations {#sec:model}
=======================
Solar spectrum {#sec:observations}
--------------
The solar spectrum is taken from the Kitt Peak National Observatory (KPNO) solar atlas published by @Kurucz1984. This solar atlas covers the spectral range of $3\,000$ to $13\,000$Å at a typical resolution ${\rm R}\equiv\frac{\lambda}{\Delta\lambda}=400\,000$. Although newer solar spectra exist such as the PEPSI spectrum provided by @Strassmeier2018, we chose to work with the former atlas as it has a very high resolution, roughly twice that of the latter. Nevertheless, comparisons of these two spectra have previously been made and they were found to be in very good agreement with one-another [@Osorio2019].
1D model atmosphere {#sec:1Datmos}
-------------------
We use the MARCS model atmosphere that was computed for the Sun from the opacity sampled grid published in @Gustafsson2008. The solar parameters of this model are ${T_{\rm eff}}/{\log{g}}/{{\rm [Fe/H]}}=5777/4.44/0.00$ and include a mixing length parameter, $\alpha_{\rm MLT}=1.50$. The solar composition used to compute the model opacities are based on those published in @Grevesse2007.
3D model atmosphere {#sec:3Datmos}
-------------------
For the work presented in this study we make use of the solar [[stagger]{}]{} [@Nordlund1994; @Nordlund1995] model with stellar parameters ${T_{\rm eff}}/{\log{g}}/{{\rm [Fe/H]}}=5777\,{\rm K}/4.44/0.0$, from the [[stagger]{}]{} model atmosphere grid [@Collet2011; @Magic2013a]. A 3D model consists of a series of computational boxes that represent a time series, which are commonly referred to a snapshots. These snapshots are selected from a larger time series of snapshots that are produced from the [[stagger]{}]{} code and are selected at a time when the simulation has reached dynamical and thermal relaxation. For our purposes – and for the sake of time – we have chosen to work with five snapshots, each consisting of $240\times240\times230$ grid points which cover a geometrical volume of $7.96\times7.96\times3.65$Mm in the $x$, $y$, and $z$ dimensions, respectively. These are used as independent input for our statistical equilibrium code [MULTI3D]{} (Sect. \[sec:multi3d\]) and then the output are averaged together, therefore applying the ergodic approximation that averaging in time is equivalent to averaging over space. In this case it is assumed that averaging in time is equivalent to averaging across the disc of the Sun. The emergent fluxes from these snapshots have an equivalent width variance of only $\sim0.75$mÅ, suggesting that including further snapshots to the study will not greatly improve the results presented in this study, only increase the computational times.
Line opacities were collected from the MARCS database and are sorted into 12 opacity bins. Continuous absorption and scattering coefficients are taken from @Hayek2010. Importantly, and unlike an equivalent 1D model, 3D models provide $x$, $y$, and $z$ velocity fields for every voxel meaning that post-processing spectrum synthesis codes provide more accurate approximations for the Doppler broadening, including asymmetric line profiles, which result from these gas flows.
We also make use of the averaged 3D model to help make qualitative comparisons between the full 3D and 1D models, however, we do not use it with [MULTI 2.3]{} or [MULTI3D]{}. A [$\langle{\rm 3D}\rangle$]{} model is computed from a 3D model by spatially averaging the thermal structure of the 3D computational box over surfaces of equal Rosseland optical depth. As this can be performed in several different ways, comparing results from different studies that do not specify their averaging techniques is ultimately self-defeating.
![3D,1D and [$\langle{\rm 3D}\rangle$]{} temperature structure.[]{data-label="fig:tstruct"}](sun){width="\linewidth"}
Figure \[fig:tstruct\] depicts the 3D solar temperature structure (blue 2D histogram), along with the 1D MARCS (dashed line) and [$\langle{\rm 3D}\rangle$]{} (solid red line) temperature structures. It is clear that the average temperature of the full 3D model and the 1D model are fairly consistent in the outermost regions of the atmosphere (as seen by comparing the 1D with the [$\langle{\rm 3D}\rangle$]{} model). However, in deeper regions of the models – where the continuum usually forms (${\log{\tau_{\rm ROSS}}}\approx0$) – the models begin to diverge. This is mostly due to the differences between the convection indicative in the 3D hydrodynamic model atmosphere – which the [$\langle{\rm 3D}\rangle$]{} model traces – and the treatment of convection theory (in this case the mixing length theory) in the 1D model atmosphere.
[MULTI 2.3]{} {#sec:multi1d}
-------------
MULTI solves the equations of radiative transfer and statistical equilibrium in 1D geometry with 1D model atmospheres. The latest release of MULTI is [MULTI 2.3]{}. However, we have made several minor changes to [MULTI 2.3]{} for our purposes including, the ability to compute the detailed balance for charge transfer processes between ions and hydrogen. We include a fixed microturbulence value of $1\,{{\rm km\,s^{-1}}}$ in our computations with the solar MARCS model. The flux data were computed using five $\mu$-angles, assuming a Gaussian quadrature scheme taken from @Lowan1942.
[MULTI3D]{} {#sec:multi3d}
-----------
[MULTI3D]{} is an message passing interface (MPI)-parallelised, domain-decomposed 3D non-LTE radiative transfer code that solves the equations of radiative transfer using the Multi-level Accelerated Lambda Iteration (ALI) method [@Rybicki1991; @Rybicki1992] for 3D model atmospheres. Every element that is modelled by [MULTI3D]{} is assumed to have no effect on the model atmosphere, as it is in [MULTI 2.3]{}. This is a good assumption for barium as it is not an electron donor nor does it have a high impact on the overall opacity, unlike magnesium or iron, for example.
At present, it will accept three types of 3D model atmospheres formats as direct input, including those computed using Bifrost [@Gudiksen2011], and [[stagger]{}]{}. While Bifrost models are read using MPI IO, the [[stagger]{}]{} models are, at present, not read this way due to complications in converting byte ordering. However, the added delay to the code’s run time is minimal, and only becomes noticeable when [MULTI3D]{} is run on several hundred CPUs. In addition to these two types of model atmosphere, the code will also accept any 3D model formatted so that the temperature, $T$, density, $\rho$, electron number density, $n_{\rm e}$, and $x$, $y$, and $z$ velocity fields are supplied on a Cartesian grid that is both horizontally periodic and equidistantly spaced. Therefore, it is relatively straightforward to convert almost any 3D model to this input format for [MULTI3D]{}.
We have introduced new coding for computing fluxes inside [MULTI3D]{} along with the appropriate post-processing routines designed to extract the flux data. All of the output flux data computed for the work presented here was calculated using a Lobatto quadrature scheme and the appropriate corresponding weights [@Abramowitz1972]. At a later stage of this paper series, other quadrature schemes will be introduced, as well as internal routines that will compute fluxes inside [MULTI3D]{} and write them as output.
[MULTI3D]{} is now capable of accepting model atoms that include hyperfine structure (HFS) and isotope shift information for any atomic transition. This means that lines with highly asymmetric profiles, caused by these effects, can now be adequately modelled by [MULTI3D]{}. To test this upgrade, and to test that we could limit the impact of systematic errors dominating the abundances and abundance corrections we provide, we compared 1D spectra computed by both [MULTI 2.3]{} and [MULTI3D]{}. This was conducted only for the vertical intensity ($\mu=1$), using a small test barium model atom, under the assumption of LTE. We use the same opacity sources, and same input model atmosphere. Systematic differences in the equivalent widths of less than $2.2$% were found between intensities computed with [MULTI 2.3]{} and intensities computed with [MULTI3D]{}. This translates to abundances differences much less than 0.01dex. The reason for these small differences is likely because of the way each code solves the radiative transfer equation; [MULTI3D]{} uses a direct 1D integration of the radiative transfer equation when computing spectra from 1D model atmospheres[^3], while [MULTI 2.3]{} utilises a faster Feautrier method. However, abundance uncertainties found here are far smaller than the errors we report in Sect. \[sec:results\]. Therefore, we were satisfied that comparing 1D output from [MULTI 2.3]{} with 3D output from [MULTI3D]{} was adequate.
We ran [MULTI3D]{} in short characteristic 3D solver mode and used the solar [[stagger]{}]{} model as input. The [[stagger]{}]{} model’s $xy$ grid points were scaled down by a factor of 64 from $240\times240$ to $30\times30$ grid points using a simple bilinear interpolation scheme. Significant tests conducted in the first paper in this paper series, [@Bergemann2019 henceforth, ], revealed no significant loss of information in the horizontal gas flows that affected the line profiles in any noteworthy way. The horizontal components were also assumed to be periodic so that rays with very low $\mu$ angles could be computed without encountering a horizontal boundaries. The vertical grid size remained consistent with the original model atmosphere at 230 grid points.
Model of Barium atom {#sec:atom}
--------------------
The model atom of barium is constructed as follows. The energy levels for the and levels are extracted from the NIST database. Of these, we include eight energy states of up to the energy of $2.86$eV, and all available levels of up to $9.98$eV. Fine structure is retained for the three lowest terms of : (ground state), ($\sim 0.65$eV), and ($\sim 2.6$eV) (Table \[tab:config\]). Transitions between these terms are typically used in the barium abundance analysis of cool stars. Other levels are merged into terms, and their energy levels are represented by the weighted sum of the individual components (weighted by the statistical weights of the levels). In total, the model is comprised of $110$ states and is restricted by the ground state of at $15.2$eV for a total of $111$ levels. Note that no lines of are observed in the spectra of FGK stars. For the Sun $n_{\ion{Ba}{i}}/n_{\ion{Ba}{ii}}\approx10^{-4} - 10^{-6}$ (see Sect. \[sec:nlte\]), hence, is a minority species. As such, a detailed treatment of the neutral stage is of no importance to the statistical equilibrium , which is the majority species in these stars. Figure \[fig:ba\_grot\] depicts the energy levels in and transitions among them in the form of a Grotrian diagram. We have colour-coded the four lines we use here as follows: gold represents the 4554Å line; green represents the 5853Å line; blue represents the 6141Å line; and red represents the 6496Å line. As there are too many energy levels in the model atom to accurately depict without overlapping energy states, this figure should be used for qualitative assessments only.
![The Grotrian diagram of atom. Gold, green, blue, and red lines indicate the diagnostic lines that we use in the solar abundance analysis at 4554, 5853, 6141, 6496Å, respectively.[]{data-label="fig:ba_grot"}](ba_plot_grotrian){width="\linewidth"}
The radiative bound-bound transitions for were extracted from the [Kurucz database](http://kurucz.harvard.edu/atoms/5601/gf5601.all), 26.03.2017. We also compared the data with the [NIST database](https://physics.nist.gov/PhysRefData/ASD/lines_form.html). For the combined terms, the lines were merged and the transition probabilities co-added as described in @Bergemann2012a. The oscillator strengths of the four diagnostic lines used for the present analysis were extracted from the experimental transition probabilities presented in @DeMunshi2015 and @Dutta2016. In total, the model contains $284$ spectral lines in the wavelength range from $1330$ to $202930$Å. The transitions with $\log gf < -10$ are not included. Most of these lines are represented by nine frequency points, except for the four diagnostic lines that we use in the abundance analysis, i.e the lines at $4554$, $5853$, $6141$, and $6496$ Å. These lines were represented with a profile containing $301$ frequency points. Several UV lines are rather strong. To test whether nine frequencies were enough to describe those strong lines, we ran a test using [MULTI 2.3]{}. The departure coefficients from the model atom were compared with an atom that contained $100$ frequency points for two strong UV transitions at $2304.247$Å and $2341.429$Å. It was found that these transitions have no effect on the populations of the four barium lines of interest. This justifies the number of frequency points chosen for transitions that were not of interest to us for this study. Damping by elastic collisions with hydrogen atoms are computed using the $\alpha$ and $\sigma$ parameters from @Barklem2000 where available. When this information was missing, we used the Uns[ö]{}ld approximation, which was scaled by $1.5$. The wavelengths are taken from @Karlsson1999. We found that for three lines (NB: not the line at 4554Å) there was a systematic offset to the solar spectrum. Once the solar spectrum was corrected for gravitational redshift they did match the observed line positions, although the 4554Å resonance line has a slightly different shift. @Karlsson1999 underline that this line might be slightly shifted in their measurements, due to the isotopic mix they used and self-absorption in this strong line. This shift is however expected to be at most 1mÅ to the blue (Litzén, private comm.) not sufficient to explain the remaining offset we observe. We discovered that the excess shift in this line was due to convective effects. We shifted the 4554Å line by 2mÅ in 1D to the blue to match the observed position, whereas we did not need to shift the 3D profile. We also introduce HFS and isotopic shifts. They were computed using the solar abundance ratios of the five barium isotopes, see @Eugster1969. The odd barium isotopes have non-zero nuclear spins that causes hyperfine splitting of the levels. The magnetic dipole and electric quadrupole constants for the five relevant energy levels were taken from @Silverans1986, and @Villemoes1993. The isotopic shifts are provided by @VanHove1982 for the 5853 and 6141Å lines, by @Villemoes1993 for the 6496Å line, and by @Wendt1984 for the 4554Å line. The diagnostic lines are hence represented by six to 15 HFS components. The complete HFS information for these lines can be found in tables in Appendix \[apdx:hfs\]. Oscillator strengths for these lines are computed from accurate experimental transition probabilities in @DeMunshi2015 @Dutta2016.
The radiative bound-free data are computed using the standard hydrogenic approximation (Kramer’s formula). This is appropriate since the first ionisation potential of is at 10eV, and the energy levels, which may contribute to radiative over-ionisation at the solar flux maximum, have very low population numbers. Also on the basis of earlier studies with strontium [@Bergemann2012b], which has a similar atomic structure, we do not expect that photo-ionisation is a significant non-LTE effect. In fact, earlier studies of barium in non-LTE showed that a majority ion and is collision-dominated [see @Mashonkina1999; @Gehren2001; @Bergemann2014]. In such ions, the statistical equilibrium is established by a competition of collisional thermalisation, photon losses in strong lines, and over-recombination. This will be discussed in detail in the next section.
One of the new features of our atom, compared to earlier studies mentioned above, is the treatment of collisions. In particular, we include the new quantum-mechanical rate coefficients for the inelastic collisions between the ions and hydrogen atoms by @Belyaev2017 [@Belyaev2018]. To the best of our knowledge, the first study of barium that employs these detailed quantum-mechanical data for collisions with hydrogen was recently published by @Mashonkina2019 for the purposes of treating isotopes. The present paper is the first time these hydrogen-collision data have been employed for full non-LTE modelling. The data are available for 686 processes, and represent collisional excitation and charge transfer reactions, i.e. ${\rm Ba + H} \leftrightarrow {\rm Ba}^{+} + {\rm H}^{-}$. The rate coefficients are typically large and may exceed $10^{-8}\,{\rm cm}^3{\rm s}^{-1}$ in the temperature regime relevant to modelling the solar atmosphere. Note that here the process ionisation refers to the ion we are interested in and a free electron, i.e. ionisation: ${\ion{Ba}{ii}}+ {\rm H} \to {\ion{Ba}{iii}}\ + {\rm H + e}$, but the process ion-pair formation reads: ${\ion{Ba}{ii}}+ {\rm H} \to {\ion{Ba}{iii}}\ + {\rm H}^{-}$, that is, loses its outer electron and it is bound with H. Its inverse process, mutual neutralisation, is the process when gains an electron from H$^{-}$. The same is valid when is replaced by and is replaced by . Charge transfer reactions do not lead to a free electron, which is the case that is usually modelled by a Drawin’s formula [@Drawin1968; @Drawin1969]. Excitation and ionisation by collisions with free electrons are computed using the @vanRegemorter1962 and @Seaton1962 formulae. A study of the impact of different collisional rates was presented in @Andrievsky2009 [Sect. 3.2]. No differences between these classical recipes were found. An earlier version of this model atom was used in [@Eitner2019]. We have since updated the oscillator strengths.
Barium line formation {#sec:formation}
=====================
non-LTE effects {#sec:nlte}
---------------
In terms of departures from LTE, is a collision-dominated ion [see @Gehren2001 and references therein]. The ionisation potential is too high for photo-ionisation to play a significant role in the statistical equilibrium (SE) in FGK type stars. On the other hand, the term structure of the ion, with several very strong radiative transitions, favours strong effects caused by line scattering. In particular, there is radiative pumping at the frequencies of optically-thin line wings, $\tau_{\rm wing} < 1$, as long as at the line centre $\tau_{\rm core} > 1$. This mechanism acts predominantly at larger depths and leads to over-population of the upper levels of the transitions, at the expense of the lower states. On the other hand, strong downward cascades occur higher up in the atmosphere, where the strong line cores become optically thin, $\tau_{\rm core} < 1$. This mechanism de-populates the upper levels via spontaneous de-excitations and this downward electron cascade causes over-population of the lower-lying energy states. As the statistical equilibrium of has been extensively studied in the literature [@Mashonkina1996; @Mashonkina1999; @Short2006; @Mashonkina2008], hence in what follows we will only describe the main features of the non-LTE line formation and discuss the differences with the earlier studies.
![Departure coefficients for the levels as a function of the continuum optical depth ${\log{\tau_{\rm 5000}}}$ computed using the reference barium model atom for the solar MARCS model atmosphere.[]{data-label="fig:deps"}](quantum){width="\linewidth"}
Quantitatively, this behaviour can be visualised as plots of the departure coefficients $b_i$[^4] as a function of the continuum optical depth at 5000Å, ${\log{\tau_{\rm 5000}}}$. Figure \[fig:deps\] depicts the $b_i$ behaviour for the solar MARCS model atmosphere. To facilitate the comparison with the detailed study by @Mashonkina1999, we have chosen the same axis range as that paper. Thick coloured curves correspond to $5$ energy states, which are involved in the radiative transitions listed in Table \[tab:config\]: the ground state of , , and the low-excitation terms and . Thin grey dotted curves show all other energy levels of in the model atom. It is interesting that despite major differences in the model and numerical methods, including the properties of the atomic model, model atmosphere, and the SE code, the agreement between our results and that of @Mashonkina1999 [see their Fig. 2] is very good. In particular, the ground state is entirely thermalised throughout the full optical depth range and develops a very modest over-population only at ${\log{\tau_{\rm 5000}}}< -4$. The first excited metastable state is also close to LTE due to strong collisional coupling with the ground state, although minor departures in the atomic number densities, of the order of a few percent, are seen at ${\log{\tau_{\rm 5000}}}< -3$ and higher up. The term at $\sim 2.6$eV shows stronger deviations from LTE already in the deep layers, ${\log{\tau_{\rm 5000}}}\sim -2$, where the non-LTE population of the level is only about $80\%$ of the LTE value (that is equivalent to $b_i = 0.8$). The pronounced depletion of the term is caused by photon losses in the lines, connecting the ground state and the lowest metastable state with . This under-population increases outwards as the lines become optically thin. All energy levels above are over-populated at ${\log{\tau_{\rm 5000}}}< 0$, the effect that @Mashonkina1999 attributed to radiative pumping.
It is interesting to briefly discuss the importance of micro-physical processes in the SE of . @Andrievsky2009 suggest that photo-ionisation cross-sections is the main source of uncertainty in the SE of . This is only true for , however, no lines of the neutral atom are observed in the optical or infra-red spectra of FGKM stars [@Tandberg1964]. Over-ionisation of has no effect on the population of , as the ratios of number densities of two ionisation stages are $n_{\rm Ba I}/n_{\rm Ba II} \sim 10^{-4}$ at ${\log{\tau_{\rm 5000}}}\approx 0$, and drops to $\sim 10^{-6}$ in the outermost atmospheric layers in the solar model. This ratio is even more extreme in the atmospheres of metal-poor stars. For example, for a model atmosphere of an RGB star with ${T_{\rm eff}}= 4600$K, ${\log{g}}= 1.6$, and ${{\rm [Fe/H]}}= -2.5$, $n_{\rm Ba I}/n_{\rm Ba II} \sim 10^{-5}$ at ${\log{\tau_{\rm 5000}}}\approx 0$, but approaches $\sim 10^{-10}$ close to the outer boundary at ${\log{\tau_{\rm 5000}}}\approx -5$. On the other hand, photoionisation in is not important. The ion has a very high ionisation potential, and its well-populated energy levels with low excitation potentials have ionisation thresholds in the far-UV, at $\lambda < 1000$ Å, where radiative flux in FGK-type stars is negligibly small. We also recomputed the departure coefficients assuming ${\sigma_{\rm photo}}/100$ and ${\sigma_{\rm photo}}\times100$. This is a very conservative estimate of uncertainty in the cross-sections, when comparing to a very similar atom, Sr, for which detailed quantum-mechanical cross-sections are available from [@Bergemann2012a]. It was found that only the non-LTE populations of the ground state change, but none of the important levels show any difference with respect to our reference model.
A more important ingredient for the SE of seems to be the accuracy of the data for inelastic collisional processes, in particular, those between and atoms. [@Short2006] suggest that the non-LTE line profiles are invariant to a factors of $0.1-10$ changes in the rates of transitions caused by collisions (NB they used approximate analytical formulae to represent these data). This may hold for a limited range of stellar parameters. For example, in the case of the Sun, using the Drawin’s recipe or QM data does not give significantly different results. However, it is known that metal-poor stars are sensitive to non-LTE effects [@Bergemann2014]. Therefore, it would be reasonable to assume that they would also be sensitive to different collisional recipes. This is particularly true in hydrodynamic model atmospheres as the decoupled non-local radiation field leads to a cooling in the outer regions of the temperature structure, relative to the equivalent 1D model [see e.g. @Gallagher2016 their Fig. 1]. We intend to explore this in the near future.
Line formation in hydrostatic and inhomogeneous models {#sec:lineformation}
------------------------------------------------------
{width="\linewidth"}
We begin with the analysis of LTE and non-LTE formation of lines in the 1D hydrostatic solar model. We will then extend the analysis to radiative transfer with 3D inhomogeneous models.
As discussed in the previous section, the non-LTE effects in are primarily dominated by line scattering. Consequently, deviations from LTE in the line source function are expected to be significant. Since the ratio of the departure coefficients for the lower $i$ and upper $j$ level of the transitions is below unity for the diagnostic lines, $b_j/b_i < 1$, the ratio of source function to the Planck function [see @Bergemann2014 for the derivation], also drops below unity. In other words, the source function in the line is sub-Planckian and the non-LTE lines profiles shall come out stronger than the LTE lines. In some cases [e.g. @Bergemann2012a for Sr], this effect is modulated by the change of the line opacity. However, since $\kappa_\nu \sim b_i$, and the population numbers for the lower levels of all lines are essentially thermal throughout the line formation depths, the line opacity is very close to its LTE value. This simple analytical picture is confirmed by comparing the LTE and non-LTE line profiles (Fig. \[fig:LTENLTE\]). The non-LTE effects are small and amount to the abundance difference of $-0.05$ (4554 Å) to $-0.1$ dex (5853Å). The other two lines show similar behaviour.
Figure \[fig:LTENLTE\] demonstrates that the 3D profiles are asymmetric and also shifted blueward relative to the 1D profiles, which is expected. This is a natural result of the convective motions inside the 3D model, that the 1D model cannot replicate [@Lohner-Bottcher2018; @Stief2019]. This is particularly obvious in the three subordinate lines, where the HFS has far less impact to the line shape than it does in the resonance line, where asymmetries are seen in both 1D and 3D. The 3D profiles, for a given barium abundance, are consistently weaker than their 1D counterparts, both in LTE and non-LTE. Therefore, positive abundance corrections are going to be needed to reproduce the same equivalent width, foreshadowing larger 3D LTE and non-LTE abundances over the 1D LTE counterpart. Like was shown in the hydrostatic case above, deviations from LTE should be significant because of line scattering.
{width="\linewidth"}
The lines of in the solar spectrum are strong, with EW’s from 207mÅ (4554 line) to 68mÅ (5853 Å) line. As such, they are not only extremely affected by damping, but are also blended. The 4554 Å line is blended by a feature close the line core, but has little impact on the line. The subordinate lines also show some blending. @Korotin2015 avoided, in particular, the 4554 Å line in their analysis of FGK metal-poor stars in stars where ${{\rm [Fe/H]}}>-1.0$. On the other hand, @Grevesse2015 included the 4554, 5853, and 6496 Å line in their analysis of the Sun. They also used the non-LTE corrections for lines. We explore the effects of line blending and damping in Sect. \[sec:results\]. The current photospheric solar abundance of barium derived by @Grevesse2015 – who applied non-LTE corrections to their 3D LTE abundance – is $2.25\pm0.03\pm0.07$ (where errors represent the statistical and systematic errors, respectively). The abundance of barium in CI meteorites [@Lodders2003] is $2.19\pm0.03$. The main goal of this paper is to explore whether ab initio atomic data from physical experiments and detailed quantum-mechanical calculations are successfully able to describe the spectrum of the Sun. Accordingly, we now move on to report our 1D non-LTE, 3D LTE and 3D non-LTE corrections relative to the average barium 1D LTE abundance.
Computing abundance corrections {#sec:method}
===============================
A grid of abundances for all four barium lines were computed using [MULTI 2.3]{}. These were fit to the solar spectrum using a $\chi^2$ code that treats abundances, macroturbulent broadening and wavelength shifts as free parameters (see Sect. \[sec:atom\] for details of line shifts). The macroturbulences found ranged from $1.5-1.9\,{{\rm km\,s^{-1}}}$. The code also normalises the fit to a local continuum for each line using two patches of spectrum either side of the line. We fixed the rotational broadening to ${v\sin{i}}=1.6\,{{\rm km\,s^{-1}}}$ [@Pavlenko2012]. The 1D LTE we attain represent the best statistical fit from this $\chi^2$ code.
We computed the four barium lines in 1D non-LTE, 3D LTE and 3D non-LTE using three abundances; ${A({\rm Ba})}=2.17$, $2.27$ and $2.35$ - covering typical abundances reported in the literature for the solar barium abundance. The abundance corrections tabulated in Table \[tab:corrections\] were determined by fitting the grid of 1D LTE profiles to the 1D non-LTE, 3D LTE and non-LTE lines so that their equivalent widths matched. Unsurprisingly, it was found that the corrections from all three abundances were identical, hence corrections we provide are robust against the typical abundance range found by most studies on the solar barium abundance.
Results {#sec:results}
=======
Our best fit profiles are compared with the solar flux profiles in Fig. \[fig:bestfits\]. The best fit 3D lines are computed by [MULTI3D]{} using the corrections given in Table \[tab:corrections\]. The 3D non-LTE abundances are remarkably consistent, apart from the 6141Å line, which is approximately $0.12$dex larger than the other three. The reasons for this will become apparent by the end of this section. Unlike in the 1D case, the best fit 3D LTE profiles show large deviations in the line cores, relative to their non-LTE counterparts, yet their equivalent widths remain very similar.
The lines and abundances discussed this far have assumed that the barium lines are unblended in the solar spectrum. In reality this is not the case. In fact, the abundances all barium lines (particularly the $6141$Å line) are dependent on line blending, as we now discuss.
Line blending corrections {#sec:blending}
-------------------------
The four barium lines we model here suffer from the effects of blending with other atomic and molecular species. Naturally, this impacts the abundances we derive when we assume that the line is clean. This is what was done when synthesising the lines with [MULTI3D]{} and [MULTI 2.3]{}, as these codes do not currently possess the capability to synthesise blends. To examine this, we used the VALD3[^5] database together with the barium line information extracted from the model atom to create new line lists for the 1D LTE spectrum synthesis code, MOOG[^6]. This code was chosen as the interactive plotting tool makes recomputing synthetic spectra and fitting it with observed data very simple. Lines were computed with and without and the abundance of the clean barium line was adjusted until its line strength matched the blended line.
The resonance line at $4554$Å is the strongest line presented here. Naturally, it would dominate most of the line depression at this spectral region. It is found that when blends are included, the 1D LTE barium abundance must be reduced by $0.01$dex. The $5853$Å line abundance was reduced by $0.03$dex. The $6141$Å line was found to be severally affected by blending, as the abundance had to be reduced by $0.16$dex. The $6496$Å barium abundance had to be reduced by $0.04$dex.
We will use these abundance corrections to determine the barium abundance in all four paradigms in Sect. \[sec:baabundance\]. First, however, it is important to determine how the barium lines are affected by systematic uncertainties, as we now present.
Line damping uncertainties {#sec:damping}
--------------------------
We have previously mentioned that differences in the radiative transfer solvers used by [MULTI 2.3]{} and [MULTI3D]{} lead to extremely small systematic uncertainties in the barium abundance. These are small ($<0.01$dex) enough to be dwarfed by uncertainties associated with the van der Waals broadening parameters, now discussed. The barium lines synthesised here have varying line strengths. This means that they will react differently to uncertainties associated with the van der Waals line damping parameters, $\sigma$ and $\alpha$. If we conservatively assume that there is a $10\%$ uncertainty associated with the damping parameters tabulated in @Barklem2000 then we can see how this affects the abundance measured in each line and derive a systematic uncertainty for each line. This does not affect the abundance corrections, as these systematics will affect all syntheses equally and so they cancel out, but it is important to determine any uncertainty associated with the abundances determined from these corrections.
We examined how both the cross sections ($\sigma$) and velocity exponent ($\alpha$) affect the line strength by varying both separately and computing new lines using [MULTI 2.3]{}. It was found that varying $\alpha$ by $\pm10\%$ had no impact on the line strengths of any of the lines, and hence we cannot attribute any abundance uncertainty to this parameter. However, a $10\%$ uncertainty in $\sigma$ was found to affect the line strength in all four lines. The 4554Å line is the strongest line measured (207mÅ). It is therefore reasonable to assume that this line will be the most affected by this damping parameter uncertainty. It was found that the barium abundance determined from this line varies by $\pm0.03$dex. The 5853Å line is the weakest line in this study (68mÅ). As such the uncertainty we assign to the van der Waals parameter of $\pm10\%$ leads to a change in abundance of only $0.01$dex. The 6141Å line is the second strongest line measured (126mÅ). This leads to abundance variations of $\pm0.03$dex. Finally, the 6496Å line is also fairly strong in the solar spectrum (102mÅ). The uncertainty we assign the damping parameter varies the barium abundance found in this line by $\pm0.02$dex. The list of associated abundance uncertainties can be seen in column four of Table \[tab:results\].
Uncertainties in line blends are also of concern when computing abundances. No uncertainty information is given by VALD, so we again conservatively assume that the $\log{gf}$ values of these lines have a $10\%$ uncertainty. The abundances attained from the 4554Å line with and without line blending were virtually identical. As previously mentioned, this is because the resonance line dominates line depression around this spectral region. Accordingly, uncertainties in blended lines of $\pm10\%$ do not affect the barium abundance. While the 5853Å line is the weakest line analysed, it suffers the least from blending. As such, there is also no sensitivity in barium abundance found from varying the blended lines. The blends around the 6141Å line have a large impact on the barium abundance. Naturally, uncertainties in $\log{gf}$ lead to an uncertainty of $\pm0.02$dex. Therefore, the inclusion of blend uncertainties increases systematic uncertainty of the 6141Å from $0.03$dex to $0.04$dex. Finally, the blending uncertainties around the 6496Å line were not found to influence the barium abundance. A break down of the associated abundance uncertainties can be seen in column five of Table \[tab:results\].
Oscillator strength uncertainties {#sec:gfuncertainties}
---------------------------------
The oscillator strengths ($f$-values) of the four diagnostic lines used in the present study are taken from @DeMunshi2015 and @Dutta2016. They also provide unique errors associated to each transition probability, which we convert in to oscillator strength uncertainties. The transition probabilities are extremely accurate, so the resulting uncertainties are very small. When these are included in our calculations we find that the propagated abundance error associated with the 4554Å line is $\pm0.00$dex. The weakest line in our sample (5853Å) has a propagated abundance uncertainty of $\pm0.01$dex. The most blended line (6141Å) is found to have a propagated abundance uncertainty of $\pm0.01$dex. Finally, the 6496Å line has an associated abundance error of $\pm0.01$. The associated abundance uncertainties for each line can be found in column six of Table \[tab:results\].
Solar barium abundance {#sec:baabundance}
----------------------
We now present the barium abundance in four paradigms: 3D non-LTE, 3D LTE, 1D non-LTE and 1D LTE. We correct the 1D LTE abundances using the corrections in Table \[tab:corrections\] and Sect. \[sec:blending\], and weight them by their uncertainties listed in Table \[tab:results\] using an inverse-variance weighted mean ($\sum{\omega_i\,X_i}$, where $\omega_i=\frac{1}{\sigma^2}$).
We report for the first time a full 3D non-LTE solar barium abundance of $2.27\pm0.02\pm0.01$, where errors given here are the systematic uncertainties and the random error determined as the standard deviation found in the line-to-line scatter of the 3D non-LTE abundances, which can also be seen in Table \[tab:results\]. This value is $0.08$dex larger than the meteoritic barium abundance of $2.19\pm0.03$ determined by @Lodders2003. We therefore find a photospheric abundance that is slightly larger than that reported from meteorites.
Using the same method just described, we find that ${A({\rm Ba})}=2.31\pm0.02\pm0.03$ in 3D LTE. Errors again represent the systematic and random errors, like above. When we use 1D model atmospheres and apply non-LTE physics to the post-process spectral synthesis we find that the 1D non-LTE barium abundance is ${A({\rm Ba})}=2.11\pm0.02\pm0.05$. Finally, when we derive the barium abundance using the classical 1D LTE approach we find that ${A({\rm Ba})}=2.24\pm0.02\pm0.02$. This abundance is similar to the 3D non-LTE abundance.
Conclusions {#sec:conclusions}
===========
We have computed new values of the solar barium abundance based upon results computed using a new barium model atom that includes quantum mechanical inelastic collisional rate coefficients between and hydrogen. We computed the barium abundance using a static 1D model atmosphere and provide 1D non-LTE, 3D LTE and 3D non-LTE corrections to this value in Table \[tab:corrections\]. Using these corrections we also present new solar photospheric barium abundances for the first time in full 3D non-LTE, as well as in 3D LTE, 1D non-LTE and 1D LTE.
The summary of this work is as follows (NB that all abundances given below are done so by adding the corrections to each 1D LTE abundance and then calculating the inverse-variance weighted mean as described in Sect. \[sec:baabundance\]):
- The 3D non-LTE barium abundance was found to be ${A({\rm Ba})}=2.27\pm0.02\pm0.01$, which is $4\sigma$ larger the meteoritic abundance published by @Lodders2003. This may suggest uncertainties in the atomic data and/or that further physics is still missing from our analyses, such as the treatment of magnetic fields. On the other hand, Ba isotopic abundance anomalies are well-documented in CI meteorites [e.g. @McCulloch1978; @Arnould2007] and it is not clear whether the meteoritic value can be indeed directly compared to the solar photospheric estimate. Nevertheless, this value represents the best photospheric solar barium abundance available from the current state-of-the-art in spectral modelling. As a result, it provides a remarkably consistent abundance for the four diagnostic lines.
- The 3D LTE abundance was determined as ${A({\rm Ba})}=2.31\pm0.02\pm0.03$. This value is larger than the meteoritic value given in @Lodders2003 and the 3D non-LTE abundance we determine, and the individual abundances are not as consistent.
- The 1D non-LTE abundance of ${A({\rm Ba})}=2.11\pm0.02\pm0.05$ suggests that the barium abundance is depleted by $0.16$dex in the solar atmosphere relative to the full 3D non-LTE. The abundance is also smaller than the meteoritic result reported in @Lodders2003 and the inconstencies between lines are larger than they are in 3D non-LTE.
The 3D non-LTE and 1D LTE abundances are very similar for barium in the Sun, but larger than that given in @Lodders2003. Conversely, the 3D LTE abundance suggests that barium abundance is even larger in the Sun, while the 1D non-LTE abundance suggests barium is slightly lower than the meteoritic value. It is clear then that the inclusion or removal of realistic treatments of line formation physics or convection has opposing effects on the barium abundance; the former strengthens the barium lines and so depletes the barium abundance, while the latter weakens the barium lines and hence a larger barium abundance is required. Therefore, the exclusion of both physical processes in the 1D LTE paradigm masks each effect, providing similar values in each line as the actual values found in 3D non-LTE. Conversely, in our work on manganese we found that the 3D and Non-LTE effects do not cancel out, but rather the effects of Non-LTE are amplified in 3D calculations with hydrodynamical models.
The previous set of transition probabilities reported by @Davidson1992 led to abundances values that were, in general, less consistent than those reported here and had larger uncertainties associated to them, with the 5853Å line being most uncertain and most inconsistent with the other three diagnostic lines. The latest published transition probabilities in @DeMunshi2015 and @Dutta2016 represent the most accurate transition parameters published. As such, the barium abundances we find from each diagnostic line used here are all in very good agreement (once the blending corrections are included).
We have presented this work as part of a larger series of papers designed to report on the development of the [MULTI3D]{} spectrum synthesis code. Up until now we have added new coding that allows it to read standard [[stagger]{}]{} model atmospheres; include the effect of charge transfer between hydrogen and ions; compute flux data based on hard-coded quadrature schemes; and compute multi-component transitions caused by HFS or isotope splittings. Further physics and mathematical techniques will be added as the project progresses that will be presented in future papers in this paper series. We also plan to extend our work on barium within this paper series to include several metal-poor benchmark stars, where the present work will be important to the relative abundances we report.
This work made heavy use of the Max Planck Computing & Data Facility (MPCDF) for the majority of the computations. This project was funded in part by Sonderforschungsbereich SFB 881 “The Milky Way System” (subproject A5) of the German Research Foundation (DFG) and by the Research Council of Norway through its Centres of Excellence scheme, project number 262622. SAY and AKB gratefully acknowledge support from the Ministry for Education and Science (Russian Federation), project Nos. 3.5042.2017/6.7, 3.1738.2017/4.6. BP is partially supported by the CNES, Centre National d’Etudes Spatiales.
HFS information {#apdx:hfs}
===============
This following tables tabulate the complete hyperfine structure information of the four barium diagnostic lines used in this study.
[^1]: ${f_{\rm odd}}\equiv\left[ N(\element[][135]{Ba}) + N(\element[][137]{Ba})\right]/N(\element{Ba})$
[^2]: both FAL-C semi-empirical models and a 3D radiative hydrodynamical model from @Asplund2000 were used.
[^3]: ordinarily [MULTI3D]{} uses a short-characteristic solver for 3D model atmospheres.
[^4]: The departure coefficient is defined as the ratio of atomic number density for a given energy level $i$ computed in non-LTE to that of LTE, $b_i = \frac{n_{i,{\rm non-LTE}}}{n_{i,{\rm LTE}}}$.
[^5]: <http://vald.astro.uu.se/>
[^6]: <https://www.as.utexas.edu/~chris/moog.html>
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'We give a group-theoretic interpretation of non-relativistic holography as equivalence between representations of the Schrödinger algebra describing bulk fields and boundary fields. Our main result is the explicit construction of the boundary-to-bulk operators in the framework of representation theory (without specifying any action). Further we show that these operators and the bulk-to-boundary operators are intertwining operators. In analogy to the relativistic case, we show that each bulk field has two boundary fields with conjugated conformal weights. These fields are related by another intertwining operator given by a two-point function on the boundary. Analogously to the relativistic result of Klebanov-Witten we give the conditions when both boundary fields are physical. Finally, we recover in our formalism earlier non-relativistic results for scalar fields by Son and others.'
---
[N. Aizawa$^{a,}$[^1], V.K. Dobrev$^{b,}$[^2] ]{}
*$^a$ Department of Mathematics and Information Sciences,\
Graduate School of Science, Osaka Prefecture University, Nakamozu Campus,\
Sakai, Osaka 599-8531 Japan*
*$^b$Institute of Nuclear Research and Nuclear Energy,\
Bulgarian Academy of Sciences\
72 Tsarigradsko Chaussee, 1784 Sofia, Bulgaria*
Introduction
============
The role of nonrelativistic symmetries in string theory was always important. In fact, being the theory of everything string theory encompasses together relativistic quantum field theory, classical gravity, and certainly, non-relativistic quantum mechanics, in such a way that it is not even necessary to separate these components, cf., e.g., [@MaMaTa; @AoNiSu; @Schvellinger; @AADV].
Thus, it is not a surprise that the Schrödinger group - the group that is the maximal group of symmetry of the Schrödinger equation - is playing more and more a prominent role, cf., e.g., [@NisSon; @Son; @BaMcG; @Goldberger; @MinPle; @DuHaHo; @Taylor; @AMSV; @DonGau; @BagGop; @ColYava; @BoKuPi; @OogPar].
Originally, the Schrödinger group, actually the Schrödinger algebra, was introduced by Niederer [@Nie] and Hagen [@Ha], as a nonrelativistic limit of the vector-field realization of the conformal algebra. In the process, the space components of special conformal transformations decouple from the system. Thus, e.g., in the case of four-dimensional Minkowski space-time from the 15 generators of the conformal algebra we obtain the 12 generators of the Schrödinger algebra.
Recently, Son [@Son] proposed another method of identifying the Schrödinger algebra in d+1 space-time. Namely, Son started from AdS space in d+3 dimensional space-time with metric that is invariant under the corresponding conformal algebra so(d+1,2) and then deformed the AdS metric to reduce the symmetry to the Schrödinger algebra.
In view of the relation of the conformal and Schrödinger algebra there arises the natural question. Is there a nonrelativistic analogue of the AdS/CFT correspondence, in which the conformal symmetry is replaced by Schrödinger symmetry. Indeed, this is to be expected since the Schrödinger equation should play a role both in the bulk and on the boundary. The posed question was studied in some of the literature above, and also in [@BroTer; @BarFue; @HeRaRo; @AdBaMcG; @Yamada; @SakYos; @RSST; @FueMor; @VolWen]. In the present paper, we examine the nonrelativistic analogue of the AdS/CFT correspondence in the framework of representation theory.Before explaining what we do let us remind that the AdS/CFT correspondence has 2 ingredients [@Maldacena; @GuKlPo; @Witten]: 1. the holography principle, which is very old, and means the reconstruction of some objects in the bulk (that may be classical or quantum) from some objects on the boundary; 2. the reconstruction of quantum objects, like 2-point functions on the boundary, from appropriate actions on the bulk. Our main focus is put on the first ingredient and we consider the simplest case of the (3+1)-dimensional bulk. It is shown that the holography principle is established using representation theory only, that is, we do not specify any action. We outline the contents of the paper below. For the implementation of the first ingredient in the Schrödinger algebra context we use a method that is used in the mathematical literature for the construction of discrete series representations of real semisimple Lie groups [@Hotta; @Schmid], and which method was applied in the physics literature first in [@DMPPT] in exactly an AdS/CFT setting, though that term was not used then.
The method utilizes the fact that in the bulk the Casimir operators are not fixed numerically. Thus, when a vector-field realization of the algebra in consideration is substituted in the Casimir it turns into a differential operator. In contrast, the boundary Casimir operators are fixed by the quantum numbers of the fields under consideration. Then the bulk/boundary correspondence forces an eigenvalue equation involving the Casimir differential operator. That eigenvalue equation is used to find the two-point Green function in the bulk which is then used to construct the boundary-to-bulk integral operator. This operator maps a boundary field to a bulk field similarly to what was done in the conformal context by Witten (cf., e.g., formula (2.20) of [@Witten]). This is our first main result. Our second main result is that we show that this operator is an intertwining operator, namely, it intertwines the two representations of the Schrödinger algebra acting in the bulk and on the boundary. This also helps us to establish that each bulk field has actually two bulk-to-boundary limits. The two boundary fields have conjugated conformal weights $\Delta$, $3-\Delta$, and they are related by a boundary two-point function.
We consider also the second ingredient of the AdS/CFT correspondence in the Schrödinger context and show how our formalism involving the Casimir differential operator relates to the case of scalar field theory discussed in [@Son; @BaMcG]. We can easily extend our considerations for the higher-dimensional cases [@AizDob]. Higher dimensional Schrödinger group has the rotation group as a subgroup. Thus our formalism can be naturally extended to the cases with arbitrary spin. The paper is organized as follows. In Section 2 we give the preliminaries on the Schrödinger algebra including the Casimir and the well-known vector-field realization. In Section 3 we make the choice of bulk using the four-dimensional space of Son and write down the vector-field realization in the bulk. In Section 4 we construct the integral boundary-to-bulk operator. In Section 5 we establish the intertwining properties of the boundary-to-bulk and bulk-to-boundary operators. We display also the intertwining relation between the two bulk-to-boundary limits of a bulk field. Finally, in Section 6 we relate our approach to earlier work on non-relativistic holography showing how we can recover those results for $d=1$.
Preliminaries
=============
The Schrödinger algebra $ {\mathfrak s}(d) $ in ($d$+1)-dimensional spacetime is generated by time translation $P_t$, space translation $ P_k$, Galilei boosts $ G_k$, rotations $
J_{k\ell} = -J_{\ell k} $ (which generate the subalgebra $ so(d)$), dilatation $D$ and conformal transformation $K$ ($k, \ell = 1,
\cdots d$). The non-trivial commutation relations are [@BarRac] $$\begin{tabular}{llll}
$[P_t, D] = 2P_t, \ $& $ [P_t, G_k] = P_k, $&$ [P_t, K] = D,$ & $[P_k, D] = P_k, \ $ \\
$ [P_k, K] = G_k, $ & $ [D, G_k] = G_k,$ & $[D, K] = 2K, $ & \\
\multicolumn{2}{l}{$[P_i, J_{k\ell}] = \delta_{i\ell} P_k - \delta_{ik} P_{\ell},$} &
\multicolumn{2}{l}{$[G_i, J_{k\ell}] = \delta_{i\ell} G_k - \delta_{ik} G_{\ell},$} \\
\multicolumn{3}{l}{$[J_{ij}, J_{k\ell}] = \delta_{ik}J_{j\ell}
+ \delta_{j\ell}J_{ik}
-\delta_{i\ell}J_{jk} - \delta_{jk}J_{i\ell},$} &
\end{tabular}
\label{commsn}$$ Actually, we shall work with the central extension $\hat {\mathfrak
s}(d) $ of the Schrödinger algebra obtained by adding the central element $M$ to ${\mathfrak s}(d)$ which enters the additional commutation relations: $
[P_k, G_{\ell}] = \delta_{k\ell} M.
$ In many physical applications the central element $M$ corresponds to mass.
For the purposes of this paper we now restrict to the 1+1 dimensional case. In this case the centrally extended Schrödinger algebra has six generators:
----- --- -------------------------- -- ----- --- -------------------
$H$ : time translation $P$ : space translation
$G$ : Galilei boost $D$ : dilatation
$K$ : conformal transformation $M$ : center
----- --- -------------------------- -- ----- --- -------------------
\
with the following non-vanishing commutation relations:
[lclcl]{} \[H, D\] = 2H, & & \[D, K\] = 2K, & & \[H, K\] = D,\
\[P, G\] = M, & & \[P, K\] = G, & & \[H, G\] = P,\
\[P, D\] = P, & & \[D, G\] = G. & &
\[S1def\]
For our approach we need the Casimir operator. It turns out that the lowest order nontrivial Casimir operator is the 4-th order one [@Perroud]: \_4 = ( 2 M D - { P, G } )\^2 - 2 { 2MK - G\^2, 2MH-P\^2 }. \[Casimir\] In fact, there are many cancellations, and the central generator $M$ is a common linear multiple.[^3]
Choice of bulk and boundary
===========================
We would like to select as bulk space the four-dimensional space $
(x, x_\pm, z) $ obtained by Son [@Son]:[^4] ds\^2 = - + . \[metric\]
We require that the Schrödinger algebra is an isometry of the above metric. Here the variable $z$ is the main variable distinguishing the bulk, namely, the boundary is obtained when $z=0$. We also need to replace the central element $M$ by the derivative of the variable $x_-$ which is chosen so that $\frac{\partial}{\partial x_-}$ continues to be central. Thus, the vector-field realization of the Schrödinger algebra now becomes: & & H = , P = , M = ,\
& & G = x\_+ + x , \[VecF1\]\
& & D = x + z + 2 x\_+ ,\
& & K = x\_+ ( x + z + x\_+ ) + (x\^2 + z\^2) . and it generates an isometry of . This vector-field realization of the Schroedinger algebra acts on the bulk fields $\phi(x_\pm,x,z) $.
In this realization the Casimir becomes: \_4 &=& M\^2 C\_4,\
C\_4 &=& \^2 - 4 - 4 z\^2\
&=& 4 z\^2 \^2 - 8 z + 5 - 4 z\^2 , \[CasReal2\]\
&& 2 - \^2 , \[schr-op\]\
&& 2 z - 1. Note that is the pro-Schrödinger operator.
Next we consider a realization of the Schrödinger algebra on the boundary. Actually, we use a well known such vector-field realization [@BarRac] in which we only modify the expression for $M$: & & H = , P = , M = ,\
& & G = x\_+ + x M , \[VecF0\]\
& & D = x + + 2 x\_+ ,\
& & K = x\_+ ( x + + x\_+ ) + x\^2 M. where $\D$ is the conformal weight. This vector-field realization of the Schroedinger algebra acts on the boundary field $\phi(x_\pm,x) $ with fixed conformal weight $\D$.
In this realization the Casimir becomes: \^0\_4 &=& M\^2 C\^0\_4,\
C\^0\_4 &=& (2 - 1) (2 - 5) \[CasBou\] As expected $ C^0_4 $ is a constant which has the same value if we replace $\D$ by $3-\D$: C\^0\_4 () = C\^0\_4 (3-) \[pequi\]This already means that the two boundary fields with conformal weights $\D$ and $3-\D$ are related, or in mathematical language, that the corresponding representations are (partially) equivalent. This will be very important also below.
Boundary-to-bulk correspondence
===============================
As we explained in the Introduction we first concentrate on one aspect of AdS/CFT [@GuKlPo; @Witten], namely, the holography principle, or boundary-to-bulk correspondence, which means to have an operator which maps a boundary field $\varphi$ to a bulk field $\phi$, cf. [@Witten], also [@Dobrev].[^5] This will be done within the framework of representation theory without specifying any action.
The fields on the boundary are fixed by the value of the conformal weight $\D$, correspondingly, as we saw, the Casimir has the eigenvalue determined by $\D$: C\^0\_4 (x\_,x) = (x\_,x) , = (2 - 1) (2 - 5)
Thus, the first requirement for the corresponding field on the bulk $\phi(x_\pm,x,z) $ is to satisfy the same eigenvalue equation, namely, we require: C\_4 (x\_,x,z) = (x\_,x,z) , = (2 - 1) (2 - 5) \[eigen\]where $C_4$ is the differential operator given in . Thus, in the bulk the eigenvalue condition is a differential equation.
The other condition is the behaviour of the bulk field when we approach the boundary: (x\_,x,z) z\^ (x\_,x) , = ,3- \[BouBeh\]
To find the boundary-to-bulk operator we follow the method of [@DMPPT], namely, we find the two-point Green function in the bulk solving the differential equation: (C\_4 - ) G(, z ; ’, z’) = z[’]{}\^4 \^3(- ’) (z-z’), \[Gdef\] where $ \chi = (x_+, x_-, x). $
As in [@DMPPT] it is important to use an invariant variable which in our case is: u = . \[udef\] In terms of $u$ the Casimir becomes: C\_4 = 4 u\^2 (1-u) - 8 u + 5. \[Casimirinu\]
We can reduce the eigenvalue equation to the equation for the hypergeometric function by the substitution: $ G(\chi, z \,;\,
\chi', z') = G(u) = u^{\alpha} F(u)$. Then the equation becomes: & & (C\_4 - ) G(, z ; ’, z’)\
& & = 4u\^[+1]{} { u(1-u) F” + 2 (- 1 - u) F’ + ( - (-1) ) F } = 0,\
& & \[HG1\] where we ignore for the moment the $\delta$-function - it will be reproduced by the singularity of the solutions at $u=1$. The parameter $\alpha$ is arbitrary, so we fix it by requiring the vanishing of the $ u^{-1} $ term, and we recover the two choices: $\alpha=\D$, $\alpha=3-\D$. Then we have: & & u(1-u) F” + 2 (-1 - u) F’ - (- 1) F = 0, ( = ), \[HG2\]\
& & u(1-u) F” + 2 (2-- (3-)u) F’ - (- 2) (- 3) F = 0, ( = 3-).\
& & \[HG3\] Since the hypergeometric equation has two independent solution, then it turns out (expectedly) that overall for the function $G(u)$ we also have a single set of two solutions: & & G(u) = u\^ F(,-1;2(-1);u), \[G1\]\
& & G(u) = u\^[3-]{} F(3-,2-;2(2-);u). \[G2\] where $F =~ _2F_1$ is the standard hypergeometric function.
As expected at $u=1$ both solutions are singular: by [@BaEr], (\[G1\]) reads: $$G(u) = \frac{ u^{\Delta} }{ 1-u } F(\Delta-2,\Delta-1;2(\Delta-1);u),$$ while (\[G2\]) reads: $$G(u) = \frac{u^{3-\Delta}}{1-u} F(1-\Delta,2-\Delta;2(2-\Delta);u).$$
Following the general method the boundary-to-bulk operator is obtained from the two-point bulk Green function by bringing one of the points to the boundary, however, one has to take into account all info from the field on the boundary. More precisely, in mathematical terms we express the function in the bulk with boundary behaviour through the function on the boundary by the formula: (,z) = d\^3’ S\_(-’,z) (’), \[Bo2Bu\] where $ d^3 \chi' = dx_+{'} dx_-{'} dx' $ and $
S_\a(\chi-\chi',z) $ is defined by S\_(-’,z) = \_[z’ 0]{} z’\^[-]{} G(u) = \^. \[Sdef\] where $\a$ is as in .
Intertwining properties
=======================
An important ingredient of our approach is that the bulk-to-boundary and boundary-to-bulk operators are actually intertwining operators.[^6] To see this we need some more notation.
Let us denote by $L_\a$ the bulk-to-boundary operator : (L\_) () \_[z 0]{}z\^[-]{}(,z), \[Buboug\] where $\a = \D,3-\D$ consistently with . The intertwining property is: L\_= \_L\_, X, \[inta\] where $\tX_\a$ denotes the action of the generator $X$ on the boundary (with $\D$ replaced by $\a$ from ), $\hX$ denotes the action of the generator $X$ in the bulk . Checking is straightforward.
Let us denote by $\tL_\a$ the boundary-to-bulk operator in : (,z) = ( \_) (,z) d\^3’ S\_(-’,z) (’). \[bobuop\]The intertwining property now is: \_\_[3-]{} = \_, X. \[intb\] The checking of requires some work, but is straightforward.
Next we check consistency of the bulk-to-boundary and boundary-to-bulk operators, namely, their consecutive application in both orders should be the identity map: L\_[3-]{} \_ &=& [**1**]{}\_[boundary]{}, \[boubu\]\
\_ L\_[3-]{} &=& [**1**]{}\_[bulk]{}. \[bubou\]
Checking means: \[invc\] (L\_[3-]{} \_) () &=& \_[z 0]{} z\^[-3]{} (\_) (,z)\
&=&\_[z 0]{} z\^[-3]{} d\^3’ S\_(-’,z) (’)\
&=& \_[z 0]{} z\^[-3]{} d\^3’ ( )\^ (’) ,\
&& A = (x-x’)\^2 -2 (x\_+-x’\_+)(x\_–x’\_-) + z\^2.
For the above calculation we interchange the limit and the integration, and use the following formula: \[delf\] \_[z 0]{} z\^[-3]{} ( )\^ = 2\^[2]{} \^[3/2]{} \^3(-’) , - 3/2 \_- The Proof of is given in the Appendix.
Using we obtain: (L\_[3-]{} \_) () = 2\^[2]{} \^[3/2]{} (). Thus, in order to obtain exactly, we have to normalize, e.g., $\tL_{\a}$.
We note the excluded values $\a - 3/2 \notin \bbz_-$ for which the two intertwining operators are not inverse to each other. This means that at least one of the representations is reducible. This reducibility was established [@DoDoMr] for the associated Verma modules with lowest weight determined by the conformal weight $\D$.[^7]
Checking is now straightforward, but also fails for the excluded values.
Note that checking we used for $\a \to\
3-\a$, i.e., we used one possible limit of the bulk field . But it is important to note that this bulk field has also the boundary as given in . Namely, we can consider the field: \_0 () (L\_ ) () = \_[z 0]{}z\^[-]{}(,z), \[Bubougg\] where $\phi(\chi,z)$ is given by . We obtain immediately: \_0 () = d\^3’ G\_(-’) (’), \[Bo2Bu2\] where G\_() = \^. \[Gdef2\]If we denote by $G_\a$ the operator in then we have the intertwining property: \_G\_= G\_\_[3-]{} . \[intc\] Thus, the two boundary fields corresponding to the two limits of the bulk field are equivalent (partially equivalent for $\a \in \bbz + 3/2$). The intertwining kernel has the properties of the conformal two-point function [@Dobrev].
Thus, for generic $\D$ the bulk fields obtained for the two values of $\a$ are not only equivalent - they coincide, since both have the two fields $\varphi_0$ and $\varphi$ as boundaries.
[**Remark:**]{} For the relativistic AdS/CFT correspondence the above analysis relating the two fields in was given in [@Dobrev]. An alternative treatment relating these two fields via the Legendre transform was given in [@KleWit].
As in the relativistic case there is a range of dimensions when both fields $\D,3-\D$ are physical: \[phys\] \_-\^0 1/2 < < 5/2 \_+\^0 .The above bounds are determined by the values at which the Casimir eigenvalue $\lambda =
(2\D-1)(2\D-5)$ becomes zero.[^8]
Nonrelativistic reduction
=========================
In order to connect our approach with that of previous works [@Son; @BaMcG; @FueMor], we consider the action for a scalar field in the background (\[metric\]): I() = -d\^3dz ( \^ \^\* + m\_0\^2 ||\^2). \[action1\] By integrating by parts, and taking into account a non-trivial contribution from the boundary, one can see that $ I(\phi) $ has the following expression: I() = d\^3dz \^\* (- m\_0\^2) - \_[z 0]{} d\^3 \^\* z . \[action2\] The second term is evaluated using (\[Bo2Bu\]). For $ z \rightarrow 0, $ one has z \~ (4z)\^ d\^3’ + O(z\^[+2]{}). \[zdphi-bou\] It follows that & & \_[z 0]{} d\^3 \^\* z = \_[z 0]{} d\^3d\^3 ’ z\^[-3]{} \^\*(,z) ( )\^ (’)\
& & = 4\^ d\^3d\^3 ’ . \[action3\]
The equation of motion being read off from the first term in (\[action2\]) can be expressed in terms of the differential operator (\[CasReal2\]): (- m\_0\^2) = ( + 2 \^2 - m\_0\^2 ) = 0. \[EoM\] The fields in the bulk (\[Bo2Bu\]) do not solve the equation of motion. Now we set an Ansatz for the fields on the boundary: $
\varphi(\chi) = e^{M x_-} \varphi(x_+,x) $ and compactify the $x_-$ coordinate: $ x_- + a \sim x_-. $ This leads to a separation of variables for the fields in the bulk in the following way: $$\phi(\chi,z) =
e^{Mx_-}
\int dx'_+ dx' \int_0^a d\xi
\left(
\frac{4z}{ (x-x')^2 - 2(x_+-x'_+) \xi + z^2 }
\right)^{\alpha} e^{-M\xi} \varphi(x'_+,x').$$ Thus we are allowed to make the identification $ \del{-} = M $ both in the bulk and on the boundary [@Son; @BaMcG]. We remark that under this identification the operator (\[schr-op\]) becomes the Schrödinger operator. Integration over $ \xi $ turns out to be incomplete gamma function: & & (,z) = e\^[Mx\_-]{} (x\_+,x,z), \[nonrel-bulk-fun\]\
& & (x\_+,x,z) = (-2z)\^ M\^[-1]{} (1-, Ma)\
& & ( - M ) (x’\_+,x’). \[nonrel-bulk-fun2\] This formula was obtained first in [@FueMor]. The equation of motion (\[EoM\]) now reads ( - m\^2 ) (x\_+,x,z) = 0, where $ m^2 = m_0^2 - 2M^2. $ Requiring $ \phi(x_+,x,z) $ to be a solution to the equation of motion makes the connection between the conformal weight and mass: \_= ( 3 ). \[Delta2mass\] This result is identical to the relativistic AdS/CFT correspondence [@GuKlPo; @Witten]. The action (\[action2\]) evaluated for this classical solutions has the following form ($\a=\D_\pm$): & & I() = -(-2)\^ M\^[-1]{} (1-,Ma)\
& & ( - M ) (x\_+,x)\^\* (x’\_+,x’). \[bou-ansatz\] The two-point function of the operator dual to $ \phi $ computed from (\[bou-ansatz\]) coincides with the result of [@Son; @BaMcG; @Henkel; @StoHen]. We remark that the Ansatz for the boundary fields $ \varphi(\chi) = \exp(Mx_- -\omega x_+ + ikx) $ used in [@Son; @BaMcG] is not necessary to derive (\[bou-ansatz\]).
One can also recover the solutions in [@Son; @BaMcG] rather simply in our group theoretical context. We use again the eigenvalue problem of the differential operator (\[CasReal2\]): C\_4 (x\_+,x,z) = (x\_+,x,z). \[EVprob\] but make separation of variables $ \phi(x_+,x,z) = \psi(x_+,x) f(z). $ Then (\[EVprob\]) is written as follows: $$\frac{1}{f(z)}
\left(
\del{z}^2 - \frac{2}{z} \del{z} + \frac{5-\lambda}{4z^2}
\right)
f(z)
=
\frac{1}{\psi(x_+,x)} \hat{S} \psi(x_+,x) = p^2 \ \mbox{(const)}$$ Schrödinger part is easily solved: $ \psi(x_+,x) = \exp(-\omega x_+ + i k x) $ which gives p\^2 = -2M + k\^2. \[enegy\] The equation for $ f(z) $ now becomes \^2 f(z) - f(z) + ( 2M- k\^2 - ) f(z) = 0. \[eqforz\] This is the equation given in [@Son; @BaMcG] for $ d=1$. Thus solutions to equation (\[eqforz\]) are given by modified Bessel functions: $
f_{\pm}(z) = z^{3/2} K_{\pm \nu}(pz)
$ where $ \nu $ is related to the effective mass $m$ [@Son; @BaMcG]. In our group theoretic approach one can see its relation to the eigenvalue of $C_4: $ $ \nu = \sqrt{\lambda+4}/2. $
We close this section by giving the expression of (\[bou-ansatz\]) for the alternate boundary field $ \varphi_0. $ To this end, we again use the Ansatz $ \varphi(\chi) = e^{M x_-} \varphi(x_+,x) $ for (\[Bo2Bu2\]). Then performing the integration over $ x'_- $ it is immediate to see that: \_0(x,x\_+) \~ e\^[Mx\_-]{} ( - M ) (x’\_+,x’). \[Bo2Bo-NR\] One can invert this relation since $ G_{3-\alpha} \circ G_{\alpha} = 1_{\rm boundary}. $ Substitution of (\[Bo2Bo-NR\]) and its inverse to (\[bou-ansatz\]) gives the following expression: I() \~ ( - M ) \_0(x\_+,x)\^\* \_0(x’\_+,x’). \[bou-ansatz2\]
Acknowledgements {#acknowledgements .unnumbered}
================
We are grateful to M. Asano for valuable discussions. Most of the work on this paper was done during the visit of V.K.D. as Guest Professor at Osaka Prefecture University. V.K.D. is supported in part by Bulgarian NSF grant [*DO 02-257*]{}.
Appendix {#appendix .unnumbered}
========
We here give a proof of . By the definition of gamma function = \_0\^ d \^[-1]{} e\^[-A/2]{}. \[gamma\] Consider the Fourier transformation of (\[gamma\]): = \_0\^ \^[-1]{} e\^[-ipX - A/2]{} , \[step1\] where & & X = x-x’, X\_= x\_- x’\_, pX = p\_x X + p\_+ X\_+ + p\_- X\_-,\
& & d\^3 X = dX dX\_+ dX\_-. Integration over $ X $ is the Guass integral and integration over $ X_+ $ gives a $ \delta$-function: $$(\ref{step1}) = \frac{1}{2^{\alpha}\, \Gamma(\alpha)} \int \int_0^{\infty}
\xi^{\alpha-3/2} \, \delta(-p_+ - i \xi X_-) \, e^{-p_x^2/2\xi - ip_- X_- - \xi z^2/2} \,
d\xi \, dX_-.$$ Applying $ \delta(\lambda x) = |\lambda|^{-1} \delta(x) $ one has: (\[step1\]) &=& \_0\^ \^[-5/2]{} e\^[-(p\_x\^2-2p\_+ p\_-)/2- z\^2/2]{} d\
&=& ( )\^[-3/2]{} K\_[-]{} ( z), where $ \rho = p_x^2 - 2p_+ p_- $ and integral representation of Bessel function was used: \_0\^ d \^[c-1]{} e\^[-(a\^2 + b\^2/)/2]{} = 2 ( )\^c K\_c(ab). Now we make the inverse Fourier transformation to (\[step1\]): = ( )\^[-3/2]{} K\_[-]{} (z) e\^[i pX]{} . It follows that \_[z 0]{} z\^[-3]{} ( )\^ &=& \_[z 0]{} z\^[-3/2]{} ()\^[-3/2]{} K\_[-]{} (z) e\^[i pX]{}\
&=& 2\^[2-3/2]{} e\^[i pX]{}\
&=& 2\^[2]{} \^[3/2]{} \^3(X). where we used $$\lim_{z \rightarrow 0} K_{\nu}(z) \sim \frac{2^{\nu}}{ 2 z^{\nu} } \Gamma(\nu).$$ Thus has been proved.
[99]{}
J. Maldacena, D. Martelli and Y. Tachikawa, JHEP [**0810**]{} (2008) 072; arXiv:0807.1100 \[hep-th\]
H. Aoki, J. Nishimura and Y. Susaki, JHEP [**0904**]{} (2009) 055; arXiv:0810.5234 \[hep-th\].
M. Schvellinger, JHEP [**0812**]{} (2008) 004; arXiv:0810.3011 \[hep-th\].
A. Akhavan, M. Alishahiha, A. Davody and A. Vahedi, JHEP [**0903**]{} (2009) 053; arXiv:0811.3067 \[hep-th\]; Fermions in non-relativistic AdS/CFT correspondence, arXiv:0902.0276 \[hep-th\].
Y. Nishida and D.T. Son, Phys. Rev. D [**76**]{} (2007) 086004; arXiv:0706.3746 \[hep-th\].
D.T. Son, Phys. Rev. [**D78**]{} (2008) 046003; arXiv:0804.3972 \[hep-th\]. K. Balasubramanian and J. McGreevy, Phys. Rev. Lett. [**101**]{} (2008) 061601; arXiv:0804.4053 \[hep-th\].
W.D. Goldberger, JHEP [**0903**]{} (2009) 069; arXiv:0806.2867 \[hep-th\].
D. Minic and M. Pleimling, Phys. Rev. [**E78**]{} (2008) 061108; arXiv:0807.3665 \[cond-mat.stat-mech\]
C. Duval, M. Hassaine and P.A. Horvathy, The geometry of Schr´odinger symmetry in gravity background/non-relativistic CFT, arXiv:0809.3128 \[hep-th\].
M. Taylor, Non-relativistic holography, arXiv:0812.0530 \[hep-th\].
A. Adams, A. Maloney, A. Sinha and S.E. Vazquez, JHEP [**0903**]{} (2009) 097; arXiv:0812.0166 \[hep-th\].
A. Donos and J.P. Gauntlett, JHEP [**0903**]{} (2009) 138; arXiv:0901.0818 \[hep-th\]; Solutions of type IIB and D=11 supergravity with Schrodinger(z) symmetry, arXiv:0905.1098 \[hep-th\].
A. Bagchi and R. Gopakumar, Galilean Conformal Algebras and AdS/CFT, arXiv:0902.1385 \[hep-th\].
E.O. Colgain and H. Yavartanoo, NR $CFT_3$ duals in M-theory, arXiv:0904.0588 \[hep-th\].
N. Bobev, A. Kundu and K. Pilch, Supersymmetric IIB Solutions with Schrödinger Symmetry, arXiv:0905.0673 \[hep-th\].
H. Ooguri and C.-S. Park, Supersymmetric non-relativistic geometries in M-theory, arXiv:0905.1954 \[hep-th\].
U. Niederer, Helv. Phys. Acta [**45**]{} (1972) 802-810.
C.R. Hagen, Phys. Rev. [**D5**]{} (1972) 377-388.
S.J. Brodsky and G.F. de Teramond, Phys. Rev. Lett. [**96**]{} (2006) 201601; hep-ph/0602252.
J.L.F. Barbon and C.A. Fuertes, JHEP [**0809**]{} (2008) 030; arXiv:0806.3244 \[hep-th\].
C.P. Herzog, M. Rangamani and S.F. Ross, JHEP [**0811**]{} (2008) 080; arXiv:0807.1099v3 \[hep-th\].
A. Adams, K. Balasubramanian and J. McGreevy, JHEP [**0811**]{} (2008) 059; arXiv:0807.1111 \[hep-th\].
D. Yamada, Class. Quantum Grav. [**26**]{} (2007) 075006; arXiv:0809.4928 \[hep-th\].
M. Sakaguchi and K. Yoshida, JHEP [**0802**]{} (2008) 092; arXiv:0712.4112 \[hep-th\].
M. Rangamani, S.F. Ross, D.T. Son and E.G. Thompson, JHEP [**0901**]{} (2009) 075; arXiv:0811.2049 \[hep-th\].
C.A. Fuertes and S. Moroz, Correlation functions in the non-relativistic AdS/CFT correspondence, arXiv:0903.1844 \[hep-th\].
A. Volovich and C. Wen, Correlation Functions in Non-Relativistic Holography, arXiv:0903.2455 \[hep-th\].
J. Maldacena, Adv. Theor. Math. Phys. [**2**]{} (1998) 231-252; hep-th/9711200.
S.S. Gubser, I.R. Klebanov and A.M. Polyakov, Phys. Lett. [**428B**]{} (1998) 105-114; hep-th/9802109.
E. Witten, Adv. Theor. Math. Phys. [**2**]{} (1998) 253-291; hep-th/9802150.
R. Hotta, J. Math. Soc. Japan, [**23**]{} (1971) 384-407.
W. Schmid, Rive Univ. Studies, [**56**]{} (1970) 99-108.
V.K. Dobrev, G. Mack, V.B. Petkova, S.G. Petrova and I.T. Todorov, [*Harmonic Analysis on the n-Dimensional Lorentz Group and Its Applications to Conformal Quantum Field Theory*]{}, Lecture Notes in Physics, Vol. 63 (Springer, 1977).
N. Aizawa and V.K. Dobrev, in preparation.
A.O. Barut and R. Raczka, [*Theory of Group Representations and Applications*]{}, (PWN, Warszawa, 1980).
M. Perroud, Helv. Phys. Acta, [**50**]{} (1977) 233-252.
H. Bateman and A. Erdelyi, [*Higher Transcendental Functions*]{}, Vol. 1 (New-York, McGraw-Hill, 1953).
V.K. Dobrev, Nucl. Phys. [**B553**]{} \[PM\] (1999) 559-582; hep-th/9812194.
V.K. Dobrev, H.-D. Doebner and C. Mrugalla, Rep. Math. Phys. [**39**]{} (1997) 201-218.
V.K. Dobrev, H.-D. Doebner and C. Mrugalla, J. Phys. [**A29**]{} (1996) 5909-5918; Mod. Phys. Lett. [**A14**]{} (1999) 1113-1122; N. Aizawa, V.K. Dobrev and H.-D. Doebner, in: ‘Quantum Theory and Symmetries II’, Proceedings of the 2nd QTS Symposium, (Cracow, 2001), (World Sci, Singapore, 2002) pp. 222-227; N. Aizawa, V.K. Dobrev, H.-D. Doebner and S. Stoimenov, Proceedings of the VII International Workshop “Lie Theory and Its Applications in Physics”, (Varna, 2007), eds. H.-D. Doebner et al, (Heron Press, Sofia, 2008) pp. 372-399; V.K. Dobrev and S. Stoimenov, invited talk by VKD at XXVII Intern. Colloquium on Group Theoretical Methods in Physics, Yerevan, August 2008, to appear in the Proceedings.
I.R. Klebanov and E. Witten, Nucl.Phys. B556 (1999) 89-114; hep-th/9905104.
M. Henkel, J. Stat. Phys. [**75**]{} (1994) 1023-1061.
S. Stoimenov and M. Henkel, J. Phys. Conf. Ser. [**40**]{} (2006) 144; math-ph/0512025.
[^1]: [email protected]
[^2]: [email protected]
[^3]: This is seen immediately by setting $M=0$, then $\tilde{C}_4\to 0$.
[^4]: In the general setting of [@Son] the space is $(d+3)$-dimensional.
[^5]: Mathematically, this means the following. We treat both the boundary fields and the bulk fields as representation spaces of the Schrödinger algebra. The action of the Schrödinger algebra in the boundary, resp. bulk, representation spaces is given by formulae , resp. by formulae . The boundary-to-bulk operator maps the boundary representation space to the bulk representation space.
[^6]: For the relativistic AdS/CFT case this was done in [@Dobrev].
[^7]: For more information on the representation theory and related hierarchies of invariant differential operators and equations, cf. [@ADDMS].
[^8]: Since the Casimir is fixed up to additive and multiplicative constants, the latter statement becomes unambiguous by the requirement that $\D_-^0 = 3-\D_+^0\,$.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract:
- 'Two channels are equivalent if their maximum likelihood (ML) decoders coincide for every code. We show that this equivalence relation partitions the space of channels into a generalized hyperplane arrangement. With this, we define a coding distance between channels in terms of their ML-decoders which is meaningful from the decoding point of view, in the sense that the closer two channels are, the larger is the probability of them sharing the same ML-decoder. We give explicit formulas for these probabilities.'
- 'Rafael G.L. D’Oliveira was supported by CAPES. Marcelo Firer was partially supported by São Paulo Research Foundation, (FAPESP grant 2013/25977-7).'
author:
- 'Rafael G. L. D’Oliveira'
- Marcelo Firer
date: 'Received: date / Accepted: date'
title: 'A Distance Between Channels: the average error of mismatched channels'
---
[example.eps]{} gsave newpath 20 20 moveto 20 220 lineto 220 220 lineto 220 20 lineto closepath 2 setlinewidth gsave .4 setgray fill grestore stroke grestore
Introduction
============
A communication channel cannot generally be chosen in application, but is rather considered to be a “fact of life”. The most that is possible is to make measurements to characterize the type of noise and to have a model for errors.
Some channels are simpler to be handled than others. If a channel is metrizable, for example, one can use methods from classical coding theory which make use of the metric[^1]. If, furthermore, the metric is translation invariant one can use syndrome decoding which greatly reduces decoding complexity.
The question of being metrizable is one that underlines many aspects in coding theory, but it is seldom stated in an explicit way, so we do it here: a channel with equal input and output set of messages is said to be *metrizable* if there is a metric such that, for any code, using maximum likelihood or minimum distance leads to the same decoding decisions[^2].
A metric structure is just one kind of structure that makes a channel more manageable, and sometimes it may be worth to consider an alternative channel model which is less accurate as a model for noise and errors but is simpler to manage in some sense, like for example with the existence of efficient decoding algorithms. In this sense, the long term goal is to develop an approximation-theory-like approach to coding theory. To do so, the first step would be to determine a distance in the space of channels which translates the probabilistic structure of channels. This is the main goal of this work.
This paper is organized in the following way:
In Section \[sec:related work\] we discuss the connections between the work presented here with previous work on mismatched decoding and the partial ordering of channels.
In Section \[sec:pre\] we establishing the notation and basic definitions used in this work. Among these is the notion of *decoding equivalence* (Definition \[def:decequ\]), a natural equivalence relation between channels. When speaking of the “space of channels” we mean the set of channels under decoding equivalence.
In Section \[sec:decequirn\] we show that the space of channels has a structure of a special kind of hyperplane arrangement known as the braid arrangement (Theorems \[teo:decequibra\] and \[teo:braidcha\]).
In the literature, a braid arrangement has a natural distance (the Kendall tau distance), but this does not attend our requirements.
In Sections \[sec:disper\] and \[sec:discha\] we present a modified version of the Kendall tau distance which is an appropriate measure from the decoding point of view: channels which are closer with respect to this distance are more probable to perform the same maximum likelihood decoding when considering arbitrary random codes (our main result, Corrolary \[cor:disrad\]).
Related Work {#sec:related work}
============
The study of the space of channels at its own sake is related (although not equivalent) to other subjects that has been studied, namely mismatched decoding and partial ordering of channels. We give a brief view of these topics, pointing the similarities and differences with our approach.
Mismatched Decoding
-------------------
Our approximation-theory-like approach is similar to the setting of *mismatched decoding*. In this setting instead of using the Maximum Likelihood decoder determined by the channel $P$ ($ML_{P}$-decoding), a different decoding criterion is used. In practice, this might occur due to inaccuracies in the measurement of a channel. In this case we are using an $ML_{Q}$-decoding, where $Q$ is the non-accurate measured channel. Another reason for mismatched decoding arises when there are no reasonable algorithms for implementing $ML_{P}$-decoding.
Mismatched decoding has an extensive literature ([@gant00] has many relevant references on the subject). The approach, however, is essentially information theoretical, guided by the fundamental question of determining *what can, in principle, be done*. This means that most of the work in the area aims to understand what is achievable asymptotically, for example, what are the achievable rates for families of channels with the input-output sets’ size going to infinity. Those are very difficult questions and hence a significant part of the effort is directed to find bounds for those rates (and other significant invariants).
Our approach is less concerned with the asymptotic aspects of achievability. Once the input and output sets $X$ and $Y$ are given (and fixed) and supposing that the actual channel is $P$, how much are we expected to loose once we decode a randomly chosen code using the ML-decoding criterion determined by a different channel $Q$. Our measure of expected loss is the overall probability of error in the whole process of encoding, transmitting, and decoding. In this sense, we may say that we are considering the mismatched decoding problem in the finite block length regime.
Partial Ordering of Channels
----------------------------
Our approach to study the geometry of the space of channels has an intersection with the concept of *channel inclusion*, as introduced by Shannon [@shan58] and as presented, for example, by Makur and Polyanskiy [@maku16]. Using the notation of Makur and Polyanskiy, given two channels with transition matrices $P$ and $Q$ of size $N_{P}\times M_{P}$ and $N_{Q}\times M_{Q}$ respectively, with $N_{Q}\leq N_{P}$ and $M_{Q}\leq M_{P}$, one says that $P$ *includes* $Q$ if there are two families $\left( A_{k}\right)
_{k=1}^{m}$ and $\left( B_{k}\right) _{k=1}^{m}$ of channels (with $A_{k}$ being an $M_{P}\times M_{Q}$ transition matrix and $A_{K}$ an $N_{Q}\times
N_{P}$ transition matrix) and probability mass function $g$ over the set $\left\{ 1,2,...,m\right\} $ such that $$Q=\sum_{k=1}^{m}g\left( k\right) B_{k}PA_{k}\text{.}$$
This concepts embraces many different situations, some of which can be understood with our definition of the space of channels with the decoding equivalence. For example, the first example introduced in Figure 1 of Shannon’s work corresponds to the situation where $m=1$ and both $A_{1}$ and $B_{1}$ is determined by a projection matrix. If we allow $A_{1}$ and $B_{1}$ to correspond to a projection or a permutation matrix (or a combination of both), we actually have a hyperplane, $\mathcal{P}_{\pi}\subset\mathbb{R}^{N\times M}$, with a braid arrangemente structure (see Section \[sec:decequirn\]) induced from the braid arrangement structure of $\mathbb{R}^{N\times M}$, by considering the intersection of a deconding cone $cone\left( x\right) \cap\mathcal{P}$.
Preliminaries {#sec:pre}
=============
In this section we start with a list of definitions and notations used throughout this work. Since these concepts are well known we present them very succinctly, citing references for details. After that, in \[subsec:space\_of\_chan\], we present the decoding equivalence between channels introduced in [@doli16a] and define the space of channels.
Notation {#cones}
--------
We consider the basic setting of information theory [@shan48] where a transmitter sends a message to a receiver passing through a channel. Let $X = \{x_1, x_2, \ldots, x_n\}$ be the set of input messages which the transmitter can send and let $Y = \{y_1, y_2, \ldots, y_m\}$ be the set of output messages which the receiver can receive. It is common for the messages to come from some alphabet in which case the sets $X$ and $Y$ are exponential on the block length with respect to the size of the alphabet.
A *channel* is a $n \times m$ probabilistic matrix $P$ such that $P_{ij} = Pr (y_j \hspace{2pt} \text{received} \hspace{3pt} \vert x_i \hspace{2pt} \text{sent})$, the probability of receiving $y_j$ given that $x_i$ was sent (the rows sum to $1$).
Given a code $\emptyset \neq C \subseteq X$, a *maximum likelihood decoder* is such that $y \in Y$ is decoded as some $c \in C$ which maximizes $ Pr (y \hspace{2pt} \text{received} \hspace{3pt} \vert c \hspace{2pt} \text{sent})$. The set of maximum likelihood (ML) decoders of the channel $P$ for a code $C$ is denoted by ${D\widehat{e}c}_C (P)$.
A *weak order* over a set $X$ is a triple $(X,\prec,\simeq)$, where $\prec$ and $\simeq$ are binary relations on $X$ satisfying, for all $x,y,z \in X$:
1. $x \prec y$ and $y \prec z$ implies that $x \prec z$,
2. $\simeq$ is an equivalence relation,
3. exactly one of $x \prec y$, $y \prec x$ or $x \simeq y$ holds.
We denote the set of all weak orders over $n$ objects by $W_n$.
We denote the symmetric group over $n$ objects by $S_n$. As usually done, we use lowercase Greek letters for elements of this set ($\sigma, \tau, \phi \in S_n$).
A set $ A \subseteq \mathbb{R}^n$ is *convex* if it contains the segment joining any two of its points, i.e. $\alpha x + (1-\alpha)y \in A$ for every $x,y \in A$ and $0 \leq \alpha \leq 1$.
A *hyperplane* is a set $H \subseteq \mathbb{R}^n$ of the form $H = \{ x \in \mathbb{R}^n : \alpha \cdot x = a \}$ where $0 \neq \alpha \in \mathbb{R}^n$, $a \in \mathbb{R}$ and $\alpha \cdot x := \sum_{i=1}^n \alpha_i x_i $ is the usual dot product.
A *hyperplane arrangement* $\mathcal{A}$ (see [@stan04] for details) is a set of hyperplanes. A region of an arrangement is a connected component of the complement of the hyperplanes, $X = \mathbb{R}^n - \bigcup\limits_{H \in \mathcal{A}} H$. The set of regions is denoted by $\mathcal{R}(\mathcal{A})$ and $r(\mathcal{A}) := \# \mathcal{R}(\mathcal{A})$.
Each hyperplane divides $\mathbb{R}^n$ into two subsets known as *half-spaces*. The two half spaces corresponding to $H = \{ x \in \mathbb{R}^n : \alpha \cdot x = a \}$ are $\{ x \in \mathbb{R}^n : \alpha \cdot x \leq a \}$ and $\{ x \in \mathbb{R}^n : \alpha \cdot x \geq a \}$.
A *convex polytope* is the intersection of a finite set of half-spaces which is bounded.
A set $C \subseteq \mathbb{R}^n$ is a *convex cone* if $\alpha x + \beta y \in C$ for every $x,y \in C$ and $\alpha,\beta \geq 0$.
We are particularly interested in the *braid arrangement,* $\mathcal{B}_n$, which consists of the $\binom{n}{2}$ hyperplanes: $x_i - x_j = 0$ for $1 \leq i < j \leq n$. Specifying to which side of the hyperplane a point $a \in \mathbb{R}^n$ belongs to is equivalent to determining whether $a_i < a_j$ or $a_j < a_i$. Doing so for every hyperplane is equivalent to imposing a linear order on the $a_i$. So to each permutation $\sigma \in S_n$ there corresponds a region $R_{\sigma} \in \mathcal{R}(\mathcal{B}_n)$ given by $R_{\sigma} = \{ x \in \mathbb{R}^n : a_{\sigma (1)} < a_{\sigma (2)} <\ldots < a_{\sigma(n)} \}$. Thus, $r(\mathcal{B}_n) = n!$.
The Iverson bracket will be used in our definitions and proofs: for a statement $S$, the bracket $[S]$ equals $1$ if statement $S$ is true and equals $0$ otherwise.
The Space of Channels {#subsec:space_of_chan}
---------------------
The results in this section appear in more detail in [@doli16a],[@doli16d] and [@doli16b].
Consider the space $\mathbb{R}_{\geq 0}^{n\times m}$ of matrices with non-negative entries. The space of all $n\times m$ channels, ${Ch}_{n\times m}$, is a subset of this space.
\[def:decequ\] Two channels $P,Q \in {Ch}_{n\times m}$ are decoding equivalent, $P \sim Q$, if, for any code $C \subseteq X$, they have the same maximum likelihood decoders, i.e. for every $C\subseteq X $, ${D\widehat{e}c}_C (P) = {D\widehat{e}c}_C (Q)$.
Our next definition will help characterize decoding equivalence.
Given a matrix $M \in \mathbb{R}_{\geq 0}^{n\times m}$, its *weak order matrix* is the matrix $O^-M$ such that ${(O^-M)}_{ij} = k$ if $M_{ij}$ is the $k$-th largest element (allowing ties) in the $j$-th column of $M$.
If $ M =
\begin{pmatrix}
9 & 2 & 1 \\
9 & 7 & 0 \\
8 & 6 & 8
\end{pmatrix}
$, then $O^- M =
\begin{pmatrix}
1 & 3 & 2 \\
1 & 1 & 3 \\
2 & 2 & 1
\end{pmatrix}
$.
Two channels $P,Q \in {Ch}_{n\times m}$ are decoding equivalent if and only if $O^-P = O^-Q$.
Corollary $3$ in [@doli16a].
With this the decoding equivalence can be extended to the whole of $\mathbb{R}_{\geq 0}^{n\times m}$ by defining $M \sim N$ if $O^-M = O^-N$.
The decoding equivalence partitions $\mathbb{R}_{\geq 0}^{n\times m}$ into $\vert W_n \vert^m$ cones, $(n!)^m$ of which are full dimensional. We denote the decoding cone containing a matrix $M$ by $Cone(M)$ and note that they are the fibers of $O^-$, i.e. $Cone = (O^-)^{-1}\circ O^-$. For details see [@doli16b Section 3].
As we shall see, the space of channels has a structure of hyperplane arrangements and the simplicial structure of hyperplane arrangements reflects the structure of maximum likelihood decoding.
Decoding Equivalence and the Braid Arrangement {#sec:decequirn}
==============================================
Maximum likelihood decoding is done comparing entries of a column of a matrix, the column corresponding to the received message. Considering a column as a vector $x\in\mathbb{R}^n_{\geq 0}$, we show that the decoding equivalence partitions $\mathbb{R}^n_{\geq 0}$ into generalized regions of the braid arrangement. We then extend this result to $\mathbb{R}^{n \times m}_{\geq 0}$.
We first define the $Order$ function.
\[def:orderfunction\] The $Order$ function, $Order:\mathbb{R}^n_{\geq 0} \rightarrow W_n$, takes a vector $x \in \mathbb{R}^n_{\geq 0}$ to the weak ordering of its coordinates.
So, for example, $Order (\sqrt{2},\frac{-1}{2},\sqrt{2}) = Order (2,1,2) = (2 \prec 1 \simeq 3)$.
\[pro:decequiord\] Two vectors $x,y \in \mathbb{R}_{> 0}^{n}$ are decoding equivalent if and only if $Order (x) = Order (y)$.
This follows because $Order (x) = Order (y)$ if and only if $O^- x = O^- y$.
The *fibers* of the $Order$ function, i.e., the inverse images $Order^{-1}(y)$ , partition $\mathbb{R}^n$ into the decoding equivalence class.
The cone function is given by $Cone : \mathbb{R}^n_{\geq 0} \rightarrow 2^{\mathbb{R}^n_{\geq 0}}$ such that $Cone (x) = ({Order})^{-1} \circ Order$. We call $Cone(x)$ the *decoding cone of* $x$.
It is clear, from the definition, that two channels in the same decoding cone determine the same maximum likelihood criteria, for every code.
We generalize the definition of the region of a hyperplane arrangement.
\[def:genreg\] A generalized region of a hyperplane arrangement $\mathcal{A}$ is a connected component of $
\bigcap\limits_{H \in \mathcal{A}_1} H - \bigcup\limits_{H \in \mathcal{A}_2} H$, where $\mathcal{A}_1,\mathcal{A}_2$ is a disjoint partition of $\mathcal{A}$. We denote the sets of generalized regions by $\mathcal{GR}(\mathcal{A})$ and $gr(\mathcal{A}) = \# \mathcal{GR}(\mathcal{A})$.
As stated in Section \[cones\], the braid arrangement consists of the $\binom{n}{2}$ hyperplanes: $H_{ij} = \{ x \in \mathbb{R}^n : x_i=x_j \}$ for $1 \leq i <j \leq n$. The next theorem shows that the decoding equivalence partitions $\mathbb{R}^n$ into generalized regions of the braid arrangement.
\[teo:decequibra\] Let $x,y \in \mathbb{R}^n_{\geq 0}$. Then, $x$ is decoding equivalent to $y$ if and only if $x,y \in R$ for some $R \in \mathcal{GR}(\mathcal{B}_n)$, where $\mathcal{B}_n$ is the braid arrangement.
Specifying to which generalized region $R_x \in \mathcal{GR}(\mathcal{B}_n)$ a point $x \in \mathbb{R}^n$ belongs to is equivalent to determining whether $x_i < x_j$, $x_i = x_j$ or $x_i > x_j$ for every $1 \leq i <j \leq n$. This is equivalent to imposing a weak order on the coordinates of $x$. But this implies that $y \in R_x$ if and only if $Order(y) = Order(x)$. The result then follows from Proposition \[pro:decequiord\].
In other words, if $R \in \mathcal{GR}(\mathcal{B}_n)$ then $x \in R$ if and only if $R = Cone(x)$, i.e. the decoding cones are the generalized regions of the braid arrangement.
\[fig:cone\] 
We now extend the results from $\mathbb{R}^{n}$ to $\mathbb{R}^{n \times m}$.
\[def:ordmat\] The $Order$ function, $Order:\mathbb{R}^{n \times m}_{\geq 0} \rightarrow W_n^m$, is defined as $$Order (M) = Order (M[\cdot][1]) \times Order (M[\cdot][2]) \times ... \times Order (M[\cdot][m]) ,$$ where $Order (M[\cdot][j])$ is the the order function in Definition \[def:orderfunction\] applied to the $j$-th column of $M$. The decoding cone of $M$ is $Cone(M) = {Order}^{-1} \circ {Order} (M)$.
The following result is an analog of Theorem \[teo:decequibra\].
\[teo:braidcha\] Let $M,M' \in \mathbb{R}^{n \times m}$. Then, $M$ is decoding equivalent to $M'$ if and only if $M[\cdot][j],M'[\cdot][j] \in R_j$ for some $R_j \in \mathcal{GR}(\mathcal{B}_n)$, where $\mathcal{B}_n$ is the braid arrangement.
The proof is equivalent to that of Theorem \[teo:decequibra\] by using Definition \[def:ordmat\].
A Decoding Distance Between Permutations {#sec:disper}
========================================
Having an appropriate model of the transmission channel is not always good enough to establish all the necessities in the communication process. Many other questions, such as the complexity of the decoding algorithms, need to be taken into consideration. For this reason, for example, the Hamming metric is many times used, even when the channel is not the binary symmetric channel.
In this sense, it may be interesting to develop an “approximation theory” for channels. The idea is that we can use an approximate simpler channel (or distance matched to it) in place of the original one.
The most basic and mandatory tool for the development of an approximation theory is a distance in the space $Cha_{n\times m}$ which is adequate in some sense. If $P$ is a channel, $P_y$ denotes the column corresponding to $y$ being a received message. We will propose a relevant distance on $Cha_{n\times m}$ which relates to the following:
*Let $P , Q \in {Cha}_{n\times m}$ be two different channels and suppose we know what output $y \in Y$ is received. Choosing a code $C \subseteq X$ from the set of all codes with uniform distribution, what is the probability that ${D\widehat{e}c}_C (P_y) \cap {D\widehat{e}c}_C (Q_y) \neq \emptyset$?*
When we say that the distance is related to that question it means that the probability that ${D\widehat{e}c}_C (P_y) \cap {D\widehat{e}c}_C (Q_y) \neq \emptyset$ decreases with the purposed distance: the closer the channels are, the more probable they are to determine the same decoders.
Since we know what output $y$ is received, only its corresponding column matters for decoding. Thus we are dealing with the decoding equivalence in $\mathbb{R}^n$.
We will only consider the cases for which $Cone(P_y)$ and $Cone(Q_y)$ are $n$-dimensional (and leave the general case for future work). We say that a channel $P$ such that $Cone(P)$ is full dimensional is a *stable channel*, since small perturbations of the channel probabilities do not affect the decoding decisions. In this case ${D\widehat{e}c}_C (P_y) \cap {D\widehat{e}c}_C (Q_y) \neq \emptyset$ is equivalent to ${D\widehat{e}c}_C (P_y) = {D\widehat{e}c}_C (Q_y)$.
By Theorem \[teo:decequibra\], each $n$-dimensional decoding cone corresponds to a region of the braid arrangement $\mathcal{B}_n$. As noted in Section \[cones\] to each $\sigma \in S_n$ there corresponds a region $R_\sigma \in \mathcal{R}(\mathcal{B}_n)$. We can therefore identify every $n$-dimensional decoding cone with a permutation in $S_n$.
Consider $\mathbb{R}_{\geq 0}^3$. The identity element $1 \in S_3$ corresponds to the cone with ordering $(1 \prec 2 \prec 3)$. The transposition $(13)\in S_3$ corresponds to the cone with ordering $(3 \prec 2 \prec 1)$.
Since decoding depends exclusively on the decoding cone, we can extend the definition of ${D\widehat{e}c}_C$ to permutations in the following way.
Let $\sigma \in S_n$, $R_{\sigma} \in \mathcal{R}(\mathcal{B}_n)$ its corresponding decoding cone and $P \in {Cha}_{n\times m}$ such that $P \in R_{\sigma}$. We define ${D\widehat{e}c}_C (\sigma) = {D\widehat{e}c}_C (P)$ for every $C \subseteq X$.
The leading question we posed in the beginning of this section can now be restated in terms of permutation groups as follows:
Given two permutations $\sigma, \phi \in S_n$, what is the probability that ${D\widehat{e}c}_C (\sigma) = {D\widehat{e}c}_C (\phi)$ if $C \subseteq X$ is chosen with uniform distribution?
More precisely, we are interested in computing the following distance:
\[def:decdist\] The decoding distance between two permutations $\sigma, \phi \in S_n$ is $$d_{dec} (\sigma, \phi) = 1 - Pr( \hspace{3pt}{D\widehat{e}c}_C (\sigma) = {D\widehat{e}c}_C (\phi)\hspace{3pt} ),$$ where a code $C\subseteq X$ is chosen randomly, with uniform distribution.
We will solve this by elementary counting.
Let $\sigma, \phi \in S_n$. We denote by $S(\sigma, \phi)$ the number of codes $C$ for which ${D\widehat{e}c}_C (\sigma) = {D\widehat{e}c}_C (\phi)$.
We aim to relate $S(\sigma, \phi)$ to $ d_{dec} (\sigma, \phi)$. We first remark that $\sigma, \phi \in S_n$ is invariant by permutations.
Let $\sigma, \phi, \tau \in S_n$. Then, $S(\tau \circ \sigma, \tau \circ \phi) = S(\sigma, \phi)$.
This follows from the fact that if you permute the rows of a channel, the same permutation on a maximum likelihood decoder of it will yield a maximum likelihood decoder of the permuted channel.
Thus, we can define $S(\sigma) = S(1,\sigma)$ and then $S(\sigma, \phi) = S(\phi^{-1} \circ \sigma)$.
We now show how to compute this function.
\[f\_i\] Let $\sigma \in S_n$ and let us define $f_i(\sigma) = \sum_{j=i+1}^n [ \sigma^{-1} (i) \leq \sigma^{-1} (j) ]$. Then, $$S(\sigma) = \sum_{i=1}^n 2^{f_i(\sigma)}.$$
We want to count how many codes such that ${D\widehat{e}c}_C (1) = {D\widehat{e}c}_C (\sigma)$. The identity element represents $(1\prec 2\prec \ldots \prec n)$ and $\sigma$ represents $(\sigma^{-1}(1)\prec \sigma^{-1}(2)\prec \ldots \prec \sigma^{-1}(n))$.
Recall that $\sigma$ corresponds to a channel $P$ with input messages $X= \{ x_1, \ldots, x_n \}$. Consider codes $C$ such that $x_1 \in C$. The identity element $1$ will decode any one of these as $x_1$. Thus ${D\widehat{e}c}_C (1) = {D\widehat{e}c}_C (\sigma)$ if and only if $\sigma$ also decodes as $x_1$. For this to happen, $C$ can only contain elements $x_i$ such that $\sigma^{-1} (i) \leq \sigma^{-1} (j)$. But $f_1(\sigma)$ counts precisely how many of these exist. So the total number of codes satisfying $x_1 \in C$ and ${D\widehat{e}c}_C (1) = {D\widehat{e}c}_C (\sigma)$ is $2^{f_1(\sigma)}$.
Now consider codes $C$ such that $x_1 \notin C$ and $x_2 \in C$. The same reasoning yields the total number of codes satisfying $x_1 \notin C$, $x_2 \in C$ and ${D\widehat{e}c}_C (1) = {D\widehat{e}c}_C (\sigma)$ as $2^{f_2(\sigma)}$.
Continuing with the same argument yields our result.
The next theorem answers the question posed in the beginning of this section.
Let $\sigma, \phi \in S_n$. If a code $C \subseteq X$ is picked uniformly distributed from the space of all codes, then $Pr({D\widehat{e}c}_C (\sigma) = {D\widehat{e}c}_C (\phi)) = \frac{S(\phi^{-1} \circ \sigma)}{2^n - 1}$.
By definition, $S(\phi^{-1} \circ \sigma)$ counts the number of codes such that ${D\widehat{e}c}_C (\sigma) = {D\widehat{e}c}_C (\phi)$. Elementary probability says we must divide this by the total number of codes.
As a direct corollary we compute the decoding distance.
\[cor:decodingdist\] The decoding distance between two permutations $\sigma, \phi \in S_n$ is $$d_{dec} (\sigma, \phi) = 1 - \frac{S(\phi^{-1} \circ \sigma)}{2^n - 1}$$
In the context of the braid arrangement there exists already a natural distance between permutations. It is known as the Kendall tau distance [@kend38], which we denote by $d_{\tau} (\sigma, \phi)$, and is defined as the minimum number of adjacent permutations $\tau_1, \tau_2, \ldots, \tau_{d_{\tau} (\sigma, \phi)}$ so that $\phi = \sigma \circ \tau_1 \circ \tau_2 \circ \ldots \circ \tau_{d_{\tau} (\sigma, \phi)}$.
Consider the graph whose vertices are the regions of the braid arrangement and such that two vertices share an edge if their corresponding regions are adjacent to each other (so that each edge corresponds to a hyperplane). Then, the Kendall tau distance is the shortest path distance of the graph.
In technical terms: if $\sigma, \tau \in S_n$ where $\tau = (r,r+1)$ and $1$ is the identity in $S_n$, then $$d_{\tau} (1, \tau \circ \sigma) - d_{\tau} (1, \sigma) = \left\{\begin{matrix}
1 & \hspace{5pt} if \hspace{5pt} \sigma^{-1} (r) < \sigma^{-1} (r+1) \\
-1 & \hspace{5pt} if \hspace{5pt} \sigma^{-1} (r) > \sigma^{-1} (r+1)
\end{matrix}\right.$$
We now show that the decoding distance behaves as a weighted version of the Kendall tau distance. We remark that the function $f_i(\sigma) = \sum_{j=i+1}^n [ \sigma^{-1} (i) \leq \sigma^{-1} (j) ]$ used in Theorem \[f\_i\] to describe the function $S( \sigma )$ differs from the Kendall tau distance, since it considers not only the number of transpositions $( r,r+1) $ but also the value of $r$: $f_r(\tau_r )=n-r-1$. As we shall see on the following theorem, diferentely from the Kendall tau distance where $|d_{\tau} (1, \tau \circ \sigma_r) - d_{\tau} (1, \sigma)|=1$ independently of $r$, the difference $|S (\tau_r \circ \sigma) - S (\sigma)|$ decreases with $r$.
\[teo:trans\]
Let $\sigma, \tau \in S_n$ where $\tau :=\tau_r= (r,r+1)$. Then,
$$S (\tau \circ \sigma) - S (\sigma) = \left\{\begin{matrix}
- 2^{f_r(\sigma) - 1} & \hspace{5pt} if \hspace{5pt} \sigma^{-1} (r) < \sigma^{-1} (r+1) \\
2^{f_{r+1}(\sigma)} & \hspace{5pt} if \hspace{5pt} \sigma^{-1} (r) > \sigma^{-1} (r+1)
\end{matrix}\right.$$
Since $(\tau \circ \sigma)^{-1}(r) = \sigma^{-1}(r+1)$ and $(\tau \circ \sigma)^{-1}(j) = \left\{\begin{matrix}
\sigma^{-1}(r) & if \hspace{5pt} j=r+1 \\
\sigma^{-1}(j) & if \hspace{5pt} j > r+1
\end{matrix}\right.$ for $j \geq r+1$ it follows that
$$\begin{aligned}
f_r (\tau \circ \sigma) &= \sum_{j=r+1}^n [(\tau \circ \sigma)^{-1}(r) \leq (\tau \circ \sigma)^{-1}(j) ] \\
&= [\sigma^{-1}(r+1) \leq \sigma^{-1}(r)] + f_{r+1}(\sigma)\end{aligned}$$
Since $(\tau \circ \sigma)^{-1}(r+1) = \sigma^{-1}(r)$ and $r+1<j \Rightarrow (\tau \circ \sigma)^{-1}(j) = \sigma^{-1}(j)$ it follows that
$$\begin{aligned}
\hspace{-50pt}
f_{r+1}(\tau \circ \sigma) =& \sum_{j=r+2}^n [(\tau \circ \sigma)^{-1}(r+1) \leq (\tau \circ \sigma)^{-1}(j) ] \\ &+ [ \sigma^{-1}(r) \leq \sigma^{-1}(r+1) ] - [ \sigma^{-1}(r) \leq \sigma^{-1}(r+1)] \\
=& f_r (\sigma) - [ \sigma^{-1}(r) \leq \sigma^{-1}(r+1)] \end{aligned}$$
Thus, we have $$\begin{aligned}
\hspace{-20pt}
S (\tau \circ \sigma) &= \sum_{i = 1}^{r-1} 2^{f_i(\sigma)} + 2^{f_r(\tau \circ \sigma)} + 2^{f_{r+1}(\tau \circ \sigma)} + \sum_{i = r+2}^{n} 2^{f_i(\sigma)} \\
=&\hspace{6pt} S(\sigma) + 2^{f_r(\sigma)} (2^{-[\sigma^{-1}(r) \leq \sigma^{-1}(r+1)]} - 1) \\
&+ 2^{f_{r+1}(\sigma)} (2^{[\sigma^{-1}(r+1) \leq \sigma^{-1}(r)]} - 1).\end{aligned}$$
As a direct corollary, using Corollary \[cor:decodingdist\]:
Let $\sigma, \tau \in S_n$ where $\tau = (r,r+1)$. Then, $$d_{dec} (1, \tau \circ \sigma) - d_{dec} (1, \sigma) = \left\{\begin{matrix}
\dfrac{2^{f_r(\sigma) - 1}}{2^n-1} & \hspace{5pt} if \hspace{5pt} \sigma^{-1} (r) < \sigma^{-1} (r+1) \\[10pt]
- \dfrac{2^{f_{r+1}(\sigma)}}{2^n-1} & \hspace{5pt} if \hspace{5pt} \sigma^{-1} (r) > \sigma^{-1} (r+1)
\end{matrix}\right.$$
A Distance Between Stable Channels {#sec:discha}
==================================
In this section we extend the results of the previous one to define a distance (in some sense) between channels. As in the last section we will only consider the case where the decoding cones are full dimensional, i.e. the channel is stable.
We could define a distance by setting $d(P,Q) = 1 - Pr ( \hspace{3pt} {D\widehat{e}c}_C (P) = {D\widehat{e}c}_C (Q) \hspace{3pt} )$, but we will see that a more refined distance can be defined.
Consider three channels $P,Q,R \in {Cha}_3$ such that $$O^- P = \begin{pmatrix}
1 & 3 & 3 \\
2 & 1 & 2 \\
3 & 2 & 1
\end{pmatrix} \hspace{7pt}
O^- Q = \begin{pmatrix}
1 & 3 & 2 \\
2 & 1 & 3 \\
3 & 2 & 1
\end{pmatrix} \hspace{7pt}
O^- R = \begin{pmatrix}
2 & 3 & 2 \\
1 & 1 & 3 \\
3 & 2 & 1
\end{pmatrix}
.$$
One can check, by doing all possible computations, that $d(P,Q) = d(P,R) = d(Q,R)= \frac{4}{7}$. But $Q$ differs from $P$ in only one position of a single column, while $R$ differs from $P$ in one position in two different columns. If $y_1$ or $y_2$ (the output messages corresponding, respectively, to the first and second columns) is received $P$ and $Q$ are essentially the same channel. Intuitively, we expect $Q$ to be closer to $P$ than $R$.
This distance does not use the fact that the received message will be known at the time of decoding. We will use this fact to define a more refined distance.
If we assume that the transmission is made through the channel $P$, and denote by $Q_y$ the column corresponding to the received message $y$ in $Q$, we can calculate $Pr ( \hspace{3pt} {D\widehat{e}c}_C (P_y) = {D\widehat{e}c}_C (Q_y)\hspace{2pt} )$, the probability that both decoders will be equal when a message $y$ is received[^3]. With this we can define the following distance:
\[defi12\] Let $P,Q \in {Ch}_{n\times m}$ and assume that $P$ is the channel being used. The radial decoding distance to $Q$ centered in $P$ is given by $$d_{dec}^P (Q) = 1 - Pr (\hspace{3pt} {D\widehat{e}c}_C (P_y) = {D\widehat{e}c}_C (Q_y) \hspace{3pt} ) .$$
The next theorem shows how to compute this distance.
\[teo:cha\]
Let $P, Q \in {Ch}_{n\times m}$ and $\sigma, \phi \in S_n^m$ be such that $\sigma_i$ and $\phi_i$ correspond to the ordering in the $i$-th column of $O^- P$ and $O^- Q$, respectively. Suppose that the channel being used is $P$. If a code $C \subseteq X$ is picked uniformly distributed from the space of all codes, then
$$Pr ( \hspace{3pt} {D\widehat{e}c}_C (P_y) = {D\widehat{e}c}_C (Q_y) \hspace{3pt} ) = \frac{1}{n(2^n - 1)} \sum_{i = 1}^m S(\sigma_i, \phi_i) \left \Vert P_i \right \Vert_1$$ where $\left \Vert P_i \right \Vert_1:=\sum_{j=1}^nP_{ji}$ is the $1$-norm of the $i$-th column of $P$.
$$\begin{aligned}
\hspace{-28pt}
Pr ( \hspace{3pt} {D\widehat{e}c}_C (P_y) = {D\widehat{e}c}_C (Q_y) \hspace{2pt} ) &= \sum_{i = 1}^m Pr ( \hspace{2pt} {D\widehat{e}c}_C (P_y) = {D\widehat{e}c}_C (Q_y) \vert \hspace{2pt} y_i \hspace{2pt} \text{received} ) Pr (y_i \hspace{2pt} \text{received}) \\
&= \sum_{i=1}^m \frac{S(\sigma_i, \phi_i)}{2^n - 1} \sum_{j=1}^n Pr (y_i \hspace{2pt} \text{received} \hspace{3pt} \vert x_j \hspace{2pt} \text{sent}) Pr(x_j \hspace{2pt} \text{sent}) \\
&= \sum_{i=1}^m \frac{S(\sigma_i, \phi_i)}{2^n - 1} \left \Vert P_i \right \Vert_1 \frac{1}{n}\end{aligned}$$
In the hypothesis of Theorem \[teo:cha\] we assume that one of the channels is the correct one. This occurs because the expression depends on the probability of receiving $y$ which may not coincide for different channels.
\[cor:disrad\] Let $P,Q \in {Ch}_{n\times m}$ and assume that $P$ is the channel being used. The radial decoding distance to $Q$ centered in $P$ is given by $$d_{dec}^P (Q) = 1 - \frac{1}{n(2^n - 1)} \sum_{i = 1}^m S(\sigma_i, \phi_i) \left \Vert P_i \right \Vert_1$$
We go back to the example discussed in the beggining of this section.
\[exe:conta\]
Suppose a channel $P = \begin{pmatrix}
\frac{5}{8} & \frac{1}{8} & \frac{2}{8} \\[1ex]
\frac{2}{8} & \frac{5}{8} & \frac{1}{8} \\[1ex]
\frac{1}{8} & \frac{2}{8} & \frac{5}{8}
\end{pmatrix}$ is used for transmission and $Q,R \in {Cha}_{3}$ are such that $$O^- Q = \begin{pmatrix}
1 & 3 & 3 \\
2 & 1 & 2 \\
3 & 2 & 1
\end{pmatrix}
\hspace{5pt}
\text{and}
\hspace{5pt}
O^- R = \begin{pmatrix}
2 & 3 & 2 \\
1 & 1 & 3 \\
3 & 2 & 1
\end{pmatrix} .$$
Then, by Theorem \[teo:cha\], $$\begin{aligned}
Pr ({D\widehat{e}c}_C (P_y) = {D\widehat{e}c}_C (Q_y) &= \frac{1}{21} (7 + 7 + 4) = \frac{6}{7}\end{aligned}$$ and $$\begin{aligned}
Pr ({D\widehat{e}c}_C (P_y) = {D\widehat{e}c}_C (R_y)) = \frac{1}{21} (5 + 7 + 4) = \frac{16}{21}.\end{aligned}$$ Thus, $${d_{dec}^P(Q) = \dfrac{1}{7} < \dfrac{5}{21} = d_{dec}^P(R)}.$$
We note that this difference is, intuitively, compatible with the simple observation that $Q$ differs from $P$ in only one position of a single column, while $R$ differs from $P$ in one position in two different columns.
The decoding distance presented in Definition \[def:decdist\] of the previous section was symmetric and only depended on the equivalence classes of the permutations. In contrast, the radial decoding distance to $Q$ centered in $P$ is not symmetrical, and although it only depends on the equivalence class of $Q$, it depends on the internal structure of $P$, i.e. if $Q \sim Q'$, then $d_{dec}^P(Q)=d_{dec}^P(Q')$, but $P\sim P'$ does not necessarily imply that $d_{dec}^P(Q)=d_{dec}^{P'}(Q)$.
Discussion
==========
In this work, we gave an explicit expression for a meaningful distance in the space of all channels over given input and output sets. This establishes the ground to study the details of what can be a kind of finite approximation approach to channels and decodification problems. A family of questions that arise in this context are the following: Let $\overline
{Ch_{n\times m}}=Ch_{n\times m} /\sim$ be the set of decoding cones and let $A\subset\overline{Ch_{n\times m}}$ be a subset of channels with some interesting property (for example, the set of channels that admits syndrome decoding). If we want to approximate a channel $P\in Ch_{n\times m}$ by a decoding cone in $A$, how much (in terms of decoding errors) should we expect to lose? From Corollary 3, we are actually interested in determining $\max\left\{ d_{dec}^{P}\left( Q\right) ;Q\in\overline{Ch_{n\times m} }\right\} $. Asymptotic versions arise naturally as we consider a family of increasing (in terms of $n=\left\vert X\right\vert $ and $m=\left\vert
Y\right\vert $) input and output sets.
This approach is similar to the one adopted in the study of mismatched channels as, for example, in [@gant00]. The approach used in this (and other works studying mismatched channels) rests on the determination of achievable rates, that is, in proving that, for $n$ sufficiently large *there are* codes that can be decoded with the approximating channel with no significant loss, that is, with probability of mis-decoding approaching $0$. In our approach we are not looking at this family of codes (asymptotically the best choice of code for the mismatched channel), but on the average loss while choosing sequences of codes with a given rate.
We also stress that *any* prescribed deterministic decision rule can be seen as a maximum likelihood decoding rule of some channel (actually an equivalence class of channels), as can be seen, for example, in [@fire14].
Besides that, we remark that we considered the case of stable channels, i.e. the case of a decoding cone $cone(P)$ that is determined by a set of strict inequalities. An unstable (non-full dimensional) cone $Cone(P)$ is determined by a set of inequalities and a non-empty set of equalities, or, in other words, $Order(P)$ contains equivalences. It inherits its decoders from its full dimensional neighbours, that is, cones corresponding to stable channels in which every inequality of $Cone(P)$ also holds.
Finding explicit expressions for a distance on the set of all decoding cones, both stable and unstable, is technically more challenging.
[100]{}
M. Deza and E. Deza, “Encyclopedia of distances,” 4th revised edition, *Springer-Verlag*, 2016.
R.G.L. D’Oliveira and M. Firer, “Channel Metrization,” *arXiv:1510.03104*, submitted, 2016.
R.G.L. D’Oliveira and M. Firer, “Geometry of Communication Channels: Metrization and Decoding,” *Symmetry: Culture and Science*, Volume 27, No. 4, 279–289, 2016.
R.G.L. D’Oliveira and M. Firer, “Minimum Dimensional Hamming Embeddings”, *Advances in Mathematics of Communications*, Volume: 11, No. 2, p. 359–366, 2017.
M. Firer and J.L. Walker, “Matched Metrics and Channels”, *IEEE Transactions on Information Theory*,Volume: 62, Issue: 3, p. 1150–1156, 2015.
E. Gabidulin, “A brief survey of metrics in coding theory,” *Mathematics of Distances and Applications*, p. 66–84, 2012.
A. Ganti, A. Lapidoth and E. Telatar, “Mismatched Decoding Revisited: General Alphabets, Channels with Memory, and the Wide-Band Limit”, *IEEE Transactions on Information Theory*,Volume: 46, Issue: 7, p. 2315–2328, 2000.
M. Kendall, “A New Measure of Rank Correlation,” *Biometrika*, 30, p. 81–89, 1938.
A. Makur and Y. Polyanskiy, “Comparison of channels: criteria for domination by a symmetric channel”, arXiv:1609.06877 \[cs.IT\], 2016.
A. Poplawski, “On Matched Metric and Channel Problem”, arXiv:1606.02763 \[cs.IT\], 2016.
C. Qureshi, “Matched Metrics to the Binary Asymmetric Channels”, arXiv:1606.09494 \[cs.IT\], 2016.
G. Séguin, “On metrics matched to the discrete memoryless channel”, *J. Franklin Inst. 309*, no. 3, p. 179–189, 1980.
C.E. Shannon, “ A Mathematical Theory of Communication”. *Bell System Technical Journal. 27*: 379–423, 1948.
C.E. Shannon, “ A Note on a Partial Ordering for Communication Channels”. *Information and Control*, Volume: 1, p. 390–397, 1958.
R.P. Stanley, “ An Introduction to Hyperplane Arrangements”. *Lecture notes, IAS/Park City Mathematics Institute*, 2004.
[^1]: For the use of general distances in coding theory see [@deza14; @gabi12].
[^2]: For a deeper look into conditions for metrization see references [@doli16a; @doli16d; @doli16b; @fire14] and [@popl16; @qure16; @segu80].
[^3]: In this case both the code $C$ and the message $y$ are random variables.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'We show the existence of an action of the free group on three generators $F_3$ on a compact metric space $X$, admitting an invariant probability measure $\mu$, such that the resulting dynamical system $(X, \mu, F_3)$ is distal, not equicontinuous, and strongly ergodic.'
address:
- |
Department of Mathematics\
Tel Aviv University\
Tel Aviv\
Israel
- |
Institute of Mathematics\
Hebrew University of Jerusalem\
Jerusalem\
Israel
author:
- Eli Glasner and Benjamin Weiss
title: Distal strongly ergodic actions
---
[July 18, 2017]{}
[March 4, 2018]{}
This note answers a question raised in an after dinner conversation during the conference entitled “Structure and Geometry of Polish groups" (17w5094) held in Oaxaca, Mexico June 11 to 16, 2017. The participants in this conversation were : Eli Glasner, Tomás Ibarlucía, François Le Maitre, Todor Tsankov and Robin Tucker-Drob. The first named author thanks BIRS and the organizers for the invitation to participate in this very successful conference. We thank Alex Lubotzky for calling the work [@BG-08] to our attention.
Let $S$ be a set of elements in $SL_2({\mathbb{Z}})$. If $\langle S \rangle$, the group generated by $S$, is a finite index subgroup of $SL_2({\mathbb{Z}})$, Selberg’s theorem [@S-65] implies (see e.g. [@L-94 Th. 4.3.2]) that $\mathcal{G}(SL_2({\mathbb{Z}}_p), S_p)$, the Cayley graphs of $SL_2({\mathbb{Z}}_p)$ with respect to $S_p$, the natural projection of $S$ modulo $p$, form a family of expanders as $p\to \infty$.
The following theorem is due to Bourgain and Gamburd [@BG-08 Theorem 1].
Let $S$ be a set of elements in $SL_2({\mathbb{Z}})$. Then the $\mathcal{G}(SL_2({\mathbb{Z}}_p),S_p)$ form a family of expanders if and only if $\langle S \rangle$ is non-elementary, i.e. the limit set of $\langle S \rangle$ consists of more than two points (equivalently, $\langle S \rangle$ does not contain a solvable subgroup of finite index).
Recall that a probability measure preserving action $(X, \mathcal{X}, \mu, G)$ is said to have the [*spectral gap property*]{} if the corresponding Koopman representation $\pi$ on $L^2_0(X, \mu)$ does not have almost invariant vectors, that is, if there is no sequence of unit vectors $\xi_n$ in $L^2_0(X, \mu)$ such that $$\lim \|\pi(g) \xi_n - \xi_n\| =0, \quad \forall g \in G.$$ The action is called [*strongly ergodic*]{} if there is no sequence of sets $A_n \in \mathcal{X}$ with $\mu(A_n)=1/2, \forall n \in {\mathbb{N}}$ such that $$\lim \mu(g A_n {\bigtriangleup}A_n) =0, \quad \forall g \in G.$$ It is not hard to see that the spectral gap property implies strong ergodicity.
We want to find compact metrizable groups $K$ and $L$ with the following properties:
- $K$ contains three elements $a, b$ and $c$ such that the subgroup ${\Lambda}= \langle a, b, c \rangle$ generated by them is free and dense in $K$, ${\overline}{{\Lambda}} = K$.
- The subgroup ${\Lambda}_0 = \langle a, b \rangle < {\Lambda}$ is also dense in $K$, $K= {\overline}{{\Lambda}_0}$.
- $L$ contains two elements $f, g$ such that the subgroup $ {\Sigma}=\langle f, g \rangle$ generated by them is free and dense in $L$, $L = {\overline}{{\Sigma}}$.
Moreover, denoting by $F_2, F_3$ the free groups on two and three generators respectively, all of the following actions (under left multiplication and with respect to the corresponding Haar measures) are strongly ergodic:
1. the action of $F_3$ via ${\Lambda}$ on $K$,
2. the action of $F_2$ via ${\Lambda}_0$ on $K$,
3. the action of $F_2$ via $ {\Sigma}$ on $L$,
4. the diagonal action of $F_2$ via $\langle (a,f), (b,g) \rangle$ on $K \times L$.
Here is a construction of such groups.
Let $x=\left(\begin{smallmatrix}1 & 2\\0 & 1\end{smallmatrix}\right)$ and $y = \left(\begin{smallmatrix}1 & 0\\2 & 1\end{smallmatrix}\right)$ be two elementary matrices generating a free group, which we call $H$, in $SL_2({\mathbb{Z}})$ (it is actually of finite index). Let $a = x^2 = \left(\begin{smallmatrix}1 & 0\\4 & 1\end{smallmatrix}\right)$ and $b = y^2 = \left(\begin{smallmatrix}1 & 4\\0 & 1\end{smallmatrix}\right)$, and let $c$ be another element in $H$ which together with $a$ and $b$ generate a free group on $3$ generators (such an element clearly exists as a and b generate a subgroup of infinite index in $H$). Now set $$K = \prod \{{\Lambda}_p : p\ {\text{prime, \ and}}\ \ p \equiv1\pmod{4}\}$$ and $$L = \prod \{{\Lambda}_p : p\ {\text{prime, \ and}}\ \ p \equiv3\pmod{4}\}.$$ As explained in [@BG-08] (see the remark on page 628 and Propositions 4 and 6) for sufficiently large $p$, the image of $S_0 = \{a, b\}$ (and a fortiori that of the set $S = \{a, b, c\}$) in $SL_2({\mathbb{Z}}_p) \cong SL_2({\mathbb{Z}})/ SL_2(p{\mathbb{Z}})$ generates $SL_2({\mathbb{Z}}_p)$. Thus, assuming our $p$’s are sufficiently large, we have $K = {\overline}{{\Lambda}} = {\overline}{{\Lambda}_0}$.
Now, with $S = \{a, b, c\}$, $S_p$ the image of $S$ in $SL_2({\mathbb{Z}}_p)$, and $K$ and $L$ as above, applying the theorem of Bourgain and Gamburd, we see that the family $\mathcal{G}(SL_2({\mathbb{Z}}_p),S_p)$ forms a family of expanders. This, in turn, is well known to imply the spectral gap property, of each of the actions (1) - (4) (with $f, g$ the maps induced on $L$ by $a, b \in SL_2({\mathbb{Z}})$). Thus they all have the spectral gap property, hence are strongly ergodic. Therefore the requirements (1) - (4) are satisfied.
Next let $K_1 = {\overline}{\langle c \rangle} < K$. By [@Z-77] there is a cocycle $\phi_0 : K_1 \to L$ such that the corresponding ${\mathbb{Z}}$-action on $K_1 \times L$ given by $$T_c(x,y) = (cx, \phi_0(x) y), \quad x \in K_1, y \in L,$$ is ergodic. Note that, as $L$ is non-commutative this distal ${\mathbb{Z}}$-action, is necessarily not equicontinuous. Define $\phi : K \to L$ by the formula $$\phi(x) = \phi_0(k_t^{-1}x) \quad {\text {for $x$ in the coset $k_tK_1$}},$$ where $t \mapsto k_t, K/K_1 \to K$ is a Borel section for the map $K \to K/K_1$.
On $X \times Y := K \times L$ define an $F_3$ action $\{T_t\}_{t \in F_3} : K \times L \to K\times L$ by: $$\begin{gathered}
T_a(x,y) = (ax, fy)\\
T_b(x,y) = (bx, gy)\\
T_c(x,y) = (cx, \phi(x)y).\end{gathered}$$
With this data at hand we can now prove our main result:
The $T$ action of $F_3$ on $K \times L$ is distal, not equicontinuous, and strongly ergodic.
Clearly the $T$ action on $K \times L$ is distal (of order $2$). It can not be equicontinuous because the restriction to the ${\mathbb{Z}}$-action $T_c : K_1 \times L \to K_1 \times L$ is not equicontinuos. Finally, by construction, the $F_2$-action on $K \times L$ is strongly ergodic and, a fortiori, so is the $T$-action. Our proof is complete.
[WWW]{}
Alexander Lubotzky, [*Discrete Groups Expanding Graphs and Invariant Measures*]{}, Progress in Math. 195, Birkhaüser, Basel, 1994.
Jean Bourgain and Alex Gamburd, [*Uniform expansion bounds for Cayley graphs of $SL_2(\mathbb{F}_p)$*]{}, Annals of Mathematics, [**167**]{} (2008), 625–642.
Atle Selberg, [*On the estimation of Fourier coefficients of modular forms*]{}, Proc. Sympos. Pure Math. VII (1965), 1–15.
Robert J. Zimmer, [*Random walks on compact groups and the existence of cocycles*]{}, Israel J. of Math., [**26**]{}, (1977), 84–90.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'Measurement of the acoustic peaks of the cosmic microwave background (CMB) temperature anisotropies has been instrumental in deciding the geometry and content of the universe. Acoustic peak positions vary in different parts of the sky owing to statistical fluctuation. We present the statistics of the peak positions of small patches from ESA Planck data. We found that the peak positions have significantly high variance compared to the 100 CMB simulations with best-fit $\Lambda$CDM model with lensing and Doppler boosting effects included. Examining individual patches, we found the one containing the mysterious “Cold Spot”, an area near the Eridanus constellation where the temperature is significantly lower than Gaussian theory predicts, displays large synchronous shift of peak positions towards smaller multipole numbers with significance lower than $1.11\times 10^{-4}$. The combination of large synchronous shifts in acoustic peaks and lower than usual temperature at the Cold Spot area results in a 4.73$\sigma$ detection (significance $p\simeq 1.11\times 10^{-6}$) against the $\Lambda$CDM model. And it was already reported in Finelli et al. (2016) that in the WISE-2MASS galaxy catalog at $z<0.3$ the Cold Spot region is surrounded by surprisingly large underdense regions around 15$^\circ$ in radius, which are found to be in the same square patch. Thus we propose there is some extra localized unknown energy to stretch out the space in the transverse direction around the Cold Spot area to simultaneously account for the Cold Spot, the excessive shift of the acoustic peaks, and the large underdense regions.'
author:
- 'Lung-Yih Chiang'
title: Excessive shift of the CMB acoustic peaks of the Cold Spot area
---
Introduction
============
The measurement of the acoustic peak positions of the cosmic microwave background (CMB) temperature anisotropies has been instrumental in deciding the geometry and content of the universe[@hu; @wmap9year], e.g. the first acoustic peak at $\l\simeq$220 indicating a flat universe. It was the Boomerang balloon experiment in 2000 that firstly shows the first acoustic peak position (in a partial sky) is measured at $\l=197\pm 6$ [@boomerang], later MAXIMA-1 at $\l\simeq220$ [@maxima], and DASI at around 200 [@dasi], VSA at 224 [@vsa], compared to full-sky measurement at $220.8\pm0.7$ [@wmap3yeartem] and the latest result to be $\l=220.0\pm0.5$ [@planckps1; @planckps2]. The discrepancy is mainly due to the fact that the acoustic peak positions vary in different parts of the sky. In this paper, we test the statistics of the peak positions from ESA Planck data. This paper is arranged as follows. In Sec.\[sec:processing\] we describe the data processing of data and we use variance as the statistics on peak positions in Sec.\[sec:variance\]. We show the excessive peak shifts near the Cold Spot area in Sec.\[sec:excessive\]. Discussion is in Sec.\[sec:discussion\].
Planck data processing and fitting {#sec:processing}
==================================
To test the statistics of the acoustic peaks, we use flat sky approximation (FSA) and take patches of 20$\times$20 deg$^2$ from the 4 CMB maps: SMICA, NILC, SEVEM and Commander maps in pixelization[@glesp], each with a specific foreground cleaning method on either harmonic or pixel domain [@planckfg1]. The map is rotated every $22^\circ.5$ (in practice it’s the $\alm$ of the map for rotation before being repixelized with with 2 arcmin/pixel) and we take the patches parallel to the $\theta$ and $\phi$ coordinate lines centered at the equator to minimize the error from pixelization. As such we separate the whole sphere into 90 patches. In our analysis we further use the GAL070 mask [@planckfg2] to block areas of heavy foreground contamination around Galactic plane, thus in total there are 45 patches available for test in each map, accounting for 43% of the whole sky.
{width="35.00000%"} {width="35.00000%"}\
{width="35.00000%"} {width="35.00000%"}\
Each patch is firstly interpolated in the $\theta$ direction as the spacing of the grid is slightly uneven due to the fact that the grids are where the Gauss-Legendre polynomial zeros are [@glesp]. Two-dimensional Fourier transform is used to analyze each patch. The scaling relation between integer Fourier wavenumber $k$ and multipole number $\l$ is $\l=2 \pi k /L$, where $L$ is the size of the patch. The amplitude of the angular power spectrum $S_\l$ can be scaled from $S_k$ at Fourier wavenumber $k$ via $$S_{\l\equiv 2\pi k/L}= L^2 S_k.$$ So for each patch with size $L$, the wavenumber $k$ is scaled into $2\pi k/L$, and there is no binning in our power spectrum and the sequence of $k$ is scaled such that the multipole interval $\Delta \l\equiv 2\pi/L$. One can also see the choice of the patch size cannot be arbitrary. The sampling interval $\Delta \l$ of angular power spectrum is inversely proportional to $L$, so we want to choose patches as large as possible to have more sampled data in the power spectrum, but it can’t be too large as the curvature of the sphere causes the flat sky approximation to break down. The scaling relation from Fourier to spherical harmonic analysis is already tested with simulations [@direct].
One issue related to Fourier transform on real signals is the non-periodic boundary (NPB) condition, which induces extra power from jump discontinuity. The condition is aggravated when there is beam convolution causing jump discontinuity more pronounced at the boundary. We have conducted simulations to estimate the extra power due to 5-arcmin beam convolution on a 20$\times$20 deg$^2$ patch to be $1.21\, \ell^{-1.38}$ ($\mu$K$^2$) (see the Appendix A). This should be subtracted before any data processing.
It is assumed that the 4 maps are fairly cleaned such that the angular power spectrum in each patch $S_k=b_k^2 C_k +N_k$, where $C_k$ denotes the CMB, $b_k$ the beam transfer function and $N_k$ the noise. We then use cross-power spectrum (XPS) to reduce the noise (see Appendix B). The XPS is a quadratic estimator between two patches or maps to provide an unbiased estimate of the power spectrum of the correlated signals: $$b^2_kC_k^{\alpha\beta}= \frac{1}{2} \langle \alpha^{*}_{{\mbox{\boldmath $k$}}} \beta_{{\mbox{\boldmath $k$}}} + \beta^{*}_{{\mbox{\boldmath $k$}}} \alpha_{{\mbox{\boldmath $k$}}} \rangle,
\label{def}$$ where $\alpha_{{\mbox{\boldmath $k$}}} $ and $\beta_{{\mbox{\boldmath $k$}}}$ are the Fourier modes of patches from halfring maps, $*$ denotes complex conjugate and the angle brackets denote an average over all $k-1/2 \le |{{\mbox{\boldmath $k$}}}| < k+1/2$ for integer k. The residual of the XPS $X_k^{\alpha\beta}$ has to be further subtracted, which stems from the lack of an ensemble and non-zero chance correlation between the “uncorrelated” signal [@xps; @direct], $$\sqrt{\langle (\xk^{\alpha\beta})^2 \rangle} \simeq \frac{\sqrt{ N^\alpha_k N^\beta_k } }{\sqrt{2\pi k}},
\label{xpsuncorr}$$ where $N^\alpha_k$ and $N^\beta_k$ are the (uncorrelated) noise power spectrum from halfring patches, respectively. The noise power spectrum can be obtain by $1/2$ of the differencing of halfring patches, assuming both are of the same level. Thus in our data pipeline, we subtract the residual $X_k^{\alpha\beta}$ from the XPS, following the correction of NPB.
It is then subsequently de-convolved with a beam transfer function at 5 arcmin [@planckfg2]. To check the validity of our data processing, in the 4 panels of Figure \[dl\] we show the mean angular power spectrum from the 45 patches of the 4 maps, which fits very well with the best-fit $\Lambda$CDM power spectrum. And in Fig \[dispersion\] we show the dispersions of the 45 power spectra from the patches, which fall within the range of the simulation up to $\l=1,350$.
We use the Levenberg-Marquardt algorithm [@recipes] as the non-linear least-square fitting routine to fit the angular power spectrum curve with sum of 5 Gaussian functions to the beginning of the 5th peak (from $\l=50$ to 1,350). The peak positions are estimated in $\l$ and only the first 3 peaks are estimated because the low amplitudes of the 4th and 5th peaks are prone to error in the estimation. The distribution of the peak positions of the SMICA map is represented in 3 dimensions of ${{\mbox{\boldmath $\l$}}}\equiv(\l^{(1)}, \l^{(2)},\l^{(3)})$ in top panel of Figure \[variance\].
![The dispersion of the 45 angular power spectra of the Planck patches with green filled circle, blue asterisk, red diamond and and light blue square sign denoting SMICA, NILC, SEVEM and Commander patches, respectively and those from the 100 simulations (black dot). One can see the dispersions from the Planck 4 maps fall within the 100 simulations up to $\l= 1,350$, which is then adopted as the the upper limit of our Gaussian-fitting range.[]{data-label="dispersion"}](sigma.pdf){width="45.00000%"}
Any statistics from the data are to be compared with lensing simulations by Consortium [@plancklensing1]. There are 100 realizations based on best-fit $\Lambda$CDM model with both the lensing and the Doppler boosting effects included [^1], which are originally employed to determine the lensing mean field, the measurement error bars, and to validate lensing reconstruction methodology. The angular resolution of the simulations is 5 arcmin, the same as the 4 Planck maps. Gravitational lensing effect [@blanchard; @cole; @linder; @seljak; @challinorlense; @lewis] is the deflection of CMB photons coming from the last scattering surface by potential gradients along our line of sight, and the team used both temperature and polarization data to measure the lensing potential at a level of 40$\sigma$ [@plancklensing1; @plancklensing2]. The Doppler boosting effect [@challinorpeculiar], on the other hand, is due to the motion of our solar system barycenter with respect to the CMB rest frame that induces not only the dipolar anisotropy, but also the Doppler modulation and aberration at high multipoles [@planckdb]. Both effects on the angular power spectrum at high multipoles are small, but they can affect significantly the angular power spectrum of a small patch [@jeong]. Since both effects are still in the 4 maps, it is thus appropriate to take for comparison the simulation with the aforementioned effects included. We process the 100 simulations with the same pipeline as the 4 maps with the same choice of patches and rescaling.
![Distribution of the acoustic peak positions of the SMICA 45 patches and histogram of the variances of the peak positions from 100 realizations of simulation. In the top panel each patch is represented with one point ${{\mbox{\boldmath $\l$}}}\equiv(\l^{(1)}, \l^{(2)},\l^{(3)})$ with the peak positions denoted in 3 coordinates. The color in each point denotes the angle between the Galactic North Pole and the patch center. In the bottom the variances of the 4 maps and 143 GHz band map are denoted with colored lines.[]{data-label="variance"}](distribution.pdf "fig:"){width="50.00000%"} ![Distribution of the acoustic peak positions of the SMICA 45 patches and histogram of the variances of the peak positions from 100 realizations of simulation. In the top panel each patch is represented with one point ${{\mbox{\boldmath $\l$}}}\equiv(\l^{(1)}, \l^{(2)},\l^{(3)})$ with the peak positions denoted in 3 coordinates. The color in each point denotes the angle between the Galactic North Pole and the patch center. In the bottom the variances of the 4 maps and 143 GHz band map are denoted with colored lines.[]{data-label="variance"}](histo_var.pdf "fig:"){width="45.00000%"}
Variance of the peak positions {#sec:variance}
==============================
We use the variance of the first 3 peaks to characterize the statistics of peak positions: $$\sum^n_{i=1} |{{\mbox{\boldmath $\l$}}_i}-\bar{{\mbox{\boldmath $\l$}}}|^2/(n-1),$$ where $\bar{{\mbox{\boldmath $\l$}}}$ is the mean of the peak positions from the n=45 patches. In bottom panel of Figure \[variance\] we plot the histogram of the variance from 100 CMB realizations and those of the 4 maps are denoted with colored lines. The result shows that the variances of the peak positions from 4 maps are all quite large with significance $p< 0.04$, in particular the variance of the SMICA map is higher than 98 realizations, which is equivalent roughly to a significance $p$-value 0.02 against the null hypothesis.
We examine the possible effect from foreground residual with the mean and the dispersion of the 45 power spectra from the patches. It is based on the fact that any extra non-correlated signal added to the CMB would increase both the mean and dispersion of the power spectra. As shown in Figure \[dl\] the mean angular power spectra fit well with the best-fit $\Lambda$CDM power spectrum, and in Figure \[dispersion\] the dispersions from the 45 patches fall within the 100 realizations up to $\l= 1,350$. Thus we are guaranteed that the data is at least free from significant foreground residual up to $\l=1,350$.
If foreground residual is the cause of the anomaly, the variance from any frequency band map should be higher than that of the SMICA map because the foreground contamination in any band map is definitely higher than that in SMICA map. However, the variance for the 143 GHz channel map containing diffuse foreground is 226.56 (where we only subtract bright point sources), smaller than that of SMICA map 236.79 instead (where the distance between the mean peak positions of 143GHz and those of SMICA is only 2.19). The foreground residual as the cause for the anomaly can then be ruled out. The influence of the Galactic plane is also tested by color-coding the points for the distribution of the peaks. There is no obvious clustering in color, hence no systematic effect from either Galactic (as shown in top panel of Figure \[variance\]) or Ecliptic influence.
![The peak distance from the mean: $|{{\mbox{\boldmath $\l$}}_i}-\bar{{\mbox{\boldmath $\l$}}}|$ of the 45 patches on the sky (top), and their histogram (black curve in the bottom panel). In the top panel we use large round symbol centered in each patch to denote the distance with color, and the patch with red color is centered at $(l,b)=(204^\circ.74,-65^\circ.48)$, where the famed Cold Spot is inside. In the bottom panel the red curve is the histogram of the 9000 patches from the simulation.[]{data-label="distance"}](dist.pdf "fig:"){width="40.00000%"}\
![The peak distance from the mean: $|{{\mbox{\boldmath $\l$}}_i}-\bar{{\mbox{\boldmath $\l$}}}|$ of the 45 patches on the sky (top), and their histogram (black curve in the bottom panel). In the top panel we use large round symbol centered in each patch to denote the distance with color, and the patch with red color is centered at $(l,b)=(204^\circ.74,-65^\circ.48)$, where the famed Cold Spot is inside. In the bottom panel the red curve is the histogram of the 9000 patches from the simulation.[]{data-label="distance"}](histo_distance.pdf "fig:"){width="30.00000%"}
Excessive peak shifts in the Cold Spot patch {#sec:excessive}
============================================
Since the variance is related to the “distance” of the peak positions of patch $i$ to their mean, $|{{\mbox{\boldmath $\l$}}_i}-\bar{{\mbox{\boldmath $\l$}}}|\equiv \sqrt{(\l_i^{(1)}-\bar\l^{(1)})^2+(\l_i^{(2)}-\bar\l^{(2)})^2+(\l_i^{(3)}-\bar\l^{(3)})^2}$, we show the distance of the 45 patches on the sky in top panel of Figure \[distance\]. One can easily notice one patch has far larger distance 40.72 than the rest. That patch is centered at $(l,b)=(204^\circ.74,-65^\circ.48)$ near the Eridanus constellation, which contains the well-known CMB “Cold Spot” (CS) [@vielva; @cruz] (see top left panel of Figure \[coldspot\]), an area where the CMB temperature has lower than Gaussian theory predicts with significance around 1% [@planckstat]. One should also note that the 3 peaks of the patch with CS not only have larger shifts than usual, but also display synchronous shifts towards smaller multipole numbers ([*i.e.*]{} larger scales) than the mean, as shown in middle panel of Figure \[coldspot\].
![The patch containing the well-known CS from SMICA map (top left panel) and with the CS subtracted (top right), where color bar range is $[-500,500]\mu$K. Middle panel shows the angular power spectra of the CS patch from 4 maps with green filled circle, blue asterisk, red diamond and light blue square sign denoting SMICA, NILC, SEVEM and Commander patches, respectively. One can see the data points of the first 3 peaks are mostly on the left side of the best-fit $\Lambda$CDM model (red curve). For comparison, we also show that of the same field from 143 GHz map (orange plus sign), which still contains foreground, and is corrected only for the extra power from non-periodic boundary condition. Even with foreground, the peak positions at 143GHz are close to those of the CMB patches. Also shown is that of the CS-subtracted square (black triangle). In the bottom panel, we show the same for the CS patch of the foreground-subtracted maps at 100, 143 and 217 GHz from SEVEM method (green filled circle, orange asterisk and blue square, respectively).[]{data-label="coldspot"}](cs_1.pdf "fig:"){width="20.00000%"} ![The patch containing the well-known CS from SMICA map (top left panel) and with the CS subtracted (top right), where color bar range is $[-500,500]\mu$K. Middle panel shows the angular power spectra of the CS patch from 4 maps with green filled circle, blue asterisk, red diamond and light blue square sign denoting SMICA, NILC, SEVEM and Commander patches, respectively. One can see the data points of the first 3 peaks are mostly on the left side of the best-fit $\Lambda$CDM model (red curve). For comparison, we also show that of the same field from 143 GHz map (orange plus sign), which still contains foreground, and is corrected only for the extra power from non-periodic boundary condition. Even with foreground, the peak positions at 143GHz are close to those of the CMB patches. Also shown is that of the CS-subtracted square (black triangle). In the bottom panel, we show the same for the CS patch of the foreground-subtracted maps at 100, 143 and 217 GHz from SEVEM method (green filled circle, orange asterisk and blue square, respectively).[]{data-label="coldspot"}](cs_3.pdf "fig:"){width="20.00000%"} ![The patch containing the well-known CS from SMICA map (top left panel) and with the CS subtracted (top right), where color bar range is $[-500,500]\mu$K. Middle panel shows the angular power spectra of the CS patch from 4 maps with green filled circle, blue asterisk, red diamond and light blue square sign denoting SMICA, NILC, SEVEM and Commander patches, respectively. One can see the data points of the first 3 peaks are mostly on the left side of the best-fit $\Lambda$CDM model (red curve). For comparison, we also show that of the same field from 143 GHz map (orange plus sign), which still contains foreground, and is corrected only for the extra power from non-periodic boundary condition. Even with foreground, the peak positions at 143GHz are close to those of the CMB patches. Also shown is that of the CS-subtracted square (black triangle). In the bottom panel, we show the same for the CS patch of the foreground-subtracted maps at 100, 143 and 217 GHz from SEVEM method (green filled circle, orange asterisk and blue square, respectively).[]{data-label="coldspot"}](dl_coldspot.pdf "fig:"){width="40.00000%"} ![The patch containing the well-known CS from SMICA map (top left panel) and with the CS subtracted (top right), where color bar range is $[-500,500]\mu$K. Middle panel shows the angular power spectra of the CS patch from 4 maps with green filled circle, blue asterisk, red diamond and light blue square sign denoting SMICA, NILC, SEVEM and Commander patches, respectively. One can see the data points of the first 3 peaks are mostly on the left side of the best-fit $\Lambda$CDM model (red curve). For comparison, we also show that of the same field from 143 GHz map (orange plus sign), which still contains foreground, and is corrected only for the extra power from non-periodic boundary condition. Even with foreground, the peak positions at 143GHz are close to those of the CMB patches. Also shown is that of the CS-subtracted square (black triangle). In the bottom panel, we show the same for the CS patch of the foreground-subtracted maps at 100, 143 and 217 GHz from SEVEM method (green filled circle, orange asterisk and blue square, respectively).[]{data-label="coldspot"}](dl_fgsub_sevem.pdf "fig:"){width="40.00000%"}
One should note that a decrement, in particular, an extended decrement in CMB temperature does not cause peak shifts, i.e. there is no correlation between a decrement in CMB temperature and peaks leaning towards smaller $\l$s. To see this, we conduct CMB simulations with best-fit $\Lambda$CDM model, in total we simulate 4,000 realizations of full-sky CMB maps (with 5-arcmin angular resolution). We use spherical Mexican hat wavelet (SMHW) [@smhw; @vielva] to analyze the set of data by convolving the SMHW on the whole simulated map via $$w(R,p)=\sumlm{\alm W_\ell(R) \ylm(p)},$$ where $\ylm$ is the spherical harmonics, $\alm$ is the spherical harmonic coefficients of the simulated map, $p$ is the position/pixel, $W_\ell(R)$ is the SMHW window function at scale $R$, and $w(R,p)$ is the SMHW coefficient. We adopt $R=300$ arcmin and found 119 simulated cold spots that have the SMHW coefficients smaller than that of the CS. We further use the CS lowest temperature as the criterion to eliminate 13, so that the 106 cold spots have deeper dip than the CS (see one example in top left panel of Fig.\[peakfromminima\]). We place the cold spots at the center of $20\times20$ deg$^2$ patches and plot in Fig.\[peakfromminima\] the normalized histograms of the 3 peak positions and of the distance from the best-fit positions $(\ell^{(1)},\ell^{(2)},\ell^{(3)})=(220.0, 537.5, 810.8)$ [@planckps2]. For comparison, we plot those from 100 patches from the lensing simulations. We further shift all the cold spots by 10 deg (top right panel as an example) so that roughly half of the cold spot is on the edge of the patch to mimic the CS patch and plot the histograms. One can see the normalized histograms of the peak positions and the distance of the full- and half-cold spot patches do not show any significant skew or shift compared to those from the ordinary CMB patches.
Moreover, extended cold areas with size $> 2^\circ$ are ubiquitous in CMB, which manifest themselves in the angular power spectrum at $\l < 90$, outside the range of the acoustic peaks. Thus, a more straightforward way to demonstrate the temperature dip in CS itself doesn’t cause peak shifts is that we subtract the CS shape from the patch as follows: we take only the upper left quarter of the CS patch, where the CS is located, and set the temperature of the rest of the patch zero. Then it is smoothed with $\sigma=1^\circ$, with which the CS patch is subtracted. The patch now is without the prominent dip, shown in the top right panel in Figure \[coldspot\], and the corresponding power spectrum is shown with black dot in the middle of the same Figure. One can see the CS decrement is mitigated with power spectrum being modified only at $\ell < 180$.
![Top left panel is an example of a simulated cold spot placed roughly in the center of a $20\times20$ deg$^2$ patch, whereas the right panel is that by shifting 10 deg such that half of the cold spot is on the edge of the patch to mimic the CS patch. Bottom 4 panels show the normalized histograms of the first 3 peak positions and the peak distance from 100 simulated full-cold spots placed at the patch center (black curves) and 100 half-cold spots (blue curves). For comparison, we plot the same for 100 patches taken from the lensing simulation (red curves). The vertical lines denote the best-fit peak positions $ (\ell^{(1)},\ell^{(2)},\ell^{(3)})=(220.0, 537.5, 810.8)$, respectively. One can see the distributions of peak positions and distance do not display any skew or shift compared to those from the ordinary CMB patches, and the distance from all 3 groups are less than 32.[]{data-label="peakfromminima"}](wv_0.pdf "fig:"){width="20.00000%"} ![Top left panel is an example of a simulated cold spot placed roughly in the center of a $20\times20$ deg$^2$ patch, whereas the right panel is that by shifting 10 deg such that half of the cold spot is on the edge of the patch to mimic the CS patch. Bottom 4 panels show the normalized histograms of the first 3 peak positions and the peak distance from 100 simulated full-cold spots placed at the patch center (black curves) and 100 half-cold spots (blue curves). For comparison, we plot the same for 100 patches taken from the lensing simulation (red curves). The vertical lines denote the best-fit peak positions $ (\ell^{(1)},\ell^{(2)},\ell^{(3)})=(220.0, 537.5, 810.8)$, respectively. One can see the distributions of peak positions and distance do not display any skew or shift compared to those from the ordinary CMB patches, and the distance from all 3 groups are less than 32.[]{data-label="peakfromminima"}](wv_1.pdf "fig:"){width="20.00000%"} ![Top left panel is an example of a simulated cold spot placed roughly in the center of a $20\times20$ deg$^2$ patch, whereas the right panel is that by shifting 10 deg such that half of the cold spot is on the edge of the patch to mimic the CS patch. Bottom 4 panels show the normalized histograms of the first 3 peak positions and the peak distance from 100 simulated full-cold spots placed at the patch center (black curves) and 100 half-cold spots (blue curves). For comparison, we plot the same for 100 patches taken from the lensing simulation (red curves). The vertical lines denote the best-fit peak positions $ (\ell^{(1)},\ell^{(2)},\ell^{(3)})=(220.0, 537.5, 810.8)$, respectively. One can see the distributions of peak positions and distance do not display any skew or shift compared to those from the ordinary CMB patches, and the distance from all 3 groups are less than 32.[]{data-label="peakfromminima"}](peak_wv.pdf "fig:"){width="47.00000%"}
To check the significance of the distance of the CS patch, we calculate the distance of the peaks from the 9000 patches from the 100 CMB realizations. In the bottom panel of Figure \[distance\], the black curve is the histogram of the SMICA map and the red 9000 simulated patches. Out of 9000 patches, only 2 with the distance larger than that of the CS patch. If one considers, however, synchronous shift towards smaller multipole numbers, there is none out of 9000 (where the highest distance with all peaks shifting towards large scales is 36.91), pushing the significance to less than $1.11\times 10^{-4}$. With the CS patch taken on the equator $\phi=[180^\circ,200^\circ]$ of the rotated SMICA map with $\theta=+67.5^\circ$, there is another slight shift from the Cold Spot patch taken at the same position from $+66^\circ$-rotated map, where the peak distance is 38.4 and shifts towards smaller multipole numbers. That an adjacent area also has the significance less than $1.11\times 10^{-4}$ shows the CS patch is not a fluke.
Below we discuss possible errors affecting the angular power spectrum. Since we compare the distance of the Cold Spot patch with those of the simulations including lensing and Doppler boosting effects, one can rule out both effect. The other common culprit is the foreground residual. One should note, however, that the patch is at high Galactic latitude, where the foreground contamination is relatively low. To demonstrate that even the foreground doesn’t affect much of the peak positions, we also plot the angular power spectrum of the same field from 143GHz channel map in middle panel of Figure \[coldspot\], which is only corrected for NPB effect. The shoot-up in power is due to the deconvolution on both the foregrounds and noise. One can see the peak positions shift very little even with the presence of the foregrounds, let alone foreground residual after the cleaning. We also plot the power spectra of the Cold Spot patch from the foreground-subtracted maps at 100, 143, 217 GHz from SEVEM method in the bottom panel, and they show the same trend. The summary of the variance of maps and estimated peak positions of the CS patch is in Table 1.
[ |p[4.4cm]{}||p[3.3cm]{}| ]{}\
Map & Variance from 45 patches\
SMICA & 236.79\
NILC & 233.45\
SEVEM &227.37\
Commander &232.07\
143 GHz & 226.56\
[ |p[2.5cm]{}||p[1.5cm]{}|p[1.5cm]{}|p[1.5cm]{}| ]{}\
Patch & 1st peak & 2nd peak &3rd peak\
SMICA & 206.08 & 504.76& 788.73\
NILC & 204.29 & 504.78 & 791.60\
SEVEM &205.84 & 505.95& 791.66\
Commander &205.56 & 505.15& 791.02\
SMICA (no CS)& 211.67 & 502.75 &791.28\
143 GHz & 206.31 & 504.27 & 792.03\
[ |p[2.5cm]{}||p[1.5cm]{}|p[1.5cm]{}|p[1.5cm]{}| ]{}
\
patch & 1st peak & 2nd peak &3rd peak\
100 GHz & 208.40 & 508.15& 796.05\
143 GHz & 205.48 & 507.09 & 790.67\
217 GHz &206.20 & 505.93& 796.52\
Even if there is still unknown residual after foreground cleaning, we can use analytical approach to estimate the error budget on the peak positions. We assume any unknown foreground or systematic residual has power-law power spectrum $B \ell^{-\lambda}$, then we can model the residual and an acoustic peak at $\ell_0$ with amplitude $A$ and spread $\sigma$ via $D_\ell=A \exp[-(\ell-\ell_0)^2/2\sigma^2]+ B \ell(\ell+1) \ell^{-\lambda} \exp[\ell(\ell+1)b^2]/2\pi $, where $b$ is the beam transfer function (see Appendix C). It is easy to see that, for $\lambda < 2$, the resultant $D_\ell$ at high $\l$ is dominated by $B \ell^{2-\lambda} \exp[\ell(\ell+1)b^2]/2\pi$. In the SMICA, NILC and SEVEM patches, in particularly the patch containing the Cold Spot, the shoot-up is not seen at high $\l$, indicating the residual coefficient $B$ has to be extremely small to counter both the deconvolution $ \exp[\ell(\ell+1)b^2]$ and $\ell^{2-\lambda}$. Thus peak shift from a small $B$ is negligible. For $\lambda >2 $, on the other hand, the peak shift is approximately proportional to $ B\sigma^2 /A \ell_0^{\lambda-1}$. The $B$ will have to be much higher than the best-fit $\Lambda$CDM model at low $\ell$ to have any significant peak shift, which again is not seen in the angular power spectrum of the CS patch, either. For $\lambda=2$ the peak positions are mostly unchanged because the $\l(\l+1)$ cancels out the residual.
![The SMICA field centred at the CS (left) and the corresponding WISE-2MASS field (right), both panels taken from @szapudiproc (see also @finelli). The gray circles from inner to outer correspond to radii of 5$^\circ$, 14$^\circ$ and 29$^\circ$, respectively. The black square indicates our CS patch as shown in top left panel of Figure \[coldspot\]. As is stated in @szapudiproc: ”[*the size of the underdense region is surprisingly large: it is detected to $\sim$20$^\circ$ with high ($\geq 5\sigma$) significance*]{}”. One can see the surprisingly large underdense region (dark blue region) is mostly included in the square patch.[]{data-label="wise2mass"}](wise2mass.pdf){width="48.00000%"}
Discussion {#sec:discussion}
==========
Investigation of the CMB CS is well documented. Since its discovery by @vielva, there have been proposals to explain such deviation from Gaussianity: foreground [@cruz06; @hansen], multiple voids [@multivoid], supervoid [@inoue06; @inoue07], cosmic texture [@zhao; @cruz], adiabatic perturbation on the last scattering surface [@valkenburg]. In particular, a supervoid is found with the size $\sim$220$h^{-1}$ Mpc to explain the lower than usual temperature by @voidszapudi. It is, however, refuted later that the model of a supervoid alone cannot explain the Cold Spot decrement [@nadathur; @finelli; @mackenzie; @marcos]. Nevertheless, in trying to justify such model, @voidszapudi use WISE-2MASS catalog to identify underdensities with $z<0.3$, from which a surprisingly large underdense region $\sim$20$^\circ$ close to the CS with high ($\geq 5\sigma$) significance is detected [@szapudiproc]. In Figure \[wise2mass\] we reproduce the figure 2 from @szapudiproc, and rotate 29$^\circ$ counterclockwise. The black square indicates exactly the CS patch we are investigating in this paper. One can clearly see that the large underdense region (dark blue) in the right panel is mostly included in the square.
The combination of lower than usual temperature of the CS (significance $\sim 0.01$) and larger than usual shifts towards large scales in the acoustic peaks of the patch ($1.1 \times 10^{-4}$) results in a 4.73$\sigma$ (significance $p\sim1.11\times 10^{-6}$) observation against the $\Lambda$CDM model. With another 5$\sigma$ significance large underdense region in the patch, there are 3 different anomalies in one patch that requires explanation. One should note that voids (either multiple voids or a supervoid) can’t shift the acoustic peaks.
Here we propose one of the possible accounts that can explain simultaneously for these anomalies: there is some localized unknown energy in the transverse direction to stretch the space around the CS region. When space is stretched, the wavelength of sinusoidal waves are elongated so that the peaks in the harmonic domain are shifted towards larger scales, and the number density of photons is reduced, while redshifting reduces their frequency, so that the temperature is dragged down. Such stretching of space could also make the density at low redshift lower. If the unknown energy is from dark energy, this implies the existence of inhomogeneity of the dark energy[@deputter; @blomqvist08; @blomqvist10].
This work is supported by ROC MoST grant 103-2112-M-001-028. The author acknowledges the use of ESA Planck Legacy Archive data [^2], [^3] package [@healpix] and [^4] package [@glesp]. The author would like to thank D.Spergel, P.Naselsky, A.Lewis, O.Doré and S.Suyu for useful discussions and suggestions, also Ya-Shiuan Lin for helping to produce top panel of Figure \[distance\].
Non-periodic boundary (NPB) correction for patches with beam convolution
========================================================================
Fourier transform presumes the signal is periodic such that the two ends are continuous, thus when one deals with real data it is inevitable to encounter jump discontinuity, which induces extra power in the power spectrum. Although the data are mostly discrete, the NPB effect nevertheless persists and should be taken into account as systematic error. A patch taken from a full-sky map has such intrinsic effect, which is uncorrectable because NPB is part of the morphology. The effect, however, is further aggravated when there is beam convolution. As the statistics and comparisons in this paper are made between beam-convolved data and simulations with best-fit $\Lambda$CDM model (without beam convolution), it is necessary to study the effect because the power spectrum of a beam-convolved patch has more induced power than the smoothed power spectrum of the non-periodic patch.
To estimate the correction, we simulate 50 full-sky CMB maps and take 30 deg$^2$ patches centered at the equator (in total 600 patches). We take the central 20 deg$^2$ part directly from the 30 deg$^2$ patches, and denote their mean power spectrum as $C_\ell^A$. Then we use 5 arcmin to convolve the 30 deg$^2$ patches and choose the central 20 deg$^2$ part, the mean power spectrum of which is denoted $C_\ell^B$. The correction is therefore $C_\ell^B - b^2 C_\ell^A$, where $b$ is the beam transfer function. One finds that the extra power due to 5-arcmin beam convolution is a rather smooth curve and can be best-fitted to be $C^{\rm NPB}_\l=1.21\, \ell^{-1.38}$ ($\mu$K$^2$) from $\ell \geq 300$. And for 143GHz band map with beam transfer function 7.30 arcmin [@planckhfibeam] the correction for NPB is $C^{\rm NPB}_\l=34.4\, \ell^{-1.80}$ ($\mu$K$^2$).
Cross-power spectrum on the data
================================
Throughout the paper we use cross-power spectrum (XPS) on the data : 4 CMB maps, 143GHz channel map, and the foreground-subtracted maps at 100, 143, and 217 GHz from the SEVEM method, to extract the peak position. The XPS is a quadratic estimator between two maps that can provide unbiased estimate of the underlying power spectrum of the correlated signals, and, at the same time, reduce the uncorrelated ones. We take from Legacy Archive the FULL MISSION RINGHALF1 and FULL MISSION RINGHALF2 maps for SMICA, NILC, SEVEM, Commander and 143GHz map. Since the patches from RINGHALF1 (denoted $\alpha$) and RINGHALF2 (denoted $\beta$) contain the same CMB signal but uncorrelated noise, we can employ the XPS on their Fourier modes $\alpha_{{\mbox{\boldmath $k$}}} $ and $\beta_{{\mbox{\boldmath $k$}}}$
$$S_k^{\alpha\beta}= \frac{1}{2} \langle \alpha^{*}_{{\mbox{\boldmath $k$}}} \beta_{{\mbox{\boldmath $k$}}} + \beta^{*}_{{\mbox{\boldmath $k$}}} \alpha_{{\mbox{\boldmath $k$}}} \rangle,$$
where $*$ denotes complex conjugate and the angle brackets denote average over all $k-1/2 \le |{{\mbox{\boldmath $k$}}}| < k+1/2$ for integer k. However, as @xps [@direct] pointed out, lack of an ensemble and non-zero chance correlation between the so-called “uncorrelated” signal results in some residual $X_k^{\alpha\beta}$ from the XPS, which has the following relation, $$\sqrt{\langle (\xk^{\alpha\beta})^2 \rangle} \simeq \frac{\sqrt{ N^\alpha_k N^\beta_k } }{\sqrt{2\pi k}},$$ where $N^\alpha_k$ and $N^\beta_k$ are the (uncorrelated) noise power spectrum from RINGHALF1 and RINGHALF2, respectively. The noise power spectrum can be obtain by $1/2$ of the differencing of HALFRING1 and HALFRING2 patches, assuming $N^\alpha_k$ and $N^\beta_k$ are of the same level. Thus in our data pipeline, we subtract the residual $X_k^{\alpha\beta}$ from the XPS, after the correction of NPB.
Foreground residual estimate
============================
The foreground residual (or any unknown residual or systematics) can take the form of $ B \ell^{-\lambda}$ after foreground cleaning such that we can model the residual and an acoustic peak at $\ell_0$ with amplitude $A$ and spread $\sigma$ by $D_\ell=A \exp[-(\ell-\ell_0)^2/2\sigma^2]+ B \ell(\ell+1) \ell^{-\lambda} \exp[\ell(\ell+1)b^2]/2\pi $, where $b$ is the beam transfer function. As deconvolution of the beam transfer function varies slowly (for the multipole range in our discussion) we assume $\exp[\ell(\ell+1)b^2]\simeq w$. Letting $\beta\equiv \pi A/\sigma^2$ and $\epsilon \equiv B/\beta$, one can solve for the peak position for $\lambda=0$, 1, 2: $\ell\simeq \ell_0 (1-\epsilon w)^{-1}$, $\ell \simeq \ell_0 + \epsilon w/ 2$, and $\ell \simeq \ell_0$, respectively. Further assuming the peak position not far from $\ell_0$ for $\lambda=3$, $\ell\sim\ell_0-\epsilon w/2\ell_0^2$.
For $\lambda=0$ the peak position is shifted by $\ell_0 \epsilon w (1-\epsilon w)^{-1}$, which could be large enough to be the source of the anomaly. One should note, however, that if there is any residual with $\lambda=0$ (flat spectrum), it should be mixed with the pixel noise and is duly subtracted in our data processing. For $\lambda=1$, the peak shift $\epsilon w/2$ is inversely proportional to $\beta$, where $\beta$ is 1.79, 0.576, 0.446 for 1st, 2nd and 3rd peak of best-fit $\Lambda$CDM model, respectively. For the shift to be significant, $B$ has to be of the same order of $\beta$. Such high level of residual $B$, together with $(\ell+1) \exp[\ell(\ell+1)b^2]$ would have caused the power to shoot up at high $\ell$, but in reality the power spectrum of the Cold Spot patch in the middle panel of Figure \[coldspot\] shows no such shoot-up at all. For $\lambda=2$ the peak position is mostly static as the $\ell(\ell+1)$ cancels out the residual.
The case of $\lambda=3$ ($\lambda > 2$) is different from the previous cases in that the power index of the residual part in $D_\ell$ is negative ($\sim \ell^{-1}$). The peak shift, inversely proportional to $\ell_0^2$, is less than unity even when $B$ is as high as 9000, making the residual at the same level of best-fit $\Lambda$CDM model quadrupole.
Although the foreground residual is unlikely to have power law with integer index, the analytical forms display the following trend: for $\lambda < 2$ the resultant $D_\ell$ is dominated by $B \ell^{2-\lambda} \exp[\ell(\ell+1)b^2]$ at high $\ell$. In the data, in particularly the patch containing the Cold Spot, the shoot-up is not seen at high $\ell$, indicating the residual coefficient $B$ has to be quite small to counter the deconvolution and $\ell^{2-\lambda}$. The peak shift is thus negligible. For $\lambda >2 $, on the other hand, the peak shift is approximately proportional to $ \epsilon w/\ell_0^{\lambda-1}$. The $B$ will have to be much higher than the best-fit $\Lambda$CDM model at low $\ell$ to have any significant peak shift, which again is not seen in the power spectrum in the middle of Figure \[coldspot\], either.
natexlab\#1[\#1]{}
å[A&A]{}
[99]{} [[de Bernardis, P., et al.,]{}]{} [2000, , 404, 955]{}
[[Blanchard, A.,]{}]{} [[Schneider, J.,]{}]{} [1987, å, 184, 1]{}
[[Blomqvist, M.,]{}]{} [[Mörtsell, E.,]{}]{} [[Nobili, S.,]{}]{} [2008, , 6, 27]{}
[[Blomqvist, M.,]{}]{} [[Enander, J.,]{}]{} [[Mörtsell, E.,]{}]{} [2010, , 10, 18]{}
[[Challinor, A.,]{}]{} [[Lewis, A.,]{}]{} [2002, , 65, 103001]{}
[[Challinor, A.,]{}]{} [[Lewis, A.,]{}]{} [2005, , 71, 103010]{}
[[Chiang, L.-Y.,]{}]{} [[Chen, F.-F.,]{}]{} [2011, , 738, 188]{}
[[Chiang, L.-Y.,]{}]{} [[Chen, F.-F.,]{}]{} [2012, , 751, 43]{}
[[Cole, S.,]{}]{} [[Efstathiou, G.,]{}]{} [1989, , 239, 195]{}
[[Cruz, M.,]{}]{} [[Martínez-González, E.,]{}]{} [[Vielva, P.,]{}]{} [[Cayón, L.,]{}]{} [2005, , 356, 29]{}
[[Cruz, M.,]{}]{} [[Tucci, M.,]{}]{} [[Martínez-González, E.,]{}]{} [[Vielva, P.,]{}]{} [2006, , 369, 57]{}
[[de Putter, R.,]{}]{} [[Huterer, D.,]{}]{} [[Linder, E.V.,]{}]{} [2010, , 81, 103513]{}
[[Doroshkevich, A.G., et al.,]{}]{} [2005, , 14, 275]{}
[[Finelli, F., et al.,]{}]{} [2016, , 455, 1246]{}
[[Górski, K.M., et al.,]{}]{} [2005, , 622, 759]{}
[[Halverson, N.W., et al.,]{}]{} [2002, , 568, 38]{}
[[Hanany, S., et al.,]{}]{} [2000, , 545, L5]{}
[[Hansen, M., et al.,]{}]{} [2012, , 426, 57]{}
[[Hinshaw, G., et al.,]{}]{} [2007, , 170, 288]{}
[[Hinshaw, G., et al.,]{}]{} [2013, , 208, 19]{}
[[Hu, W.,]{}]{} [[Dodelson, S.,]{}]{} [2002, , 40, 171]{}
[[Inoue, K.T.,]{}]{} [[Silk, J.,]{}]{} [2006, , 648, 23]{}
[[Inoue, K.T.,]{}]{} [[Silk, J.,]{}]{} [2007, , 664, 650]{}
[[Lewis, A.,]{}]{} [[Challinor, A.,]{}]{} [2006, , 429, 1]{}
[[Linder, E.V.,]{}]{} [1990, , 243, 353]{}
[[Jeong, D., et al.,]{}]{} [2014, , 89, 023003]{}
[[Mackenzie, R., et al.,]{}]{}
[[Marcos-Caballero, A.,]{}]{} [[Fernández-Cobos, R.,]{}]{} [[Martínez-González, E.,]{}]{} [[Vielva, P.,]{}]{} [2016, , 460, L15]{}
[[Martínez-González, E., et al.,]{}]{} [2002, , 336, 22]{}
[[Nadathur, S., et al.,]{}]{} [2014, , 90, 103510]{}
[[Naidoo, K.,]{}]{} [[Benoit-Levy, A.,]{}]{} [[Lahav, O.,]{}]{} [2016, , 459, L71]{}
[Planck Collaboratoin: Ade P.A.R. ,]{} [2014a, å, 571, A15]{}
[Planck Collaboratoin: Ade P.A.R. ,]{} [2014b, å, 571, A12]{}
[Planck Collaboratoin: Ade P.A.R. ,]{} [2014c, å, 571, A17]{}
[Planck Collaboratoin: Aghanim N. ,]{} [2014, å, 571, A27]{}
[Planck Collaboratoin: Adam R. ,]{} [2016a, å, 594, A7]{}
[Planck Collaboratoin: Adam R. ,]{} [2016b, å, 594, A9]{}
[Planck Collaboratoin: Aghanim N. ,]{} [2016, å, 594, A11]{}
[Planck Collaboratoin: Ade P.A.R. ,]{} [2016a, å, 594, A15]{}
[Planck Collaboratoin: Ade P.A.R. ,]{} [2016b, å, 594, A16]{}
[[Press, W.H., et al.,]{}]{} [1994, *Numerical Recipes in Fortran*, Cambridge University Press, 2nd Ed.]{}
[[Seljak, U.,]{}]{} [1996, , 463, 1]{}
[[Scott, P.F., et al.,]{}]{} [2003, , 341, 1076]{}
[[Szapudi, I., et al.,]{}]{} [2014, [*Proceedings of 49th Rencontres de Moriond Conference Cosmology 2014*]{}, edited by E. Augé, J. Dumarchez and J. Tran Thanh Van, 33]{}
[[Szapudi, I., et al.,]{}]{} [2015, , 450, 288]{}
[[Valkenburg, W.,]{}]{} [2012, , 01, 47]{}
[[Vielva, P., et al.,]{}]{} [2004, , 609, 22]{}
[[Zhao, W.,]{}]{} [2013, , 433, 3498]{}
[^1]:
[^2]: http://pla.esac.esa.int/pla/\#home
[^3]: http://healpix.jpl.nasa.gov/
[^4]: http://www.glesp.nbi.dk/
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: |
We explore in some detail the scenario proposed to explain the observed knee of the cosmic ray (CR) spectrum as due to the effects of photodisintegration of the CR nuclei by interactions with optical and soft UV photons in the source region. We show that the photon column densities needed to explain the experimental data are significantly lower than those obtained in previous estimations which neglected multinucleon emission in the photodisintegration process. We also treat more accurately the photodisintegration thresholds, we discuss the effects of photopion production processes and the neutron escape mechanism, identifying the physical processes responsible for the qualitative features of the results. This scenario would require the CR nuclei to traverse column densities of $\sim 5 \times 10^{27}-
2 \times 10^{28}$ eV/cm$^2$ after being accelerated in order to reproduce the observed knee, and predicts that the CR composition should become lighter above $\sim 10^{16}$ eV.
author:
- |
Julián Candia, Luis N. Epele and Esteban Roulet\
[*Phys. Dept., U. of La Plata, CC67, 1900, La Plata, Argentina.*]{}
title: Cosmic ray photodisintegration and the knee of the spectrum
---
.5cm .5cm 21truecm 15truecm
Introduction
============
It is widely accepted nowadays that cosmic rays (CRs) with energies per particle up to about $10^{18}$ eV are protons and nuclei of galactic origin. Furthermore, it is well established that the full CR energy spectrum has a power-law behavior with a steepening taking place at the so-called knee, corresponding to an energy $E_{knee}=3\times 10^{15}$ eV. Although it is well known that the CR composition below the knee has a significant heavy component, its behavior beyond the knee remains somewhat controversial. Indeed, above $10^{14}$ eV/nucleus the information about the CR mass composition has to be drawn from extensive air shower observations at ground level with the results from different experiments not always being compatible. Moreover, the theoretical predictions depend on the hadronic model adopted and this can introduce further uncertainties on the inferred composition. The model dependence in certain observables can however be useful to discriminate among the hadronic models using the experimental data, and in this comparison QGSJET and VENUS turn indeed to be favoured against other models such as SIBYLL [@fowler]. Thus, the CR mass composition beyond the knee is not definitively established yet, with some observations [@swordy] suggesting that it turns lighter and others instead suggesting that the heavier components become dominant [@amen1; @ag00; @ka01; @ar00].
The attempts made so far to explain the physical origin of the knee of the cosmic ray spectrum can be roughly classified into three kind of models. One of them exploits the possibility that the acceleration mechanisms could be less effective above the knee [@fich; @joki; @koba], while a second one assumes that leakage from the Galaxy plays the dominant role in suppressing CRs above the knee [@syro; @wdow; @ptus]. The third scenario, originally proposed by Hillas [@hillas], considers nuclear photodisintegration processes (and proton energy losses by photomeson production), in the presence of a background of optical and soft UV photons in the source region, as the main responsible for the change in the CR spectrum. The first two scenarios are based on a rigidity dependent effect which consequently produces a change in the spectral slope of each nuclear component with charge $Z$ at an energy $\sim E_{knee}Z$. Hence, it removes first the protons and the lighter nuclei and only at larger energies affects the heavier nuclei, predicting that the CR mass composition should become heavier beyond the knee. On the contrary, the third scenario predicts a lighter CR composition above the knee due to the disintegration of the propagating nuclei. The predictions of this last scenario have been worked out by Karakula and Tkaczyk [@karak] some years ago. The aim of the present paper is to re-evaluate in detail the propagation of nuclei and protons in the source region following their approach, but taking into account some additional features that turn out to be important for the determination of the source parameters required for the scenario to work. In particular, we re-evaluate the nuclear photodisintegration rates taking into account multinucleon emission processes, we introduce more accurate threshold energies in the giant dipole resonance for every stable nucleus (according to recent remarks by Stecker and Salamon [@steck1]), and discuss the impact of the neutron mechanism [@neutron] which would allow the escape of neutrons from the source without further energy losses. Moreover, we show that by introducing a non negligible lower cutoff to the power-law photon distribution, the abrupt suppression of the all-particle CR spectrum above $10^{17}$ eV previously obtained [@karak] is prevented, and hence the underlying reason for that effect is identified. Let us also mention that although we focus our study here on the scenario proposed to explain the observed steepening at the knee, we expect that our general considerations about the treatment of photodisintegration processes should also be relevant in other contexts.
The propagation of cosmic rays
==============================
The source and the photon background spectra
--------------------------------------------
As already mentioned, we deal with a model that considers that the CR nuclei are accelerated inside discrete sources which are surrounded by a very strong background of optical and soft UV photons. If the typical energy of the photons is in the optical range ($1-10$ eV), the photodisintegration of CR nuclei will start to be efficient at CR energies $E \geq A \times 10^{15}$ eV. This is so because in the CR rest frame the optical photon (which is boosted by a relativistic factor $\gamma \geq 10^6$) appears as an energetic gamma ray with $E \geq 1-10$ MeV, i.e. capable of photodisintegrating the nucleus. Hence, the basic assumption is that the CR spectrum emitted by the source is a featureless extrapolation of the spectra measured below the knee and that the observed change in slope is actually due to the CR interactions (i.e. the process of nuclear photodisintegration and the photopion production by protons) with the surrounding photon background.
We then assume that the source emits nuclei with mass number $A$ (with $A$ ranging from 1 to 56). Since there is only one stable isotope for a nucleus of a given mass $A$ along most of the decay chain from $^{56}$Fe to $^{1}$H, we take for definiteness a unique charge $Z$ associated to any given value of $A$. The differential fluxes emitted by the source are assumed to be given by power-law distributions
$$\phi_i^0 (E) = \Phi_i^0 E^{-\gamma_{i}}$$
(where $i=56-A$ throughout). The intensities $\Phi_i^0$ and spectral indices $\gamma_{i}$ were taken from the detailed knowledge about CR mass composition below the knee (for energies per particle above $\sim few \ \ Z \times 10^{10}$ eV) [@wiebel].
For the photon background that surrounds the source we consider that its spectrum follows either a (thermal) Planckian-type distribution given by
$$n(\epsilon) = {\rho \over 2 \zeta (3) (k_{B}T)^{3}}
\frac{\epsilon^{2}}{\exp(\epsilon/k_{B}T)-1}$$
where $k_{B}$ is the Boltzmann constant, $T$ the absolute temperature and $\zeta$ the Riemann zeta function (such that $\zeta (3) \approx 1.202$), or a power-law distribution given by
$$n(\epsilon) = \rho (\alpha-1)
\left(\frac{1}{\epsilon_{m}^{\alpha-1}}-\frac{1}{\epsilon_{M}^{\alpha-1}}\right)^{-1} \epsilon^{-\alpha}$$
where $\epsilon_{m}$, $\epsilon_{M}$ are the lower and upper energy cutoffs and $\alpha$ the assumed spectral index. For both distributions, $\rho$ represents the total number density and $n(\epsilon)$ the differential number density corresponding to a photon energy $\epsilon$. Defining $\rho_E$ as the total energy density, one has for the Planckian spectrum
$$\rho_E \approx 2.631 k_BT \rho \ \ ,$$
while the analogous expression for the power-law spectrum is
$$\rho_E = {\alpha - 1 \over 2- \alpha } \left( \epsilon_M^{2-\alpha} -
\epsilon_m^{2-\alpha} \right) \left( {1 \over \epsilon_m^{\alpha - 1}} -
{1 \over \epsilon_M^{\alpha - 1}} \right)^{-1} \rho \ \ .$$
We adopt $\alpha =1.3$ as in ref. [@karak], but the introduction of a non negligible lower cutoff $\epsilon_{m}$ in the power-law spectrum excludes the presence of an abundant number of low energy (infrared) photons that would otherwise play the dominant role in the photodisintegration above $10^{17}$ eV, as was the case in ref. [@karak]. Regarding the spatial dependence of the radiation background we do not make any particular assumption, but the results will depend only on the integrated photon column density along the trajectory of the CR since its production at the source until it leaves the region filled with radiation.
Propagation of nuclei: photodisintegration
------------------------------------------
The main mechanism of energy loss for nuclei with energy per particle below $10^{18}$ eV propagating through these photon backgrounds is the process of photodisintegration. In fact, a nucleus of mass $A=56-i$ and Lorentz factor $\gamma=E/Am_pc^2$ that propagates through a photon distribution $n(\epsilon)$ (as, for example, those of eqs. (2) and (3)) has a probability of fission with emission of $j$ nucleons given by
$$R_{ij}(E)={1\over 2\gamma^2}\int_{\epsilon'_{thr,ij}/2\gamma}^\infty
{\rm d}\epsilon \ \
{n(\epsilon)\over \epsilon^{2}}\int_{\epsilon'_{thr,ij}}^{2\gamma\epsilon}
{\rm d}\epsilon'\epsilon'\sigma_{ij}(\epsilon'),$$
where $\sigma_{ij}$ is the corresponding photodisintegration cross section, $\epsilon$ the photon energy in the observer’s system and $\epsilon'$ its energy in the rest frame of the nucleus. In order to calculate $R_{ij}$ we fitted $\sigma_{ij}$ with the parameters given in Tables I and II of ref. [@puget], while the reaction thresholds $\epsilon'_{thr,ij}$ were taken from Table I of ref. [@steck1]. A useful quantity to take into account all reaction channels is the effective emission rate given by
$$R_{i,eff}=\sum_{j\geq 1}jR_{ij} \ \ .$$
At low energies ($\epsilon'_{thr,ij} \leq \epsilon' \leq 30$ MeV), the cross section $\sigma_{ij}(\epsilon')$ is dominated by the giant dipole resonance and the photodisintegration proceeds chiefly by the emission of one or two nucleons. Note that, according to recent remarks pointed out in ref. [@steck1], we employ threshold energies that depend both on the nucleus and on the number of emitted nucleons. The improvement consists in shifting the threshold energy, that was previously assumed to be $\epsilon'_{thr}=2$ MeV for all reaction channels and all nuclei [@karak; @puget], to higher energies, such that single-nucleon emission has now a typical threshold of $ \sim 10$ MeV, while the double-nucleon emission energy threshold becomes typically $ \sim 20$ MeV. Although the appearance of the knee in this scenario is ultimately a threshold effect, the improved calculation including the change in the specific threshold energies has however a minor impact on the resulting emission rates when compared with those obtained with a fixed threshold energy $\epsilon'_{thr} = 2$ MeV. This can be appreciated in Figure 1, where plots of $R_{i,eff}/\rho_{E}$ vs energy per nucleon $E^*$ are shown for a $^{56}$Fe nucleus ($i=0$) propagating through a thermal photon spectrum (with $k_{B}T=10$ eV). The emission rate $R_{0,eff}$ (calculated with the specific threshold energies) shows a small shift towards higher energies with respect to that obtained by means of the fixed threshold energy (labeled $R_{0,eff}^{(2)}$). The reason for this reduced effect of the threshold shifts is that the dominant photodisintegration effects are due anyhow to the energies around the giant resonance peak, which for single-nucleon emission are typically peaked around $ \sim 20$ MeV and for double-nucleon emission around $ \sim 26$ MeV, and hence clearly well above the threshold energies.
[=3.0in ]{}
At higher energies, $\sigma_{ij}(\epsilon')$ is approximately flat and the multinucleon emission acquires a higher probability, becoming actually dominant for heavy nuclei, for which the probability for single nucleon emission is of only $10 \%$. To estimate the relevance of multinucleon emission processes, we compare the effective rate from eqs. (6) and (7) with the emission rate that neglects multinucleon processes (as calculated in ref. [@karak]). Figure 2(a) shows plots of $R_{i,eff}/\rho_{E}$ versus energy per nucleon $E^*$ for a $^{56}$Fe nucleus ($i=0$) propagating through a thermal photon spectrum (with $k_{B}T=1.8$ eV). From the figure it becomes evident that multinucleon processes play an important role, increasing the emission rate at high energies by a factor of $\sim 4$ with respect to the single nucleon emission results. Analogously, Figure 2(b) shows the corresponding plots for a power-law photon spectrum with upper energy cutoff $\epsilon_M = 20$ eV and spectral index $\alpha = 1.3$. From this figure we can also see the effect of introducing a non negligible lower cutoff in the spectrum. In fact, we observe that with a negligible cutoff (e.g. $\epsilon_m = 10^{-6}$ eV) the emission rate grows steadily, because increasingly abundant low energy (infrared) photons give in this case the dominant contribution to the rates. They also keep the influence of the giant resonance dominant even at high particle energies, and hence in this case the multinucleon emission has little impact. However, if we take a non negligible cutoff into account (for instance, $\epsilon_{m}=0.8$ eV), we find that the flat high energy regime of $\sigma_{i}$ now dominates beyond a given value of $E^*$ and hence the emission rate saturates at high energies. Comparing the two emission rate curves with non negligible cutoff, we observe again that multinucleon emission processes lead to a significant increase (by the same factor of $\sim 4$ we already encountered) in the emission rate. Thus, introducing a non negligible lower cutoff energy in the power-law photon spectrum the results should not differ much from those with the thermal photon spectrum and, being the rates with multinucleon emission larger, they should also require smaller photon densities than those previously estimated in ref. [@karak] to produce similar overall effects.
[=3.0in ]{}
[=3.0in ]{}
It should be noticed that in Figures 1 and 2 we have chosen to plot the emission rates as functions of the energy per nucleon $E^*$, instead of energy per nucleus $E$, because it is $E^*$ that remains constant during photodisintegration and also because for different nuclei the shapes of the cross section are similar with the maxima occurring approximately at the same values of $E^*$.
The equations that describe the propagation of nuclei are:
$$\frac{\partial \phi_{i}(E^{*},x)}{\partial x} = -\phi_{i}(E^{*},x)
\sum_{j\geq 1} R_{ij}(E^{*}) +
(1-\delta_{0i}) \sum_{j=1}^{i} R_{(i-j)j}(E^{*})
\phi_{i-j}(E^{*},x),$$
where $\phi_{i}(E^{*},x)$ is the differential flux of a nucleus of mass $A=56-i$ (for $0 \leq i \leq 54$) with an energy per nucleon $E^{*}$ at propagation distance $x$ from the source, $R_{ij}(E^{*})$ is the above defined emission rate and $\delta_{ij}$ is the Kronecker delta. The exact solution of these equations (i.e., the differential fluxes at propagation distance $L$) is given by:
$$\phi_{i}(E^{*},L) = \sum_{j=0}^{i} b_{ij}(E^{*})
\exp (-\sum_{k\geq 1} R_{jk}(E^{*}) L) ,$$
where
$$b_{ii}=\phi_i^0 (E^{*})-(1-\delta_{0i})\sum_{j=0}^{i-1} b_{ij}(E^{*})$$
and
$$b_{ij} = {\sum_{k=1}^{i-j} b_{(i-k)j}(E^{*}) R_{(i-k)k}(E^{*})\over
\sum_{k \geq 1} R_{ik}(E^{*})-\sum_{k \geq 1} R_{jk}(E^{*})},\ \
{\rm for}\ \ i>j.$$
Recalling that the emission rates $R_{ij}$ are linear in the photon energy density $\rho_E$, it should be noted that, as expected, the fluxes only depend on the column density $\rho_E L$. It also should be pointed out that, since the equations describing the CR propagation only depend on the energies per nucleon $E^{*}$, the final CR spectra require the re-calculation of the solution (eqs.(9)-(11)) in terms of energy per nucleus.
Propagation of protons
-----------------------
The main energy loss mechanism that must be taken into account for the propagation of protons is the photomeson production process. We have to distinguish between the primary protons produced at the source with an initial flux $\phi_{i=55}^0$ (as given by eq.(1)) and the nucleons released as by-products of the nuclear photodisintegration. The fluxes corresponding to these (secondary) protons and neutrons ($\phi^{sp}$ and $\phi^n$, respectively) can be easily determined from the solution for the nuclear fluxes, as we will see in the following. Let $\phi_i^N (E,x)$ be the differential flux of nucleons of energy $E$ produced within a distance $x$ away from the source by the photodisintegration of nuclei of mass number $A=56-i$ and energy per nucleon $E^{*}=E/A$. Then, the equation governing the evolution of the fluxes is
$$\frac{\partial \phi_i^N(E,x)}{\partial x} =
\phi_{i}(E^{*},x) \sum_{j \geq 1} j R_{i,j}(E^{*})
\ \ , \ \ {\rm for}\ 0 \leq i \leq 54 \ \ ,$$
(with the initial condition $\phi_i^N (E,x=0) = 0$) and its solution reads
$$\phi_i^N (E,L) = \sum_{j=0}^{i}
\left(\frac{\sum_{k \geq 1} k R_{ik}(E^{*})}{\sum_{k \geq 1}
R_{jk}(E^{*})}\right)
b_{ij}(E^{*})
\left[ 1-\exp \left(-\sum_{k\geq 1} R_{jk}(E^{*}) L\right) \right] \ \ .$$
Thus, the total fluxes of secondary protons and neutrons are found by performing the weighted sums over index $i$:
$$\phi^{sp}(E,L) = \sum_{i=0}^{54} \frac{Z_i}{56-i} \phi_i^N (E,L) \ \ ,$$
$$\phi^{n}(E,L) = \sum_{i=0}^{54} \left(1 -\frac{Z_i}{56-i}\right)
\phi_i^N(E,L) \ \ ,$$
where $Z_{i}$ is the charge of the $i$-nucleus.
We will also discuss the possible impact of the so-called neutron mechanism[@neutron], which could take place in the presence of a magnetic field in the source region. Indeed, a magnetic field would retain the charged particles within that region for some period of time, while non-charged particles would be able to escape rapidly. In that case, we will consider that the flux of neutrons $\phi^n$ adds to the final all-particle CR flux neglecting any energy losses, while the flux of secondary protons $\phi^{sp}$, combined with the initial flux of primary protons $\phi_{i=55}^0$, will undergo energy losses due to the well studied process of photopion production. Indeed, since a background photon looks like a high-energy gamma ray in the proton rest frame, the photopion reactions
$$p + \gamma \to p + \pi^{0}$$
$$p + \gamma \to n + \pi^{+}$$
proceed with an interaction probability given by
$$g(E)={1\over 2\gamma^2}\int_{\epsilon'_{thr}/2\gamma}^\infty
{\rm d}\epsilon
{n(\epsilon)\over \epsilon^{2}}\int_{\epsilon'_{thr}}^{2\gamma\epsilon}
{\rm d}\epsilon'\epsilon'\sigma_{\gamma p}(\epsilon')K(\epsilon'),$$
where, analogously to eq. (6), $\gamma=E/m_pc^2$ is the Lorentz factor of the proton, $\sigma_{\gamma p}$ is the $\gamma p$ interaction cross section (parametrized from experimental data compiled in ref.[@table]), $K$ is the coefficient of inelasticity (defined as the average relative energy loss of the proton), $\epsilon$ the photon energy in the observer’s system and $\epsilon'$ its energy in the rest frame of the proton. The energy threshold for photopion production is $\epsilon'_{thr}=145$ MeV.
Then, if we call $\phi^{p}(E,x)$ and $\phi^{*n}(E,x)$ the differential fluxes of protons and neutrons of energy $E$ (produced by the reactions given by eqs.(16) and (17), respectively) after tarversing a distance $x$ from the source, the corresponding propagation equations are
$$\frac{\partial \phi^{p}(E,x)}{\partial x} =
-g(E) \phi^{p}(E,x) + \frac{1}{2}\frac{1}{(1-K)} g\left(\frac{E}{1-K}\right)
\phi^{p}\left(\frac{E}{1-K},x\right)$$
and (neglecting neutron energy losses)
$$\frac{\partial \phi^{*n}(E,x)}{\partial x} =
\frac{1}{2}\frac{1}{(1-K)} g\left(\frac{E}{1-K}\right)
\phi^{p}\left(\frac{E}{1-K},x\right) \ \ ,$$
where the factors $\frac{1}{2}$ arise from the relative probability of occurrence of reactions (16) and (17). The coefficient of inelasticity $K$ was given a typical value $K=0.3$, irrespective of proton energy [@karak]. The initial conditions to solve these differential equations are
$$\phi^{p}(E,x=0) = \phi_{i=55}^0(E) + \phi^{sp}(E,L)$$
and
$$\phi^{*n}(E,x=0) = 0 \ \ ,$$
respectively. In the first one we assumed for simplicity that the secondary protons produced by photodisintegration of nuclei add directly to the initial proton spectrum. This is justified since whenever the photopion processes are relevant, the photodisintegration is very efficient and hence nuclei are rapidly disintegrated. The solution to eq.(19) turns out to be
$$\phi^{p}(E,L) = \sum_{l=0}^{\infty} \phi_l^p(E,L) \ \ ,$$
where
$$\phi_0^p(E,L) = \phi^{p}(E,0) \exp(-g(E) L) \ \ ,$$
while for $l>0$ $$\phi_l^p(E,L)= {\phi^{p}\left(\frac{E}{(1-K)^{l}},0\right)\over (2(1-K))^{l}}
\prod_{j=1}^{l} g\left(\frac{E}{(1-K)^{j}}\right)
\sum_{n=0}^{l} {\exp \left(-g\left(\frac{E}{(1-K)^{n}}\right) L\right)\over
\prod_{^{m=0} _{m \neq n}}^{l}
\left[
g\left(\frac{E}{(1-K)^{m}}\right)-g\left(\frac{E}{(1-K)^{n}}
\right)\right]} \ \ .$$
For the neutron component, the solution to eq.(20) reads
$$\phi^{*n}(E,x) = \sum_{l=0}^{\infty} \phi_l^{*n}(E,x) \ \ ,$$
where
$$\phi_0^{*n}(E,L)= \frac {\phi^{p}\left(\frac{E}{1-K},0\right)}{2(1-K)}
\left(1-\exp\left(-g\left(\frac{E}{1-K}\right) L\right)\right) \ \ ,$$
while for $l> 0$ $$\phi_l^{*n}(E,L)= \frac {\phi^{p}\left(\frac{E}{(1-K)^{l+1}},0\right)}
{(2(1-K))^{l+1}}
\prod_{j=0}^{l} g\left(\frac{E}{(1-K)^{j+1}}\right)\times$$ $$\sum_{n=0}^{l} {1-\exp\left(-g\left(\frac{E}{(1-K)^{n+1}}\right) L\right)
\over g\left(\frac{E}{(1-K)^{n+1}}\right)
\prod_{^{m=0}_{ m \neq n}}^l
\left[g\left(\frac{E}{(1-K)^{m+1}}\right)-g\left(\frac{E}{(1-K)^{n+1}}
\right)\right]}.$$
The linearity of the photopion reaction rate $g(E)$ with respect to the photon energy density $\rho_E$ (as implied by eq.(18)) causes the proton and neutron fluxes to depend only on the column density $\rho_E L$, similarly as for the fluxes of nuclei.\
Results and comparison with observations
========================================
So far, we wrote down the appropriate equations governing the propagation of CR particles, and we found their exact solutions: eqs.(9)-(11) for nuclei, eqs.(13),(15) and (26)-(28) for neutrons (that eventually decay into protons outside the source region), and eqs.(23)-(25) for protons (that undergo energy losses due to photopion production), the sum of all these contributions being the total CR flux $\phi_{total}$. We can now investigate how $\phi_{total}$ depends on the assumed photon distributions around the source, namely on the column density $\rho_E L$ and on the parameters involved in the photon spectral distributions (eqs.(2) and (3)).
Figure 3(a) shows the total CR differential flux as a function of the energy per particle for a Planckian spectrum with $k_{B}T=1.8$ eV and several values for the column density $\rho_E L$. Analogously, Figure 3(b) shows the results corresponding to a Planckian spectrum with $k_{B}T=10$ eV. To compare our results with experimental data, the figures also show the observed spectra measured by different experiments (CASABLANCA, DICE, Tibet, PROTON satellite and AKENO) [@fowler; @swordy; @amen2].
[=3.0in ]{}
[=3.0in ]{}
The effect of increasing the temperature is clearly that of shifting the knee towards lower energies, while the effect of increasing $\rho_E L$ is that of intensifying the steepening at the knee. This behavior is easily explained by the fact that it is the nuclear photodisintegration the responsible for the occurrence of the knee (since photopion production processes show up at higher energies). Then, if the photon temperature (and thus, the mean photon energy) is larger, the particle energy required for the nuclei to disintegrate decreases. Furthermore, CR propagation cannot be affected by $\rho_E L$ below the knee, where the photodisintegration has not set up yet, but above the knee the disintegration rates have a linear dependence on $\rho_E$. Hence, we should expect $\rho_E L$ to control the steepening in the knee region, in agreement with the results shown in the figures.
In ref. [@karak] the best fit to observations for the Planckian spectrum was that corresponding to $k_{B}T=1.8$ eV and $\rho_E
L=2.25\times10^{29}$ eV/cm$^2$. For the sake of comparison, Figure 3(a) shows our results for the same temperature. It is clearly observed that the new results are compatible with the AKENO experimental data for low column density values $\rho_E L\simeq 5\times 10^{27}$ eV/cm$^2$, i.e. $\sim 50$ times lower than the previous result. Figure 3(b) shows that the corresponding results for $k_{B}T=10$ eV are in somewhat better agreement with CASABLANCA observations for column densities in the range $\rho_E L=5\times 10^{27}-2\times 10^{28}$ eV/cm$^2$. One has to keep in mind however that the model discussed is oversimplified in many respects: we assume that the same photon column density is traversed by all nuclear species and for all CR energies, we neglect additional effects due to change in diffusion properties at these energies, which in alternative scenarios [@ptus] are actually the responsible for the appearance of the knee, we assume that CR interactions in the interstellar medium are negligible, etc.. Hence, an accurate agreement with observational data is not to be expected, although it is reassuring to see that the results have the right tendency to account for the main qualitative features of the data in the knee region.
Figures 4(a) and 4(b) show plots of mean mass composition $\langle \rm{ln} A \rangle$ versus $E$ corresponding to the Planckian photon spectra with $k_BT=1.8$ eV and $k_BT=10$ eV, respectively. For comparison, we also show the CR mass composition measured by different experiments [@fowler; @swordy; @ka01; @chat]. One can observe that solutions corresponding to the largest plotted values of column density $\rho_E L$ imply that CRs above the knee are chiefly constituted by protons, in disagreement with experimental data; the non-negligible contribution of heavier CR components beyond the knee imposes an upper limit on the column density around $(\rho_EL)_{max} \approx
2 \times 10^{28}$ eV/cm$^2$. This result stresses again the fact that multinucleon emission processes play a significant role in CR propagation within this scenario. We also see that, although the solutions with lower column density ($\rho_EL \leq (\rho_EL)_{max}$) are in reasonable agreement with some data sets within error bars, the composition beyond the knee seems to be somewhat suppressed. Nevertheless, we should bear in mind that we have employed just one source with a fixed column density, and that in a realistic situation there would be many sources, each with a different environment of radiation. Avoiding any detailed calculations, we may think about the contribution of one additional source. If it happens to have a column density $\rho_E L$ somewhat smaller, it would lead to a flatter flux; then, whereas the first source would be responsible for the occurrence of the knee, the second one would keep a non-vanishing nuclear component beyond the knee. Furthermore, in the computations we assumed that all nuclei traverse the same distance $L$, thus disregarding the fact that the magnetic field present in the source region should retain more effectively the heavier nuclei at any given CR energy. Taking this into account (for instance, assuming that a nucleus traverses a distance in the photon background proportional to its charge) we would find that, while the disintegration would be very effective for heavy nuclei (and would be, in fact, responsible for forming the knee), it would be less effective for the lighter nuclei (since the column density for them would be much smaller) and they could then survive beyond the knee. Hence, we see that this scenario can easily account for a change in the composition towards lighter nuclei above the knee, which is apparent in some of the existing data. On the contrary, rigidity dependent scenarios predict the opposite trend, and may have difficulties to account for some measurements. It has indeed been pointed out that they would require the introduction of an ad-hoc additional light component above the knee to reproduce the observations [@arm].
[=3.0in ]{}
[=3.0in ]{}
The results for the power-law photon spectrum (with a lower cutoff energy in the infrared range) should not differ much from the ones obtained for the Planckian spectrum, and this behavior was in fact corroborated. Let us fix the spectral index and the upper cutoff energy at the values found in ref. [@karak] to correspond to the best fit to experimental data, namely $\alpha=1.3$ and $\epsilon_M=20$ eV. Figure 5 shows $\phi_{total}$ versus $E$ for power-law distributions with column density $\rho_E L=10^{28}$ eV/cm$^{2}$ and several values for the lower cutoff energy ($\epsilon_m=0.01,0.1,0.8$ eV). It is clear that the overall spectrum displays a steepening which has the right qualitative features to explain the knee although it does not provide an accurate fit to neither AKENO nor CASABLANCA data individually. Turning now to discuss the dependence of the results on the lower cutoff energy, it is found that, apart from minor differences arising from spectrum normalization, lower values of $\epsilon_m$ yield lower fluxes in the high energy end. The explanation for this is that from the expression for the photodisintegration emission rate (eq.(6)) it turns out that the role of $\epsilon_m$ is that of determining a value of the particle energy $E$, such that below $E$ the giant dipole resonance dominates, while above $E$ the flat high energy regime of $\sigma$ prevails. As we lower $\epsilon_m$, more photons with lower energy become available so that the flux suppression imposed by the giant resonance extends to higher particle energies.
[=3.0in ]{}
This analysis provides a suitable explanation of the results shown in Figure 5, as well as those of Figure 2(b). The results previously obtained for power-law spectra [@karak] show an extremely abrupt flux suppression that seems to be in contrast with experimental data (CR energy spectra and, especially, CR mass composition) above $E=10^{17}$ eV, and can be attributed to the negligible value for $\epsilon_m$ adopted there. The use of a non negligible cutoff prevents the flux suppression and provides a better fit to observations in this energy region. It should also be remarked that a solution yielding an excessive CR flux at energies above $10^{17}$ eV is by no means troublesome since it is clear that at these high energies efficient leakage of CRs from the Galaxy cannot be disregarded, and that a less effective acceleration could also be playing a role. These mechanisms would clearly contribute to lower the CR flux, so that it does not seem reasonable to expect a reliable explanation of the CR spectrum by photodisintegration processes alone beyond $E\sim 10^{17}$ eV.
[=3.0in ]{}
Figure 6 shows plots of the proton and neutron components ($\phi^p+
\phi^{sp}$ and $\phi^n + \phi^{*n}$, respectively) as well as the total CR flux $\phi_{total}$ for the Planckian photon spectrum with $k_BT=1.8$ eV and $\rho_E L=2 \times 10^{29}$ eV/cm$^2$. In order to estimate the possible effects of the neutron mechanism, we also plotted the total CR flux that is obtained ignoring the escape of neutrons, labelled $\phi_{total}^*$. From the figure it is seen that the neutron mechanism raises the CR total flux by $\sim 100 \%$ for $E \geq 5 \times 10^{16}$ eV. Had we adopted a column density $\rho_E L=10^{29}$ eV/cm$^2$, the corresponding flux increment would be only $\sim 10 \%$. Nevertheless, it should be noted that these results correspond to very high column densities (actually inconsistent with the observed CR composition), and that effects arising from the neutron escape mechanism turn out to be negligible for lower values of the column density $\rho_EL$, for which also the photopion losses of protons can be altogether ignored.
To see what are the implications for the sources in order to produce the required photon densities for this scenario to work, let us make the following simple estimate. An isotropic source with photon energy density $\rho_{E,0}$ at its surface (with radius $r_0$) will have around it a photon energy density given by $\rho_E(r)=\rho_{E,0}(r_0/r)^2$. This means that the photon column density measured radially from the surface is $\langle \rho_E r\rangle\simeq \rho_{E,0} r_0$. On the other hand, if the CRs are magnetically confined near the source, the path they travel across the dense photon field will be larger, i.e. $\langle \rho_E L\rangle\equiv \lambda \langle \rho_E r\rangle$, with the parameter $\lambda\gg 1$. If to obtain an estimate of the maximum achievable densities we assume that the source radiates with a luminosity $L=4\pi r_0^2\rho_{E,0} c$ smaller than the Eddington limit, i.e. $L<L_{Edd}=4\pi GMm_pc/\sigma_T\simeq 1.3\times
10^{38}M/M_\odot$ erg/s, with $M$ the mass of the source, $m_p$ the proton mass and $\sigma_T$ Thomson’s cross section, we find that the radius of the source needs to satisfy $$r_0<\frac{GMm_p}{\sigma_T\rho_{E,0}r_0}\simeq \left(\frac{\lambda}
{10}\right)\left(\frac{10^{28}{\rm eV/cm}^2}{\langle \rho_E L\rangle}\right)
\left( \frac{M}{M_\odot}\right)2\times 10^9 {\rm m},$$ and hence the photon source must be compact (like a pulsar) and in general smaller than the typical size of a white dwarf. It has to be mentioned however that in non-steady-state situations like in supernova explosions, luminosities larger than $L_{Edd}$ are possible. Another consideration is that if the source has a black-body spectrum with temperature $T$ and it is radiating at the Eddington limit, the relation $4\pi r_0^2\sigma T^4=L_{Edd}$ implies that $T\simeq 1.8\times 10^3$ eV$ (M/M_\odot)^{1/4}(10$ km/$r_0)^{1/2}$, and hence for the photon energies to be in the optical/UV region, radii somewhat larger than those of neutron stars would be favored (or the luminosity should be smaller than $L_{Edd}$).
As a summary, we have considered in more detail the photodisintegration of CR nuclei in the scenario in which their interaction with optical and soft UV photons around the source is responsible for the steepening of the spectrum and for the change in composition above the knee. The more accurate treatment of the nuclear processes, in particular the inclusion of the multinucleon emission, implies that the required photon column densities are lower (by an order of magnitude) than previously obtained. The main difference between this scenario and other (rigidity dependent) explanations proposed for the knee is the resulting CR composition, which here becomes lighter above $\sim 10^{16}$ eV.
Acknowledgments {#acknowledgments .unnumbered}
===============
Work supported by CONICET, ANPCyT and Fundación Antorchas, Argentina. E. R. thanks A. Dar for fruitful discussions on the subject.
[99]{}
J.W. Fowler et al., astro-ph/0003190.
S.P. Swordy and D.B. Kieda, Astropart. Phys. **13** (2000) 137.
M. Amenomori et al. (Mt Fuji Emulsion Chamber Collaboration), in: Proc. 19th Int. Cosmic Ray Conf., La Jolla (1985) Vol.2, p. 206.
M. Aglietta et al. (EAS-TOP Collaboration), Nucl. Phys. Proc. Suppl. [**85**]{} (2000) 318.
K.H. Kampert et al. (Kascade Collaboration), OG.1.2.11 Proc. 26th ICRC (Salt Lake City, 1999); K.H. Kampert et al., astro-ph/0102266.
F. Arqueros et al. (HEGRA Collaboration), Astron. and Astrophys. [**359**]{} (2000) 682.
C.E. Fichtel and J. Linsley, ApJ **300** (1986) 474.
J.R. Jokipii and G.E. Morfill, ApJ **312** (1986) 170.
K. Kobayakawa et al., astro-ph/0008209.
S.I. Syrovatsky, Comm. Astrophys. Space Phys. **3** (1971) 155.
J. Wdowczyk and A.W. Wolfendale, J. Phys. G. **10** (1984) 1453.
V.S. Ptuskin et al., Astron. Astrophys. **268** (1993) 726.
A.M. Hillas, in: Proc. 16th Int. Cosmic Ray Conf., Kyoto (1979) Vol.8, p. 7.
S. Karakula and W. Tkaczyk, Astropart. Phys. **1** (1993) 229.
F.W. Stecker and M.H. Salamon, ApJ **512** (1999) 521.
V. S. Berezinsky and O. F. Prilutsky, Sov. Astron. Lett. [**3**]{} (3) (1977) 141; V.S. Berezinsky, Proc. 15$^{th}$ ICRC (Plovdiv, 1977) [**10**]{} p. 84.
B. Wiebel-Sooth et al., Astron. and Astrophys. **330** (1998) 389.
J.L. Puget et al., Astrophys. J **205** (1976) 638.
Rev. of Part. Phys., Eur. Phys. Jour. C **15** (2000).
M. Amenomori et al., ApJ **461** (1996), 408.
E. Chatelet et al., in: Proc. 22nd Int. Cosmic Ray Conf., Dublin (1991) Vol.2, p.45.
S.V. Ter-Antonyan and L.S. Haroyan, hep-ex/0003006.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'Deep learning models are often trained on datasets that contain sensitive information such as individuals’ shopping transactions, personal contacts, and medical records. An increasingly important line of work therefore has sought to train neural networks subject to privacy constraints that are specified by differential privacy or its divergence-based relaxations. These privacy definitions, however, have weaknesses in handling certain important primitives (composition and subsampling), thereby giving loose or complicated privacy analyses of training neural networks. In this paper, we consider a recently proposed privacy definition termed *$f$-differential privacy* [@dong2019gaussian] for a refined privacy analysis of training neural networks. Leveraging the appealing properties of $f$-differential privacy in handling composition and subsampling, this paper derives analytically tractable expressions for the privacy guarantees of both stochastic gradient descent and Adam used in training deep neural networks, without the need of developing sophisticated techniques as [@deep] did. Our results demonstrate that the $f$-differential privacy framework allows for a new privacy analysis that improves on the prior analysis [@deep], which in turn suggests tuning certain parameters of neural networks for a better prediction accuracy without violating the privacy budget. These theoretically derived improvements are confirmed by our experiments in a range of tasks in image classification, text classification, and recommender systems.'
author:
- 'Zhiqi Bu[^1]'
- 'Jinshuo Dong[^2]'
- 'Qi Long[^3]'
- 'Weijie J. Su[^4]'
bibliography:
- 'ref.bib'
title: Deep Learning with Gaussian Differential Privacy
---
*University of Pennsylvania*\
November 25, 2019
Acknowledgments {#acknowledgments .unnumbered}
---------------
We are grateful to David Durfee, Ryan Rogers, Aaron Roth, and Qinqing Zheng for stimulating discussions in the early stages of this work. This work was supported in part by NSF through CAREER DMS-1847415, CCF-1763314, and CCF-1934876, the Wharton Dean’s Research Fund, and NIH through R01GM124111.
[^1]: Graduate Group in Applied Mathematics and Computational Science. Email: [[email protected]]{}.
[^2]: Graduate Group in Applied Mathematics and Computational Science. Email: [[email protected]]{}.
[^3]: Department of Biostatistics, Epidemiology and Informatics. Email: [[email protected]]{}.
[^4]: Department of Statistics. Email: [[email protected]]{}.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
author:
- Jakob Meineke
- 'Jean-Philippe Brantut'
- David Stadler
- Torben Müller
- Henning Moritz
- Tilman Esslinger
title: 'Interferometric Measurement of Local Spin-Fluctuations in a Quantum Gas'
---
=1
[**The subtle interplay between quantum-statistics and interactions is at the origin of many intriguing quantum phenomena connected to superfluidity and quantum magnetism [@Auerbach_Quantum_Magnetism]. The controlled setting of ultracold quantum gases is well suited to study such quantum correlated systems [@varenna_lectures]. Current efforts are directed towards the identification of their magnetic properties [@jordens_quantitative_2010; @Nascimbene2011; @sommer_universal_2011], as well as the creation and detection of exotic quantum phases [@lewenstein_ultracold_2007; @eckert_quantum_2008; @roscilde_quantum_2009]. In this context, it has been proposed to map the spin-polarization of the atoms to the state of a single-mode light beam [@bruun_probing_2009]. Here we introduce a quantum-limited interferometer realizing such an atom-light interface [@hammerer_quantum_2010] with high spatial resolution. We measure the probability distribution of the local spin-polarization in a trapped Fermi gas showing a reduction of spin-fluctuations by up to 4.6(3)dB below shot-noise in weakly interacting Fermi gases and by 9.4(8)dB for strong interactions. We deduce the magnetic susceptibility as a function of temperature and discuss our measurements in terms of an entanglement witness.**]{}
Quantum mechanics manifests itself in the correlations between the constituent parts of a physical system. These correlations quantify the probability of joint measurements, and are experimentally observable in the statistical distribution of the outcomes of repeated measurements. Experiments studying density fluctuations have successfully demonstrated the potential of these techniques as a tool to study local thermodynamic properties of quantum gases [@esteve_observations_2006; @gemelke_situ_2009; @muller_local_2010; @sanner_suppression_2010; @bakr_probing_2010; @sherson_single-atom-resolved_2010; @serwane_deterministic_2011]. In another context, interferometric methods have been used to study spin-fluctuations in atomic vapors, leading to the observation of entanglement and spin-squeezing [@hammerer_quantum_2010; @oblak_quantum-noise-limited_2005; @appel_mesoscopic_2009]. More recently, several authors have proposed to apply similar techniques to map quantum fluctuations, generated by the many-body dynamics in a quantum gas [@eckert_quantum_2008; @roscilde_quantum_2009; @bruun_probing_2009], on to the optical field of a single mode probe beam. In a different approach, speckle noise originating from out-of-focus regions in off-resonant imaging was related to the spin-fluctuations of a Fermi gas [@sanner_speckle_2011]. In this letter, we use a shot-noise limited interferometer to directly measure the probability distribution of the local spin-fluctuations in a two-component quantum degenerate Fermi gas.
![Interferometer Setup (a) Interferometer beams in the vicinity of the atomic cloud. While the probe passes through the cloud shown in grey, the local oscillator passes by the side of it. The beams (minimum $1/e^2$-radius $1.2\, \mu{\rm m}$) overlap in the far field giving rise to an interference pattern as shown. See methods for creation of the interferometer beams. (b) Optical setup to obtain two interference patterns, only one of which is affected by the atoms. Using a quarter-wave retardation plate ($\lambda/4$) and two polarizing beam splitters (PBS), the $\sigma^-$-component of the polarization, which interacts with the atoms, is separated from the $\sigma^+$-component. This yields two far-field interference patterns on one image as shown in the lower part of the figure, see methods. The lens adjusts the size of the patterns on the camera. (c) Level-scheme illustrating that only $\sigma^-$ light interacts with the atoms. $\sigma^+$ light is far detuned from resonance. Transitions from both states to their respective excited states are nearly closed cycling transitions.[]{data-label="fig1"}](fig1.pdf){width="50.00000%"}
Our interferometer is analogous to Young’s double slit experiment. Two tightly focused beams, the probe and the local oscillator, are focused to separate points as shown in Figure \[fig1\] and overlap in the far field. Position and visibility of the resulting interference pattern are determined by changes in phase and amplitude of the probe beam, which passes through an atomic cloud, while the local oscillator does not. The analysis of the interference pattern thus allows the reconstruction of both quadratures of the probe beam, phase and amplitude, which carry information about the local properties of the atomic cloud.
A probe beam passing through a mixture of $^6$Li atoms in the lowest two hyperfine states $|1\rangle$ and $|2\rangle$ acquires a phase-shift given by $$\phi=\sigma_0\left(\frac{n_1\delta_1}{1+s+4\delta_1^2}+\frac{n_2\delta_2}{1+s+4\delta_2^2}\right),
\label{eq1}$$ with $\sigma_0$ the resonant scattering cross-section, $s$ the saturation parameter and $n_i$ and $\delta_i$ the line-of-sight integrated density and the frequency detuning in atomic linewidths for state $|i\rangle$, respectively. By choosing the detuning exactly in between the two resonances, $\delta_1=-\delta_2=6.4$, the phase shift $\phi$ is proportional to the line-of-sight integrated spin-polarization density $m=n_1-n_2$. A single measurement of the phase-shift yields the spin-polarization of the given experimental realization. Consequently, the full probability distribution of spin-polarization can be reconstructed from repeated measurements.
![Interferometer Performance (a) Measured phase () and optical density () as a function of detuning. Maximum saturation of the probe beam $s=1.2$, duration of the probe pulses $1.2\, \mathrm{\mu s}$. Solid lines result from a fit to the phase data using the model described in the text yielding $n_2\sigma_0=3.2$, effective saturation $0.6$ and an effective linewidth that is 20% broader than the natural linewidth of $5.9\,\mathrm{MHz}$, which we attribute to the probe laser. Errorbars show standard deviations. (b) Measured phase-variance (without atoms) $\delta \phi^2$ () as a function of photon number in the probe beam determined from 100 measurements for each point. The empty circle () indicates the phase-variance for the intensity at which the spin-polarization measurements were made. Errorbars are an uncorrelated sum of statistical and systematic uncertainties. Expected phase-noise () for quantum efficiency $\eta=0.6$. The width of the line corresponds to 20% errors estimated from the uncertainty of our determination of $\eta$. The gray lines indicate the square of the phase shift expected for a single atom fixed in space at a detuning of half the atomic linewidth for the indicated $1/e^2$-waists of the probe beam [@aljunid_phase_2009]. []{data-label="fig2"}](fig2.pdf){width="50.00000%"}
To validate our procedure we show that (i) equation (\[eq1\]) holds in the parameter regime of the experiment, and (ii) that the phase measurement is only limited by photon shot-noise of the probe beam. Then, each additionally detected photon leads to a projection of the atomic state into a smaller subspace. To verify (i), we measure the frequency-dependent phase shift and optical density for a Fermi gas comprised of atoms in state $|2\rangle$ only, see methods for preparation and data-processing. Figure \[fig2\]a shows the characteristic asymmetric profile of the phase whereas the optical-density is well-described by a Lorentzian. The solid lines result from fitting the phase data to equation (\[eq1\]) with $n_1=0$ and agree with the measurement provided the probe duration is $\sim 1\,{\rm \mu s}$ and the saturation is less than $\sim 10$. The use of stronger (longer) pulses leads to a systematic shift of measured phases to larger values, due to the light-forces resulting from the strong focusing of the probe beam and scattering of photons. To verify (ii), we measure the phase-variance as a function of the photon number as shown in Figure \[fig2\]b. From the power-law behavior of the phase-variance we deduce that photon shot-noise limits the sensitivity of the phase measurement. Indeed, the phase-noise is expected to be given by $\delta \phi^2=\frac{1}{\eta N}$ [@lye_nondestructive_2003], where $N$ is the number of photons and $\eta=0.6$ the quantum efficiency, which is determined in an independent measurement.
{width="\textwidth"}
We first measure the distribution of the spin-polarization for weakly interacting Fermi gases at three different temperatures given by $T_1=8.5(5)\,{\rm \mu K}$, $T_2=2.0(2)\,{\rm \mu K}$ and $T_3=0.58(5)\,{\rm \mu K}$, see methods for preparation. Figure \[histo\]a shows that the probability distributions of the spin-polarization have a narrower width the lower the temperature of the gas. The distributions exhibit no significant asymmetry and are well described by Gaussian functions, as expected from the large number of atoms in the probe beam, $\sim\,500$ in each state. For weakly interacting Fermi gases, number fluctuations in each hyperfine state are independent, so that fluctuations of the spin-polarization are given by $\delta m^2=\delta n_1^2+\delta n_2^2$. Consequently, with the onset of quantum-degeneracy, spin-fluctuations are reduced, since in each state only atoms at the Fermi edge contribute to the fluctuations, which is a manifestation of antibunching due to Pauli’s principle [@muller_local_2010; @sanner_suppression_2010]. The measured variances of the spin-polarization, in order of decreasing temperature, are $\delta m_{T_1}^2=35(6)\,{\rm \mu m^{-4}}$, $\delta m_{T_2}^2=27(3)\,{\rm \mu m^{-4}}$ and $\delta m_{T_3}^2=15(3)\,{\rm \mu m^{-4}}$. Here, the background corresponding to a standard deviation of $\pm 13(2)$ atoms in the probed volume has been subtracted, see methods. The column density in the probed region was $n_1=n_2=n_{{\rm col}}/2=110\,{\rm \mu m^{-2}}$ for the gases at $T_2$ and $T_3$ and $20\,\%$ lower for the gas at $T_1$. From our measurement we determine a reduction of the spin-fluctuations as compared to a thermal gas prepared with the same column density by [dB]{} for the gas at $T_2$ and [dB]{} for the gas at $T_3$, as shown in Figure \[histo\]b, which is in quantitative agreement with theory for a noninteracting Fermi gas, see methods.
We now turn to the study of a gas with strong repulsive interactions, prepared close to a Feshbach resonance at temperature $T_{{\rm mol}}=0.36(10)\,{\rm \mu K}$, see methods. The resulting histogram is displayed in Figure \[histo\]a and shows a distribution of the spin-polarization that is significantly narrower than for the weakly interacting gases. This reflects the fact that for interacting gases correlations are present between the different spin states. In particular, for strong repulsive interactions close to a Feshbach resonance, weakly bound molecules form [@varenna_lectures] and the number of atoms in the two states will always be equal. Spin-fluctuations in the atom number difference are created at the cost of breaking molecules and are consequently suppressed. We measure $\delta m_{{\rm mol}}^2=5(2)\,{\rm \mu m^{-4}}$ at $n_{\rm col}/2=110\,{\rm \mu m^{-2}}$, with the background subtracted as before. This corresponds to a reduction by [dB]{} as compared to a noninteracting thermal gas. A reduction by $\sim 18 \rm{dB}$ is expected from ref. [@bruun_probing_2009] for a molecular BEC at zero temperature. The observation of a lower value could be caused by pair-breaking or fluctuations of the probe frequency, see methods.
The measured values for the spin-fluctuations can be used to determine the magnetic susceptibility $\chi$ via the fluctuation-dissipation theorem (FDT) [@seo_compressibility_2011], provided the probed system is in grand-canonical equilibrium with its surroundings. Due to column-integration, the fluctuation-dissipation theorem here reads $kT \int\chi=kT \chi_{{\rm col}}=A \delta m^2$, where $\chi_{{\rm col}}$ is the column-integrated magnetic susceptibility, $k$ Boltzmann’s constant and $A$ the effective area of the probed column. For small volumes and at low temperatures corrections to the FDT are expected, because correlations between the probed system and its surroundings cannot be neglected [@klawunn_local_2011; @hung_observation_2011]. Using ref. [@klawunn_local_2011] and including column-integration, we estimate the corrections to be less than 10% even for our coldest samples. Figure \[histo\]c shows the column-integrated magnetic susceptibility per particle $\chi_{{\rm col}}/n_{{\rm col}}$ as a function of temperature. It relates the spin imbalance to the energy needed to create it. For high temperatures, the susceptibility decreases inversely proportional to the temperature as expected. For low temperatures and with the onset of quantum degeneracy, the susceptibility saturates to a value depending on the trap details: the stiffer the trap, the lower the susceptibility. The solid line shows theory for noninteracting fermions calculated for the trap parameters of the gas at $T_3$. The susceptibility for the strongly interacting gas of molecules is lower than for the weakly interacting gas at $T_3$, despite its lower temperature and weaker trap.
The link between the magnetic susceptibility and spin-fluctuations has been proposed as a macroscopic entanglement witness in solid state systems [@wiesniak_magnetic_2005]. We now apply this concept to our measurement of the spin-fluctuations. For this purpose we describe each particle as a two-level system, which together form an effective spin $\mathbf{M}$ [@wiesniak_magnetic_2005; @toth_spin_2009], where $\langle M_z\rangle$ is proportional to the atom number difference and $\delta M_z^2=A^2 \delta m^2$. For all separable states, the ratio of the spin-fluctuations to the total atom number is bounded from below. For a system in grand-canonical equilibrium described by a Hamiltonian which is invariant under rotations of the spin state, this bound is expressed by the inequality $$\frac{A \delta m^2}{n_{{\rm col}}}\geq\frac{2}{3}, \label{eq:inequality}$$ see methods and supplementary informations for details. Figure \[histo\]d shows the spin-fluctuations $A \delta m^2$ as a function of the column density $n_{\rm{col}}$. The gas at $T_3$ violates the above inequality, which is expected because the Pauli principle leads to non-separable states at low temperatures [@Vedral2003]. The notion of entanglement applied to indistinguishable particles is subject of current investigations [@Horodecki2009].
For the case of the strongly interacting gas, we find $\frac{A \delta m^2}{n_{{\rm col}}}=0.11(4)$, violating the bound by more than 10 standard deviations. The fluctuations are $2.4\,\rm{dB}$ lower than expected for a noninteracting gas at the same temperature, see bold line in Figure \[histo\]d. This corresponds to $2.4\,\rm{dB}$ of spin-squeezing following ref. [@kheruntsyan_quantum_2006].
Our analysis uses the following assumptions: (i) We create a gas with equal atom number in the two states, which leads to $\langle M_z\rangle=0$. (ii) We probe the system in thermal equilibrium. As a consequence, the transverse components have dephased to the amount permitted by Pauli’s principle and we have $\langle M_x \rangle=\langle M_y \rangle=0$. (iii) The time-evolution is described by a Hamiltonian that is invariant under rotations of the spin [@Zwierlein2003]. It is therefore sufficient to measure only the $z$-component.
The rigorous application of the entanglement witness [@wiesniak_magnetic_2005; @toth_spin_2009] requires equal coupling of all atoms to the probe beam. Here, we take into account the inhomogeneity of the probe beam by introducing the effective area $A$.
In conclusion, we have measured the probability distribution of the spin-flucutations in a trapped Fermi gas. We have discussed our measurement in terms of an entanglement witness. Our work constitutes a first step towards the detection of entanglement in many-body states in quantum gases. The detection of higher order correlations could be achieved by extracting higher moments of the probability distribution for the spin-polarization [@cherng_quantum_2007].
[**Methods:**]{}
#### Generation of interferometer beams
The interferometer beams are generated by applying two radio frequencies differing by 20 MHz along each of the axes of a two-axis acousto-optical deflector (AOD) very similar to previous work [@zimmermann_high-resolution_2011]. This results in four beams in the diffraction order of the AOD which, after passing through a high-resolution microscope objective, are arranged in a square. Two of these beams, probe beam and local oscillator, have exactly the same frequency and form the interference pattern on the camera. The other two are detuned by $\pm20$ MHz and their interference patterns average out over the duration of the probe pulses. The intensities of the beams are controlled via the power in the individual radio frequencies so that the local oscillator is 20 times as intense as the probe beam. Phase stability is ensured by deriving each radio frequency from the same source for both axes. The light beams are elliptically polarized. Due to the birefringence of the atomic cloud in a magnetic field, also used in polarization-contrast imaging [@bradley_bose-einstein_1997], only the $\sigma^-$-polarized component of the light interacts with the atoms, while the $\sigma^+$- component passes undisturbed, see Figure \[fig1\]b. The power ratio of $\sigma^+$- and $\sigma^-$-component is about 10.
#### Experimental sequence
An equal mixture of $^6$Li atoms in the two lowest hyperfine states, denoted by $|1\rangle$=$|m_J$=$-1/2,m_I$=$1\rangle$ and $|2\rangle$=$|m_J$=$-1/2,m_I$=$0\rangle$ is prepared similar to previous work [@muller_local_2010]. A second dipole trap with a wavelength of $767\, {\rm nm}$ and a $1/e^2$-radius of $10\,{\rm \mu m}$ is then switched on over 500ms and is used to locally increase the total column-density. Finally the magnetic field is ramped to 475G where the scattering length is $a=-100\,a_0$, with $a_0$ the Bohr radius. The final trap depths of the large dipole trap are $84\,{\rm \mu K}$ for $T_1$, $19\,{\rm \mu K}$ for $T_2$ and $10\,{\rm \mu K}$ for $T_3$ corresponding to $1050\, {\rm mW}$, $235\, {\rm mW}$ and $130\, {\rm mW}$, respectively. The trap depths of the second dipole trap are $9\,{\rm \mu K}$, $4.5\,{\rm \mu K}$ and $1.5\,{\rm \mu K}$, respectively. The central region of the cloud is probed interferometrically with a pulse of 1.2$\mu$s duration at a maximum saturation of 9. Experiments are repeated 400 times. Images of the whole cloud are taken after the interferometric measurement to determine the total atom number as well as the temperature from the radial expansion after $1\,{\rm ms}$ time of flight. Experiments that show a larger deviation of the total atom number than 5% are discarded, amounting to $10-20\%$ of the images. For the preparation of the strongly interacting gas, we ramp directly to $800\, {\rm G}$, where $a=7000\,a_0$ and the trap depth is lowered to $17\, {\rm mW}$ followed by recompression to $30\, {\rm mW}$, which corresponds to a trap depth of $4.8\, {\rm \mu K}$ for the molecules. The experiment is repeated 200 times. We estimate the final temperature to be one tenth of the relevant trap depth leading to $T_{{\rm mol}}=0.36(10)\,{\rm \mu K}$ after recompression. For the preparation of a gas containing only atoms in state $|2\rangle$, we hold the gas for 100ms close to a $p$-wave Feshbach resonance at $159\,{\rm G}$, which leads to the loss of nearly all particles in state $|1\rangle$. Subsequent evaporation of the remaining atoms leads to a non-degenerate gas of atoms in state $|2\rangle$.
#### Theory
We compare our measurements for the weakly interacting gases with theory for noninteracting fermions [@Huang_Statistical]. We determine the temperature and the chemical potential from fits to the time-of-flight images as in previous work [@muller_local_2010]. We then calculate the density distribution in the combined trap given the fitted temperature and atom number to determine $T_F$ in the center of the combined trap. This assumes full thermalization in the presence of the second trap, which is supported by the measured spin-fluctuations. The knowledge of $T$, $\mu$ and the trap shape allows us calculate the mean and the variance of the atomic density along the line of sight using column-integration.
#### Data processing
Three images are obtained in each experiment, one with atoms in the interferometer and two without. The images are averaged along the direction parallel to the fringe-pattern and a Fourier filter is applied to suppress the low spatial frequencies. For the determination of the mean phase, a simultaneous sinusoidal fit to the $\sigma^-$-pattern on images with and without atoms is used to determine the phase. Free parameters are amplitude and wavelength of the fringe pattern, as well as a common phase for both patterns and a phase-difference for the picture with atoms. Residual mechanical motion, e.g. due to the microscopes, is compensated for by using the same simultaneous fit applied to the $\sigma^+$-pattern. For determination of the phase-fluctuations we find it advantageous to analyze the correlations between the $\sigma^+$- and the $\sigma^-$-pattern on each image. This allows us to exploit the similarity of the two patterns and use the $\sigma^+$-pattern to noiselessly amplify the signal contained in the $\sigma^-$-pattern, analogous to homodyne techniques. The phase-shift due to the atoms causes a displacement of the zero-crossings of the correlation function in the image with atoms as compared to the image without atoms. See supplemental materials for further details on data processing.
#### Determination of effective area
The effective area $A$ relates the spin-fluctuations to the mean atomic density and corresponds to the area of a beam with uniform intensity giving the same result. For the nearly non-degenerate gas at $T_1=1.02 T_F$ the spin-fluctuations are well described by Poissonian statistics because the atoms are uncorrelated. The fluctuations are thus proportional to the average atom number in the probe volume. This allows the determination of the effective area $A=0.97 n_{{\rm col}}/\delta m_{T_1}^2=4.9(8)\,{\rm \mu m^2}$, corresponding to an effective waist of the probe beam of $1.8(3)\,{\mu m}$. The factor $0.97$ accounts for the residual suppression of the spin-fluctuations at $T_1=1.02 T_F$.
#### Background
The contribution to the phase variance originating from photon shot-noise is determined from the two images without atoms using the above described procedure. This yields $\delta m_{{\rm bgr}}=6.7(1.0)\,{{\rm \mu m^{-4}}}$ and a standard deviation of the atom number difference in the probed volume of $\sqrt{A^2 \delta m_{\rm bgr}^2}=13(2)$. Frequency fluctuations of the probe can cause fluctuations in the measured phase. A variance of the probe frequency of 2MHz$^2$ would correspond to apparent spin-fluctuations of $5\,{\rm \mu m^{-4}}$ at the column-density in our experiment and could contribute to the difference of our results compared to the theory in ref. [@bruun_probing_2009].
#### Entanglement Witness
Each atom realizes a two-level system represented by Pauli matrices $\sigma_{x,y,z}$. We define the spin-polarization or magnetization of the probed region as $M_{x,y,z}=\sum_i \sigma_{x,y,z}^i$, where the sum extends over all atoms in the probe volume. For all separable states the inequality $\delta M_x^2+\delta M_y^2+\delta M_z^2\geq 2 \langle N\rangle$ is fulfilled, where $\langle N\rangle$ is the average number of atoms probed [@wiesniak_magnetic_2005]. Using the invariance of the Hamiltonian under spin rotations and that we measure expectation values in the grand-canonical ensemble, which show the same symmetry as the Hamiltonian, this reduces to $\frac{\delta M_z^2}{N}=\frac{A \delta m^2}{n_{\rm col}} \geq\frac{2}{3}$. See Supplementary Information for more details.
[10]{}
Auerbach, A. . Springer, (1998).
Inguscio, M., Ketterle, W., and Salomon, C., editors. , volume Course CLXIV, (2008).
Jördens, R., Tarruell, L., Greif, D., Uehlinger, T., Strohmaier, N., Moritz, H., Esslinger, T., Leo, L. D., Kollath, C., Georges, A., Scarola, V., Pollet, L., Burovski, E., Kozik, E., and Troyer, M. Quantitative determination of temperature in the approach to magnetic order of ultracold fermions in an optical lattice. (18), 180401, May (2010).
Nascimbène, S., Navon, N., Pilati, S., Chevy, F., Giorgini, S., Georges, A., and Salomon, C. behavior of the normal phase of a strongly interacting gas of cold atoms. (21), 215303, May (2011).
Sommer, A., Ku, M., Roati, G., and Zwierlein, M. W. Universal spin transport in a strongly interacting fermi gas. (7342), 201–204, April (2011).
Lewenstein, M., Sanpera, A., Ahufinger, V., Damski, B., Sen, A., and Sen, U. Ultracold atoms in optical lattices: Mimicking condensed matter physics and beyond. , 243–379 (2007).
Eckert, K., [Romero-Isart]{}, O., Rodriguez, M., Lewenstein, M., Polzik, E. S., and Sanpera, A. Quantum non-demolition detection of strongly correlated systems. (1), 50–54, January (2008).
Roscilde, T., Rodriguez, M., Eckert, K., [Romero-Isart]{}, O., Lewenstein, M., Polzik, E., and Sanpera, A. Quantum polarization spectroscopy of correlations in attractive fermionic gases. (5), 055041 (2009).
Bruun, G. M., Andersen, B. M., Demler, E., and [Sørensen]{}, A. S. Probing spatial spin correlations of ultracold gases by quantum noise spectroscopy. (3), 030401, January (2009).
Hammerer, K., [Sørensen]{}, A. S., and Polzik, E. S. Quantum interface between light and atomic ensembles. (2), 1041, April (2010).
Estève, J., Trebbia, J., Schumm, T., Aspect, A., Westbrook, C. I., and Bouchoule, I. Observations of density fluctuations in an elongated bose gas: Ideal gas and quasicondensate regimes. (13), 130403, April (2006).
Gemelke, N., Zhang, X., Hung, C., and Chin, C. In situ observation of incompressible mott-insulating domains in ultracold atomic gases. (7258), 995–998 (2009).
Müller, T., Zimmermann, B., Meineke, J., Brantut, J., Esslinger, T., and Moritz, H. Local observation of antibunching in a trapped fermi gas. (4), 040401, July (2010).
Sanner, C., Su, E. J., Keshet, A., Gommers, R., Shin, Y., Huang, W., and Ketterle, W. Suppression of density fluctuations in a quantum degenerate fermi gas. (4), 040402, July (2010).
Bakr, W. S., Peng, A., Tai, M. E., Ma, R., Simon, J., Gillen, J. I., Fölling, S., Pollet, L., and Greiner, M. Probing the [SuperfluidtoMott]{} insulator transition at the [Single-Atom]{} level. (5991), 547 –550, July (2010).
Sherson, J. F., Weitenberg, C., Endres, M., Cheneau, M., Bloch, I., and Kuhr, S. Single-atom-resolved fluorescence imaging of an atomic mott insulator. (7311), 68–72 (2010).
Serwane, F., [Zürn]{}, G., Lompe, T., Ottenstein, T. B., Wenz, A. N., and Jochim, S. Deterministic preparation of a tunable [Few-Fermion]{} system. (6027), 336 –338, April (2011).
Oblak, D., Petrov, P. G., Alzar, C. L. G., Tittel, W., Vershovski, A. K., Mikkelsen, J. K., [Sørensen]{}, J. L., and Polzik, E. S. Quantum-noise-limited interferometric measurement of atomic noise: Towards spin squeezing on the cs clock transition. (4), 043807, April (2005).
Appel, J., Windpassinger, P. J., Oblak, D., Hoff, U. B., [Kjærgaard]{}, N., and Polzik, E. S. Mesoscopic atomic entanglement for precision measurements beyond the standard quantum limit. (27), 10960 –10965, July (2009).
Sanner, C., Su, E. J., Keshet, A., Huang, W., Gillen, J., Gommers, R., and Ketterle, W. Speckle imaging of spin fluctuations in a strongly interacting fermi gas. (1), 010402, January (2011).
Aljunid, S. A., Tey, M. K., Chng, B., Liew, T., Maslennikov, G., Scarani, V., and Kurtsiefer, C. Phase shift of a weak coherent beam induced by a single atom. (15), 153601, October (2009).
Lye, J. E., Hope, J. J., and Close, J. D. Nondestructive dynamic detectors for [Bose-Einstein]{} condensates. (4), 043609, April (2003).
Seo, K. and [Sá de Melo]{}, C. A. R. Compressibility and spin susceptibility in the evolution from [BCS]{} to [BEC]{} superfluids. , May (2011).
Klawunn, M., Recati, A., Pitaevskii, L. P., and Stringari, S. Local atom-number fluctuations in quantum gases at finite temperature. (3), 033612 (2011).
Hung, C., Zhang, X., Gemelke, N., and Chin, C. Observation of scale invariance and universality in two-dimensional bose gases. (7333), 236–239, February (2011).
Wieśniak, M., Vedral, V., and Časlav Brukner. Magnetic susceptibility as a macroscopic entanglement witness. , 258–258 (2005).
Tóth, G., Knapp, C., Gühne, O., and Briegel, H. J. Spin squeezing and entanglement. (4), 042334, April (2009).
Vedral, V. Entanglement in the second quantization formalism. (2), 289–306, June (2003).
Horodecki, R., Horodecki, P., Horodecki, M., and Horodecki, K. Quantum entanglement. , 865–942, Jun (2009).
Kheruntsyan, K. V. Quantum atom optics with fermions from molecular dissociation. (11), 110401, Mar (2006).
Zwierlein, M. W., Hadzibabic, Z., Gupta, S., and Ketterle, W. Spectroscopic insensitivity to cold collisions in a two-state mixture of fermions. , 250404, Dec (2003).
Cherng, R. W. and Demler, E. Quantum noise analysis of spin systems realized with cold atoms. (1), 7 (2007).
Zimmermann, B., Müller, T., Meineke, J., Esslinger, T., and Moritz, H. High-resolution imaging of ultracold fermions in microscopically tailored optical potentials. , 043007, April (2011).
Bradley, C. C., Sackett, C. A., and Hulet, R. G. condensation of lithium: Observation of limited condensate number. (6), 985, February (1997).
Huang, K. . Wiley, New York, 2nd ed. edition, (1987).
Giorgini, S., Pitaevskii, L. P., and Stringari, S. Theory of ultracold atomic fermi gases. (4), 1215–60 (2008).
#### Acknowledgments
We acknowledge enlightening discussions with M. Christiandl, A. Imamoglu, K. Mølmer, E. Polzik, R. Renner, A. Sørensen, M. Ueda and V. Vuletic and funding from NCCR MaNep, NCCR QSIT, ERC SQMS and FP7 FET-open NameQuam. JPB acknowledges support of European Union under Marie Curie IEF fellowship.
#### Author Contributions
JM and JPB analyzed the data, JM, JPB, DS and TM carried out the experimental work, all authors contributed to project planning and to writing the manuscript.
####
Correspondence should be addressed to TE.
####
The authors declare no competing financial interests.
Beam generation
===============
The interferometer beams are generated using a two-axis acousto-optic deflector (2D AOD). Radio-frequency (RF) signals at two different frequencies are sent to both axes of the 2D AOD, resulting in the deflection of four beams in the -1/-1 order. Two RF sources (USRP 2) generate monochromatic signals, one at 47.5 MHz and the other at 67.5 MHz. The splitting and subsequent recombination depicted in Fig. \[figRF\] allows to tune the power ratio between the two frequencies, for each axis of the 2D AOD.
![RF circuit for the generation of the interferometer beams. Two RF sources (USRP 2) generate signals at 47.5 and 67.5 MHz, respectively. After splitting and selective attenuation, the two frequencies are combined and injected on each axis of the 2D AOM. Four beams are generated having a variable power ratio. The four spots in the image represent the four beams after passing through the microscope. The sizes of the spots represent the relative powers, and the colors represent different frequencies of the deflected light.[]{data-label="figRF"}](RF.pdf){width="45.00000%"}
After passing through a high-resolution microscope objective, the beams deflected in the -1/-1 order are arranged in a square and have a variable power, as depicted on Figure \[figRF\]. The most (least) powerful of those beams has been deflected by the non-attenuated (attenuated) signals from each source. These two beams have the same frequency, differing by 115 MHz from that of the incoming laser beam, and they lead to the interference pattern observed in the experiment. Conversely, the two beams having the same power have frequencies differing by 40 MHz, and contribute only to the background.
Data processing using correlations
==================================
{width="\textwidth"}
In order to take full advantage of the similarity between the probe signal ($\sigma^-$ polarized) and the reference signal ($\sigma^+$ polarized), we process the data using a method inspired by heterodyne detection. Figure \[fig1\] illustrates the different steps of this data processing algorithm.
The two parts of the experimental images contain the probe and reference signals (Figure \[suppfig1\]a). A line sum along the direction of the fringes is first computed, yielding a one dimensional fringe pattern signal (Figure \[suppfig1\]b). The right part of the signal is scaled so that the probe and reference have roughly the same intensity. A filter in Fourier space is then applied to the full scaled signal, conserving only the Fourier components around the fringe spacing. The precise shape of the filter does not influence the obtained results, provided the low frequency components (containing the envelope of the two fringe patterns and the background) are removed. Figure \[suppfig1\]c presents a typical signal obtained after those processing steps.
We then compute the autocorrelation function of the processed fringe pattern. A fixed spacing is inserted between the probe and reference signal in the processed fringe pattern before the correlation function is computed. We do this so that when computing the autocorrelation function of the picture, the $\sigma^+$ - $\sigma^-$ correlation is isolated from the other contributions ($\sigma^+$ - $\sigma^+$ and $\sigma^-$ - $\sigma^-$ ) Figure \[suppfig1\]d presents a typical autocorrelation signal. The autocorrelation function displays two separated parts. The sum of the correlation functions of the reference with itself and the probe with itself appears at the center. Conversely, the correlation function of the probe with the reference appears at the sides. The center part of this probe-reference correlation signal is selected (in order to avoid finite size effects), and the positions of the zero crossings are extracted by linear interpolation of the discrete signal, as depicted on Figure \[suppfig1\]d. The error made by linear interpolation is typically two orders of magnitude lower than the photon shot-noise limit in our experiments. The position of a zero-crossing of the correlation function corresponds to the displacement of the $\sigma^+$-pattern with respect to the $\sigma^-$-pattern that is required to overlap the maxima of one pattern with the zero-crossings of the second pattern. The determination of the zero-crossings of the correlation function is thus a direct measurement of the [*relative displacement*]{} of the $\sigma^+$-pattern with respect to the $\sigma^-$-pattern.
Each run of the experiment yields three pictures : one taken in the presence of the atoms, and two pictures taken in the absence of atoms (hereafter named pictures 1,2 and 3). The position of the crossings and thus the fringe displacements are measured on each of those pictures. The mean over the crossings yields the mean position of the fringe pattern in each picture.
To obtain the distribution of the fringe displacements, the experiments are repeated up to 400 times, over about 2 hours. Over this period, slow drifts of the phase on the “atoms” pictures, of up to 5 degrees, are observed, most probably due to temperature variations in the environment. In order to compensate those drifts, a sliding average (over 15 runs) of the positions of the crossings on picture 3 is computed. We take this as a measurement of the drift, and subtract this averaged signal from the positions measured on pictures 1 and 2. This subtraction operation amounts to measuring the displacement of each crossing of picture 1 to the corresponding crossing of picture 3, corrected for long term deviations. Taking the average of this corrected quantity over the crossings (i.e. averaging the position measured for all the pattern), we obtain the relative displacement of the fringes on pictures 1 and 2 compared to 3. The ratio of this mean displacement to the period of the fringes yields the phase shift observed on pictures 1 and 2 compared to 3.
We now have two phase shifts measured for each run of the experiment, with and without atoms. Thus, we obtain the statistical distribution of the phase shifts in the presence of the atoms from all the shifts of picture 1, together with the distribution of shifts on pictures 2, taken in exactly the same conditions but without atoms. To measure the phase variance due to the atoms, we subtract the phase variance on picture 2 (background variance in the text) to the phase variance of the picture 1.
Entanglement Witness
====================
In reference [@wiesniak_magnetic_2005] it is shown that the magnetic susceptibility is a macroscopic entanglement witness. If the magnetic susceptibility is lower than a certain bound, the state describing the thermal state of the bulk system cannot be a convex combination of product states (separable state), which means that the state must contain entanglement. The arguments in ref. [@wiesniak_magnetic_2005] are based on the fluctuation-dissipation theorem which links the magnetic susceptibility to the fluctuations of the spins that are the constituents of the system. The experimental signal is proportional to the difference in atomic density in each of the two spin states. Let us define the collective spin describing the magnetization as $$\vec{M}=\sum \vec{\sigma}_i,$$ where $\vec{\sigma}_i$ are Pauli matrices. The calculation described in ref. [@wiesniak_magnetic_2005] leads to $$\delta M_x^2+\delta M_y^2+\delta M_z^2\geq 2 N. \label{eq:ineq}$$ Assuming that the fluctuations are the same along all three axes, as we will argue below, the fluctuations in a given direction are bounded by $$\label{eq:rotsymineq}
\delta M_z^2\geq \frac{2}{3} N.$$ In the experiment we have access to the column-integrated total density $n_{\rm{col}}=N/A$ and the column-integrated fluctuation density $\delta m^2=\delta M_z^2/A^2$, where $A$ is the effective area of the probed region. The effective area $A$ is determined experimentally and takes into account the intensity profile of the probe beam. Using this in we get $$\frac{A\delta m^2}{n_{\rm{col}}}\geq \frac{2}{3}.$$ The measurement of a value lower than $2/3$ means that the state cannot be separable.
Invariance under Spin Rotations {#invariance-under-spin-rotations .unnumbered}
-------------------------------
The Hamiltonian describing a two-component Fermi gas is given by [@giorgini_theory_2008]: $$\begin{aligned}
H&=&\sum_{\sigma}\int d\vec{r}\Psi(\vec{r})_{\sigma}^{\dagger}\left(-\frac{\hbar^2 \nabla^2}{2m}+V_{ext}(\vec{r})-\mu_{\sigma}\right)\Psi(\vec{r})_{\sigma}\nonumber \\
&+ &\iint d\vec{r}d\vec{r'}V(\vec{r}-\vec{r'})\Psi(\vec{r})_{\uparrow}^{\dagger}\Psi(\vec{r'})_{\downarrow}^{\dagger}\Psi(\vec{r'})_{\downarrow}\Psi(\vec{r})_{\uparrow}.\end{aligned}$$ Here the field operators for the two spin states denoted by $\sigma\in\{\uparrow, \downarrow\}$ obey the fermionic anti-commutation relation $\{\Psi(\vec{r})_{\sigma},\Psi(\vec{r'})_{\sigma'}^{\dagger}\}=\delta_{\sigma,\sigma'}\delta(\vec{r}-\vec{r'})$, $V$ is the interaction potential, $V_{ext}$ the trapping potential and $\mu_{\sigma}$ is the chemical potential for the two components. At the energy scale of our experiment, the interaction term in the Hamiltonian describes the s-wave scattering of two particles, the strength of which is characterized by the scattering length $a$. The wavefunction describing the relative position of the two particles is symmetric under the exchange of particles. As a result the spin wavefunction must be antisymmetric (singlet) and is thus invariant under rotations of the spin. For a balanced gas, $\mu_{\uparrow}=\mu_{\downarrow}$, the first part of the Hamiltonian is invariant under rotations of the spin state. The external trapping potential for the two components differs by the energy of the hyperfine splitting. However, since no transitions between the states occur on experimentally relevant timescales and since we prepare a balanced gas, this energy difference is irrelevant for the dynamics of the system. Since we prepare the system in a thermal state, the transverse components of the collective spin $\mathbf{M}$ have completely decohered. The invariance under rotations of the spin relies both on the structure of the Hamiltonian and on the preparation of a balanced, thermal state [@Zwierlein2003]
Entanglement Witness in Grand-Canonical Ensemble {#entanglement-witness-in-grand-canonical-ensemble .unnumbered}
------------------------------------------------
#### Symmetry for grand-canonical expectation values
In our experiment we measure the spin-fluctuations of a small subsystem of the whole atomic cloud. This subsystem is in thermodynamic equilibrium with the rest of the system. In this situation the expectation values in the grand-canonical ensemble of a given observable have the same symmetry as the Hamiltonian, which might not be the case for the ground state. For the specific case considered we have $\langle M_z\rangle=\langle U M_z U^{\dagger}\rangle$, where $U$ is an arbitrary rotation of the spin state. The state of the system is given by the density matrix $\rho=\frac{1}{Z}e^{-\beta H}$. We then have $$\begin{aligned}
\langle U M_z U^{\dagger}\rangle&=& Tr \rho U M_z U^{\dagger}\nonumber\\
&=& \frac{1}{Z}Tre^{-\beta H}U M_z U^{\dagger} \nonumber\\
&=& \frac{1}{Z}Tr Ue^{-\beta H} M_z U^{\dagger}\quad([H,U]=0) \nonumber\\
&=& \frac{1}{Z}Tr e^{-\beta H} M_z U^{\dagger}U\quad(\mathrm{trace\ cyclic}) \nonumber\\
&=& \frac{1}{Z}Tr e^{-\beta H} M_z \nonumber\\
&=& \langle M_z \rangle.\end{aligned}$$ As an example, consider the situation of a ferromagnetic ground state where domains of spins pointing in a certain direction would exist. In each realization the experiment the symmetry is broken, but the direction in which the spins are pointing is different. Every direction must be equally probable, which leads to the situation that the ensemble averages have the same symmetry than the Hamiltonian.
#### Non-fixed particle number
The derivation of the inequalitiy \[eq:ineq\] assumes a fixed number of particles $N$ and the authors of ref. [@wiesniak_magnetic_2005] work in the canonical ensemble. However, the argument carries through for measurements in the grand-canonical ensemble where $N$ is replaced by its expectation value $\langle N\rangle$. For this we express the density matrix in the grand-canonical ensemble as $$\rho=\sum_N w_N \rho_N,$$ where $\rho_N$ is the density matrix in the subspace of fixed particle number $N$ with the weight factors $w_N$. As before, we write $\langle \cdot \rangle=Tr \rho \cdot$ for the grand-canonical expectation value and similarly we define $\langle \cdot \rangle_N=Tr \rho_N \cdot$ as the expectation value in the $N$-particle subspace. Similarly, $\delta\cdot$ and $\delta_N \cdot$ denote the variance in the grand-canonical ensemble and with fixed particle number $N$, respectively. We now calculate the variance of the operator $M_z$ $$\begin{aligned}
\delta M_z^2&=&\langle (M_z -\langle M_z \rangle)^2\rangle\nonumber\\
&=&\sum_N w_N Tr \rho_N (M_z -\langle M_z \rangle)^2\nonumber\\
&=&\sum_N w_N Tr \rho_N (M_z -\langle M_z \rangle_N -(\langle M_z \rangle-\langle M_z \rangle_N))^2\nonumber\\
&=&\sum_N w_N Tr \rho_N \Big{[}(M_z -\langle M_z \rangle_N)^2 \nonumber\\
&&-2( M_z -\langle M_z \rangle_N)(\langle M_z \rangle-\langle M_z \rangle_N)\nonumber\\
&&+(\langle M_z \rangle-\langle M_z \rangle_N)^2\Big{]}\nonumber\\
&=&\sum_N w_N \delta_N M_z^2 + \sum_N w_N(\langle M_z \rangle -\langle M_z \rangle_N)^2\nonumber\\
&\geq&\frac{2}{3}\langle N \rangle +\sum_N w_N(\langle M_z \rangle -\langle M_z \rangle_N)^2.\end{aligned}$$ In the last step we have used the known result from \[eq:rotsymineq\]. The last term in the last line is small, because the width of the weights $w_N$ is centered around $N=\langle N \rangle$ with a width $\sqrt{\langle N\rangle}$ and because $\langle M_z\rangle \simeq \langle M_z\rangle_{\langle N\rangle}$. More importantly, the last term is positive and if the fluctuations per particle are measured to be smaller than $2/3$ the state cannot be separable.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'We present a novel distribution-free approach, the data-driven threshold machine (DTM), for a fundamental problem at the core of many learning tasks: choose a threshold for a given pre-specified level that bounds the tail probability of the maximum of a (possibly dependent but stationary) random sequence. We do not assume data distribution, but rather relying on the asymptotic distribution of extremal values, and reduce the problem to estimate three parameters of the extreme value distributions and the extremal index. We specially take care of data dependence via estimating extremal index since in many settings, such as scan statistics, change-point detection, and extreme bandits, where dependence in the sequence of statistics can be significant. Key features of our DTM also include robustness and the computational efficiency, and it only requires one sample path to form a reliable estimate of the threshold, in contrast to the Monte Carlo sampling approach which requires drawing a large number of sample paths. We demonstrate the good performance of DTM via numerical examples in various dependent settings.'
bibliography:
- 'bibfile.bib'
---
Introduction
============
Selecting threshold is a key step in many machine learning tasks, such as anomaly detection by scan statistics [@Scan09], sequential change-point detection [@annals2013], and extreme $K$-arm bandit [@extremeBandits14]. Broadly speaking, determining threshold is the central problem for statistical hypothesis testing and estimating confidence intervals. The goal of setting the threshold include controlling the $p$-value or the significance level, controlling the false-alarm rate, or establishing the upper or lower confidence bounds for the max $K$-arm bandits. This goal can usually be cast into setting a threshold $x$ such that the maximum of a random sequence $S_1, S_2, \ldots, S_n$, which typically corresponds to statistics or outputs of a learning algorithm, is less than the threshold with a pre-specified level $\alpha$, i.e., $$\mathbb{P}\cbr{\max_{t=1,\ldots n} S_t > x} \leqslant \alpha, \label{key_problem}$$ under the assumed distribution in the hypothesis setting etc. These $S_t$ are dependent in many settings. For instance, in scan statistics, there are $n$ scanning regions, $S_t$ corresponds to a statistic formed for each region. An anomaly is detected if any of the regional statistics exceeds the threshold, and $\alpha$ is the significance level.
Despite its importance, setting threshold remains one of the most challenging parts in designing a learning algorithm. This is commonly done by Monte Carlo simulations and bootstrapping, which requires repeating a large number of experiments to generate sequences either from the assumed distribution or by bootstrapping the original sequence; this can be computationally extensive. Since $\alpha$ is usually set to a small number (for the algorithm to be reliable), this means that we have to estimate a small probability. To obtain a high precision, a large number of repetitions are needed. What exacerbates this is that in many settings generating samples are not easy. For instance, the assumed distribution can be difficult to draw samples, and it is common to use the computationally extensive Markov-Chain Monte Carlo techniques. In the learning setting, this can mean to run the algorithms many times, and running the algorithm even once (such as deep-learning) over a large-scale reference dataset even once can be time-consuming.
In other cases, analytical approximations are found to relate the tail probability to the threshold (e.g., ARL approximation in the sequential change-point detection setting [@annals2013]). However, these results typically make strong parametric assumptions on the data to make the problem tractable. In practice it is hard to model the distribution for the sequence precisely, being the output of a learning algorithm, and the distribution may vary from one dataset to the next. Moreover, the random sequence has non-negligible dependence, while theoretical approximations are usually developed for samples. For instance, in scan statistics [@Scan09] for anomaly detection, a detection statistic is formed for each local scanning region, and the statistics for overlapping scanning regions are correlated since they are computed over common data. In sequential hypothesis testing and change-point detection, given a streaming data sequence, each time we form a statistic over a past sliding window to detect a change. Due to overlapping of the sliding window, the detection statistics at each time are highly correlated. In the bandit setting, the empirical rewards may be estimated from a sliding window which again results in dependence. Without taking into account the dependence, threshold choice is usually inaccurate and cannot meet the targeted level.
Therefore, there is a substantial need for developing a unifying framework for designing threshold addressing the above issues. The proposed approach should be computationally efficient, distribution free, accurate, and robust to the dependence inherent to the sequence of statistics.
[**Our contributions:**]{} In this paper, we present a novel distribution-free approach to choosing a threshold for a broad class of algorithms given a pre-specified level, which we call the data-driven threshold machine (DTM). DTM takes a (possibly dependent but stationary) sequence $\{S_1, \ldots, S_n\}$ generated by a learning algorithm, a pre-specified level $\alpha$, and returns a threshold $x$ such that $\mathbb{P}\{\max_{i=1}^n S_i > x\} \leqslant \alpha$ (illustrated in Figure \[fig:dtm\] below).
\[fig:dtm\]
We make a connection between the threshold design and the extreme value theory (EVT) since the threshold design can be cast into a problem of determining the tail of the extreme over a random sequence. The classic literature of EVT [@Leadbetter83; @Hsing88] has been focused on developing the limiting distribution of the extreme and the estimation parts based on theory are written obscurely in various scatted places. In the past, EVT has been largely used for domains such as finance [@EVTRisk06; @Rocco12] and environmental science [@Smith89]. In this paper, we focus on estimation by using the forms of the limiting distributions from the classic references [@Leadbetter83; @Hsing88], but also take advantage of the advances in EVT [@Ferro03; @Suveges07] to handle dependent sequence via estimating the extremal index, which is quite important to settings such as scan statistics, online change-point detection, and extreme bandits, where dependent between the sequence of statistics can be significant. Moreover, EVT directly focuses on the tail of distribution, thus avoiding a drawback of using statistical approximation, whose estimates are somehow biased by the central part of the distribution. In principle, EVT-based estimates of threshold can be more precise.
In a nutshell, our approach is to relate threshold to the tail probability of an arbitrary dependence sequence. We leverage the forms of the limiting distribution functions to parameterize the tail probability using four parameters including the extremal index, which explicitly captures the dependence. DMT is a three-stage method. In Stage I, we bootstrap from the original sequence of samples to generate an sequence with the same marginal distribution. In Stage II, we estimate the parameters of the extreme value distributions, using the heights of the exceedance points in this sequence given a pre-specified threshold. In Stage III, we estimate the extremal index using the inter-exceedance times of the original sequence. To summarize, the features of DMT include
- DMT is [*distribution-free*]{} without making any parametric assumption about the sequence. To get around assuming the parametric distribution for data, we use the asymptotic distribution of the maximum of a sequence. The ideas leverage the powerful extreme type theorem, which states that the limiting distribution of the maxima will be one of the three distributions, Weibull, Gumbel or Fréchet law. Hence, this reduces the task of estimating the tail probability to a much simpler one of estimating three parameters of the extreme value distributions. The asymptotic kicks in with a moderate sample size [@Leadbetter83]. The samples are utilized to estimate these parameters as well as the [*extremal index*]{} described below, via a “Poisson process” trick: when the threshold value high, the exceedance events are rare and can be well modeled as a Poisson process.
- DTM is [*robust to dependence*]{} of the sequence. It can obtain accurate threshold even when the sequence is dependent and works well as long as the sequence does not have infinite memory.
- DTM is [*computationally efficient*]{}. Since it only takes the original sequence, without performing any Monte Carlo simulation. The main computation involves maximum likelihood estimation of fours parameters where many standard optimization procedures can be employed.
Closely related work
--------------------
Choosing threshold using EVT has been studied in [@threshold10]; however, they assume samples, which cannot be applied to the settings we consider here such as scan-statistic, change-point detection since the dependence in the sequence of statistics is very significant. In other settings, EVT has been used to understand the theoretical basis of machine learning algorithms: recognition score analysis [@ECCV10; @PAMI11], for novelty detection [@Novelty09], and for satellite image analysis [@Shao13].
Motivating examples
-------------------
[**Scan statistics [@Scan09].**]{} There are $n$ scanning regions, $S_t$ corresponds to a statistic formed for each region; an anomaly is detected when any of the regions has statistic exceeds the threshold. The probability is over the null distribution assuming there is no anomaly, and $\alpha$ is pre-specified type-I error or significance level. Thus, the definition of significance level is (\[key\_problem\]).
[**Online change-point detection [@Siegmund1985].**]{} Given a sequence of mutually independent data $\{x_1, x_2, \dots\}$, there may be a change-point such that the distribution of the data changes. Our goal is detect such a change as quickly as possible after it occurs. The well-known CUSUM procedure uses a log-likelihood ratio statistic $S_t:= \max_{k<t} \sum_{i=k+1}^t \ell(x_i)$ for each time $t$, where $\ell(x_i)$ is the log-likelihood for each individual sample and the maximizing over $k$ corresponds to searching for the unknown change-point location. The detection procedure is a stopping time $T = \inf\{t: S_t > x\}$. To control the false-alarm-rate, one will specify the so-called average-run-length (ARL) so that $\mathbb{E}_0(T) \leq {\rm ARL}$. It can be shown that $T$ is asymptotically exponential when $x$ is large [@SiegmundVenkatraman1995], and hence the ARL requirement can be translated into $\mathbb{P}_0\{\max_{1\leqslant i \leqslant n} S_i > x\} = 1-e^{-n/{\rm ARL}}$. The sequential change-point detection can be viewed as a special case of the sequential likelihood ratio test (SPRT) [@Siegmund1985], in which similar relations between the threshold and the specified levels occur.
[**Extreme bandits.**]{} The extreme bandits [@extremeBandits14], also known as the max-$K$ bandit in [@kArm05], models a scenario in outlier detection, security, and medicine. It considers the following learning setting. The learner chooses one arm each time and then receives only the sample for that arm. For each pull, the $k$th arm generates a reward following a distribution $f_k$ with unknown parameters. Let $S_{k, t}$ be the estimate for the true reward. The estimate for $S_{k, t}$, if using sliding window, will have non-negligible dependence. The performance of a learner is evaluated by the [*most extreme value*]{} she has found. In this setting, to use the classic upper confidence bound rule (see, e.g., [@UCB2014]), one has to find $\mathbb{P}_k\{\max_{1\leqslant i\leqslant t}S_{k, i} > x\} < \alpha$ for each arm $k$ for a pre-specified confidence level $1-\alpha$.
Data-Driven Threshold Machine
=============================
Given a sequence of (possibly dependent but stationary) observations of length $n$ $$\mathbb{S} = \{S_1, S_2, \ldots, S_n\}$$ generated as the output of a learning algorithm, our data-driven threshold machine (DTM) returns the threshold $x$ for a certain target level $\alpha$ in three steps:
- The algorithm first bootstraps (or samples with replacement) from the original sequence $\mathbb{S}$ to generate a new sequence $$\mathbb{S}^* = \{S_1^*, S_2^*, \dots, S_n^*\}.$$ Due to the sampling technique, the new sequence preserves the marginal distribution but breaks the local dependence in the original stationary sequence.
- The algorithm selects [*exceedant*]{} samples which are greater than a large pre-set cutoff value $u$ from $\mathbb{S}^*$. The index and height of these exceedant sample will follow a marked Poisson process approximately, and we use them to estimate the (location, scale, and type) parameters of the extreme value distribution. (Illustrated in Fig. \[demo1\]).
- The algorithm returns to the original sequence $\mathbb{S}$, and apply the same pre-set cutoff value $u$. This is based on the profound theory that the threshold exceeding events converges in distribution to a compound Poisson process [@Leadbetter83]. Then the algorithm estimates the extremal index to capture inter-dependence between samples, using the temporal intervals between adjacent exceedant points. (Illustrated in Fig. \[demo2\]).
The overall algorithm is summarize in Algorithm \[alg:Framwork\]. The DTM algorithm can be applied when:
- Sequence satisfies the so-called $\alpha$-mixing condition (\[condition\]) , which is a moderate requirement. Sequences that satisfy the condition (\[condition\]) include the $m$-th order Markov chain and the $m$-dependent sequence (i.e., two samples are independent if their indices are at least $m$ apart) [@Lehmann04]. Most machine learning algorithms with a finite memory of data will satisfy this requirement.
- Threshold $u$ should be chosen large enough so that the points exceed $u$ can be approximated as a Poisson process. Theorem 2.4.4 in [@Leadbetter83] states condition for the convergence of the exceedant points to a Poisson process. In practice, we choose $u$ as the .95 or .99 quantile of the data.
- The number of samples $n$ should be large enough, so that we have enough samples exceeding a large $u$ for the estimation to be accurate, and also that asymptotic distribution of the maximum converges. In theory, the number of samples $n$ should be at least $O(\tau^2 e^{-\tau})$, where $\tau = - \log (1-\alpha)$, as a consequence of Theorem 2.4.2 in [@Leadbetter83]. In practice, when the number of exceedant sample has a moderate size, say, 10 to 100, the estimate for the threshold will still be accurate.
A sequence $\mathbb{S} = \{S_1, S_2, \dots, S_n\}$;\
Tail probability level $\alpha$;\
Parameter $u$ to select exceedant sample. threshold $x$\
Bootstrap $\mathbb{S}$ to form sample $\mathbb{S}^*$.\
Select from $\mathbb{S}^*$ that exceeds $u$ and record their “time” (index) and heights $\{ (i_1, S^*_{i_1}), (i_2, S^*_{i_2}), \dots \}$;\
Use exceedant heights $\{S^*_{i_1}, S^*_{i_2}\ldots\}$ to estimate location $\hat{\mu}$, scale $\hat{\sigma}$, and shape $\hat{\xi}$ parameters that maximize the marked Poisson process likelihood function (\[likelihood\]); Select from $\mathbb{S}$ that exceeds $u$ and record their “time” (index) and heights $\{ (i_1, S_{i_1}), (i_2, S_{i_2}), \dots \}$;\
Use exceedant times $\{i_1, i_2, \ldots\}$ to estimate the extremal index $\hat{\theta}$ by maximizing the mixture model likelihood (\[likelihood\_2\]);\
$x = \hat{C}^{-1}(-(1/\hat{\theta})\log(1-\alpha))$ where $$\label{C_hat}
\vspace{-2mm}
\hat{C} =
\left\{
\begin{array}{ll}
\left[1 + \hat{\xi} \left( \frac{x -\hat{\mu}}{\hat{\sigma}} \right) \right]^{- \frac{1}{\hat{\xi}}},
& \hat{\xi} \neq 0 \\
\mbox{exp} \left\{ - \frac{x-\hat{\mu}}{\hat{\sigma}} \right\} & \hat{\xi} = 0.
\end{array}\right.
\vspace{-1mm}$$
Theoretical Derivation
======================
DMT is based on the profound extreme value theory. We will show why DMT works. We first present the background of extreme value theory. Then we present how to estimate the three parameters $\mu$, $\sigma$, and $\xi$ for the so-called extreme value distributions using the heights of exceedant sample. Finally, we present how to estimate the extremal index $\theta$ using the time intervals between exceedant samples.
Parametrizing tail probability for extreme value
------------------------------------------------
Essentially, the problem we want to solve is to estimate the tail probability of extreme values. Surprisingly, as we show in Theorem \[theorem\], these extreme value distributions will follow specific parametric forms irrespective of the original distribution for $S_t$ and the dependence structure. Hence, our problem can be tackled by estimating the parameters of these parametric distributions. We will first describe the mixing condition needed for the theorem.
A stationary sequence $\{S_1,\dots,S_n\}$ is said to satisfy the distributional mixing condition, if for any integers $ 1<i_1 < i_2< \dots < i_p< j_1 <j_2 < \dots < j_q<n$ for which $j_1-i_p>l$, and for any real $u$ $$\label{condition}
\begin{split}
&\left| \mathbb{P}\left\{ S_{i_1} \leq u, \dots, S_{i_p} \leq u, S_{j_1} \leq u, \dots, S_{j_q} \leq u \right \} \right. \\
& \quad - \mathbb{P}\left\{ S_{i_1} \leq u, \dots, S_{i_p} \leq u \right\} \cdot \\
&\quad\quad \quad \left. \mathbb{P} \left\{ S_{j_1} \leq u, \dots, S_{j_q} \leq u \right\} \right| \leq g(l),
\end{split}$$ where $g(l)\rightarrow 0$ as $l\rightarrow \infty$.
The distributional mixing condition is a mild condition, which ensures that the dependence between $S_i$ decay fast enough. It is satisfied in most learning scenarios. For instance, the sequence, the order $M$ Markov chain, and the $M$ dependent sequence all satisfy (\[condition\]) [@Lehmann04]. Most scan statistics satisfy (\[condition\]) since the detection statistics are computed locally and any statistic computed over non-overlapping regions are mutually independent. With the above mixing condition, we can state the following fundamental extreme type theorem [@Fisher28; @Gnedenko43; @Leadbetter83]
\[theorem\] Let $\{ S_1, \dots, S_n\}$ be a stationary process with marginal distribution $F$ and satisfying the distributional mixing condition (\[condition\]). Let $\{S_1^*, S_2^*, \dots\}$ be another sequence of independent variables with the same marginal distribution $F$. Let $$M_n = \max_{1 \leq t \leq n} S_t, \quad \mbox{and} \quad M_n^* = \max_{1 \leq t \leq n } S_t^*.$$ Then there exist a sequence of positive $\{a_n\}$ and positive $\{b_n\}$ such that $$\begin{split}
& \mathbb{P}\left\{ \frac{M_n^*-b_n}{a_n} \leqslant x \right\} \xrightarrow{n\to \infty} G(x) \quad \mbox{and}\\
& \mathbb{P}\left\{ \frac{M_n-b_n}{a_n} \leqslant x \right\} \xrightarrow{n\to \infty} [G(x)]^{\theta},
\end{split}$$ where $\theta \in (0, 1]$ is the constant called the extremal index. Depending on the marginal distribution $F$, $G(x)$ is a member of the generalized extreme-value-distribution parameterized as $$\begin{aligned}
\label{gevd}
G(x) =
\begin{cases}
\exp\{- \left[ 1 + \xi \left( \frac{x-\mu}{\sigma} \right) \right]^{-\frac{1}{\xi}} \}, & \xi \neq 0; \\
\exp\{ - e^{ - \frac{x-\mu}{\sigma} } \}, & \xi =0,
\end{cases}\end{aligned}$$ defined over the set $\{x: 1 + \xi (x-\mu) / \sigma >0\}$, with location parameter $\mu$, scale parameter $\sigma>0$, and shape parameter $\xi$: $\xi>0$ corresponds to Fréchet distribution, $\xi <0$ corresponds to Weibull distribution, and $\xi = 0 $ corresponds to the Gumbel distribution.
In plain words, this extreme type theorem states that for sequences, the extreme value has to converge to one of three functional forms of the extreme value distribution, under the so-called “distributional mixing condition”. For a dependence sequence, the asymptotic distribution can be constructed from an sequence with the same marginal distribution and a so-called extremal index $\theta$, which is related to the local dependence of the sequence $\{S_i\}$ at a extreme level. This theorem motivates our approach to estimate the tail probability in . Essentially, we will first construct an sequence to estimate the parameters in $G(x)$, and then estimate the extremal index $\theta$ using the original dependent sequence.
One may wonder how to find $a_n$ and $b_n$. In fact, it can be shown that $G( (x-b_n)/a_n)$ remains to be one of the three extreme value distributions just with different parameter values [@Leadbetter83]. Hence, we may estimate $\mathbb{P}\{M_n^* \leqslant x\}$ directly by estimating the three parameters of the extreme value distribution $G(x)$, without worrying about the specific form of $a_n$ and $b_n$.
Learning parameters for $G(x)$ {#sec:g}
------------------------------
Thus, given the observed data $\mathbb{S}=\{S_1, \dots, S_n\}$, which are dependent and stationary, we will first construct a sequence of data $\mathbb{S}^*=\{S_1^*, \dots, S_n^*\}$ with the same marginal distribution to learn the extreme value distribution $G(x)$. Thus, we will first bootstrap (or sample with replacement) from the original sequence $\mathbb{S}$ to generate the new i.i.d. sequence $\mathbb{S}^*$. This sampling scheme preserves the marginal distribution $F(x)$ but breaks the local dependence in the original stationary sequence.
Next, given $\mathbb{S}^*$, we choose a high cutoff value $u$ to obtain the sequence of exceedant samples (as illustrated in Figure \[demo1\]). In practice, $u$ is set to $.95$ or $.99$ quantile of the data. Let $n_u$ denote the random number of samples that exceed the cutoff $u$. Since this number depends on the choice of $u$, we use $u$ as the subscript. Let $\{i_1, i_2,\dots, i_{n_u}\}$ denote the index of these exceedant sample, and then $$\{S^*_{i_1}, S^*_{i_2}, \dots, S^*_{i_{n_u}}\}$$ are the selected exceedant points.
[**Marked Poisson process approximation.**]{} To estimate parameters for $G(x)$, the key idea is a “Poisson trick”: the normalized index of the exceedant sample can be approximated by a Poisson process, and the marks of the events will correspond to the heights of the exceedant sample. The precise statement can be found in Theorem 5.2.1 of [@Leadbetter83]. Below, we present a simple argument to show that the intensity of the process is related to the extreme value distribution $G(x)$.
Since $S_t^*$ is an sequence, we have that $\mathbb{P}\left\{M_n^*\leq u \right\} = F^n(u)$. Alternatively, based on Theorem \[theorem\], we have that for large $n$, $\mathbb{P}\left\{M_n^*\leq u \right\}\approx G(x)$. By relating these two, and taking log on both sides, we obtain $n\log F(u) \approx \log G(x)$. Furthermore, for large $u$, $F(u)$ is close to 1, and $\log F(u) \approx -(1-F(u))$ using Taylor expansion. Hence, we obtain $$1-F(u)\approx -(1/n) \log G(u),$$ which means that for every data point, the probability to exceed the threshold $u$ is $-\log(G(u))/n$, a small number for large $u$. If we define a point process $N_n$ on the unit interval $(0, 1]$ consisting of events corresponding to normalized index of the exceedant sample, $\{i_1/n,\ldots,i_{n_u}/n\}$, then the point process converges to a Poisson process with intensity equal to $n(-\log(G(u))/n) = -\log G(u)$.
Further taking into account the heights of the exceedant sample, we can model the sequence of pairs, $\{ (\smallfrac{i_1}{n},S^*_{i_1}),\ldots,(\smallfrac{i_{n_u}}{n},S^*_{i_{n_u}} )\}$, as a [*marked Poisson process*]{} where the heights corresponds to the markers of the events. The intensity measure of the process for any set $\mathcal A=[\tau,t] \times (x, \infty)$ is hen given by $$\begin{aligned}
\Lambda^*(\mathcal A) =
\begin{cases}
(t-\tau) \left[ 1+ \xi \left( \frac{x-\mu}{\sigma} \right) \right]^{-\frac{1}{\xi}}, & \xi \neq 0; \\
(t-\tau) e^{ - \frac{x-\mu}{\sigma} }, & \xi =0.
\end{cases}\end{aligned}$$ Taking derivative, for any $t$ and $x \geqslant u$, we have the intensity function of the process given by $$\begin{aligned}
\lambda^*(t, x) =
\begin{cases}
\sigma^{-1} \left[ 1+ \xi \left( \frac{x-\mu}{\sigma} \right) \right]^{-\frac{1}{\xi} -1} & \xi \neq 0; \\
\sigma^{-1} e^{ - \frac{x-\mu}{\sigma} } & \xi =0.
\end{cases}\end{aligned}$$
[**Likelihood function.**]{} Therefore, the likelihood function for $\Ecal=\{ (\smallfrac{i_1}{n},S^*_{i_1}),\ldots,(\smallfrac{i_{n_u}}{n},S^*_{i_{n_u}} )\}$ under the [*marked Poisson process*]{} model is given by $$\label{likelihood}
\begin{split}
& \mathcal{L}(\mu, \sigma, \xi; \Ecal) = \exp \left\{ -\Lambda_0^* \right\} \prod_{k=1}^{{n_u}} \lambda^*(\smallfrac{i_k}{n}, S^*_{i_k}) \\
& \propto
\left\{
\begin{array}{ll}
\exp \left\{ - \left[ 1 + \xi \left( \frac{u-\mu}{\sigma} \right) \right]^{-\frac{1}{\xi}} \right\}\cdot \\
\quad \prod_{k=1}^{{n_u}} \frac{1}{\sigma} \left[ 1+ \xi \left( \frac{S^*_{i_k}-\mu}{\sigma} \right) \right]^{-\frac{1}{\xi}-1}, &
\xi \neq 0 \\
\exp \left\{ - \exp \left\{ - \frac{u-\mu}{\sigma} \right\} \right\}
\prod_{k=1}^{{n_u}} \frac{1}{\sigma} \cdot & \\
\quad \exp \left\{ - \frac{S^*_{i_k}-\mu}{\sigma} \right\}, & \xi = 0.
\end{array}
\right.
\end{split}$$ where $\Lambda_0^*:=\Lambda^*((0,1]\times(u,\infty))$. From (\[likelihood\]), we find that the likelihood function only depends on the heights of the exceedant sample. Once $u$ is fixed, the index of the exceedant sample does not change the likelihood function.
Maximization of the likelihood function over the parameters does not lead to an analytical solution, but it can be done via standard optimization since only three variables are involved. Initialization is done with the method-of-moments, which relate the mean and variance of the exceedant sample to the three parameters-to-be-estimated, to avoid the discontinuity at $\xi = 0$. More details can be found in [@Coles03].
Learning extremal index $\theta$. {#sec:t}
---------------------------------
In this section, we focus on learning the extremal index $\theta$, which captures the dependence of the original sequence $\mathbb{S}$. Now we will apply the cutoff $u$ to $\mathbb{S}$ and obtain a new set of index $\{i_1,\ldots,i_{n_u}\}$, and the corresponding heights $$\{S_{i_1}, \ldots, S_{i_{n_u}} \}.$$ We will use the inter-exceedance times to estimate $\theta$, based on a theory in [@Hsing88].
[**Compound Poisson process approximation.**]{} Basically, when $\{S_1, \dots, S_n\}$ is stationary, the inter-exceedance times $\{i_1, i_2,\dots, i_{n_u}\}$ will converge to a [ *compound Poisson process*]{}. A compound Poisson process is a continuous-time stochastic process with jumps. The jumps occur randomly according to a Poisson process, and the size of the jumps is also random according to a probability distribution (as illustrated in Figure \[demo2\]).
Based on this theory, [@Ferro03; @Suveges07] give a more refined characterization. They proved that the limit distribution of the inter-exceedance times would be a mixture of an exponential distribution and a point mass on zero; the mixing proportion for the point mass will be equals to $\theta$. Intuitively, when there is a dependency in the sequence, even in the extremal level, the data points tend to exhibit a clustering structure. If one data point reaches a high level, then the successive data tend to reach a high level as well. Hence, $\theta$ characterizes the clustering behaviors of the data at the extreme level and it can be interpreted as the inverse of the limiting mean cluster size.
More specifically, let $T_k(u)$ denote the $k$-th inter-exceedance time, with $T_k(u) = i_{k+1}-i_k$, $k = 1, \dots, n_u-1$ (see Figure \[demo2\] for an illustration). When $T_k(u)-1$ is nonzero, then the value of $T_k(u)-1$ can be interpreted as a distance between two adjacent clusters. Let $F(u)$ be the marginal probability that $S_i\leq u$. [@Ferro03; @Suveges07] proved that when $n$ tends to infinity, the limiting distribution of the variable $(1-F(u)) (T(u)-1)$ converges a mixture distribution $$\label{compound_density}
\begin{split}
& \mathbb{P} \left\{ (1-F(u))(T(u)-1) \in (t, t+dt) \right\}\\
&= \left\{\begin{array}{ll}
1-\theta, & t = 0; \\
\theta^2 e^{-\theta t}, & t>0
\end{array}\right.
\end{split}$$ This means that with probability $\theta$ the inter-exceedance time is an exponential variable with rate $\theta$, and otherwise it is of length zeros. Note that all zero observations of $T_k(u)-1$ will attribute to the point mass component of the likelihood.
[**Likelihood function.**]{} Using (\[compound\_density\]), we can write the likelihood function of the sequence of inter-exceedance time, $\{T_1(u)-1,\ldots, T_{n_u-1}(u)-1\}$, from which we can estimate $\theta$ $$\label{likelihood_2}
\begin{split}
& \mathcal{L}( \theta; S_1, \dots, S_n)
= (1-\theta) ^{ (n_u - n_c - 1) } \theta^{2 n_c } \\
& \quad \mbox{exp} \left\{ -\theta \sum_{i=1}^{n_u-1}
(1-F(u))(T_i(u)-1)
\right\},
\end{split}$$ where $n_c =\sum_{i=1}^{n_u-1} \mathbb{I}\{(T_i-1) \neq 0\}$ corresponds to the number of non-zero inter-exceedance times, and $(1-F(u))$ can be replaced by its estimate $n_u/n$. A closed form expression for the maximum likelihood estimator $\hat{\theta}$ can be easily derived: $$\hat{\theta} = 1 - \frac{n_u - n_c - 1}{
2n_c - \sum_{j=1}^{n_u - 1}(1-F(u))(T_j(u) - 1)}.$$ Thus, this estimator of $\hat \theta$ together with the estimators for $\hat \mu$, $\hat \sigma$ and $\hat \xi$ from Section \[sec:g\] completes the major work of our data-drive threshold machine. Last, we set $G(x; \hat \mu, \hat \sigma, \hat xi)^{\hat \theta} = \alpha$ and solve for $x$ and obtain $x = {\hat C}^{-1} (-(1/{\hat \theta})\log(1-\alpha))$ as used in Algorithm \[alg:Framwork\].
Numerical Examples
==================
We will conduct two set of examples in this section investigating the accuracy of tail probability estimation and applying our method to a few machine learning problems.
Accuracy of tail probability modeling.
--------------------------------------
We study the accuracy of the DTM in estimating of $\mathbb{P}\{ \max_{1 \leq i \leq n} S_i \leq x \}$ by comparing with the simulation results. First, we generate a total number of $L$ sequences each with sample size $n$. For each sequence, we record the maximal value. Then for the $L$ sequences, we will have $L$ such maximums. In this way, we can get the empirical distribution for $\mathbb{P}\{ \max_{1 \leq i \leq n} S_i \leq x \}$. If $L$ is a large number, we can regard this empirical distribution as the true distribution. On the other hand, we apply our algorithm to just one sequence of data with sample size $n$, and select the data points exceeding the predetermined $u$ to fit the model. Then substitute the estimated parameters into the parametric form to get the approximation. Note that our algorithm only uses $1/L$ of the amount the data compared to simulation.
[**Adaptive to data distribution.**]{} We arbitrarily select one distribution from the three types of distributions, with exponentially decaying tails, heavy tails, and short tails respectively, and show DTM is agnostic to the underlying distribution. Specifically, we consider the following cases: (1) $S_i \sim Beta(2, 5)$, which is short-tailed and the random variables is upper bounded by 1; (2) $S_i \sim \chi^2$ with degree 1, which has exponentially decaying tail, and (3) $S_i \sim $ Student-$t$ with degree 4, which is a heavy tail distribution. Let $n=10^4$, $L=10^4$, and $u$ be the $.99$ quantile of the data. The comparison results of the empirical and the approximated distributions for $\mathbb{P}\{ \max_{ 1\leq i \leq n} S_i\}$, under the three cases are demonstrated in Figure [\[tail\]]{}. Note that our algorithm only uses $10^{-4}$ of the amount the data compared to simulation, but get almost the same results.
In this example, we consider the following cases: (1) $S_i \sim Beta(2, 5)$, which is short-tailed and the random variables is upper bounded by 1; (2) $S_i \sim \chi^2$ with degree 1, which has exponentially decaying tail, and (3) $S_i \sim $ Student-$t$ with degree 4, which is a heavy tail distribution. Let $n=10^4$, $L=10^4$, and $u$ to be the $.99$ quantile of the data. The results are demonstrated in Figure [\[tail\]]{}. Note that our algorithm only uses $10^{-4}$ of the amount the data compared to simulation, but get almost the same results. Moreover, our algorithm does not need to know any prior knowledge about the tail of the underlying distribution. That is, we don’t need to know whether the data $\{S_i\}$ are a heavy tail, short tail or exponentially decaying tails. The algorithm can adaptively and accurately learn this information from the data.
[**Adaptive to dependence.**]{} We study the accuracy of the DTM on stationary sequences $\{S_t\}$ with local dependence. Specifically, we consider the following random sequences $$S_t = e^{-1/m} S_{t-1}+ \sqrt{1- e^{-2/m}} Z_t$$ where $\{Z_t\}$ are independent standard normal variables. For this sequence, it has such properties that $\{S_t\}$ is a Gaussian process with $\mathbb{E}[S_t]=0$, and $\mbox{Cov} (S_t, S_{t'})= \mbox{exp}(-|t'-t|/m)$. By adjusting $m$, we can control the strength of local dependence. If $m=0$, $\{S_t\}$ is an sequence. Increasing $m$ will enhance the local dependence.
Consider $m=0$ and $m=50$, respectively. The values of $n, L$ are the same as previous examples. The comparison results of the empirical and the approximated distributions for $\mathbb{P}\{ \max_{ 1\leq i \leq n} S_i\}$ are displayed in Figure \[tail\_m\]. Our algorithm shows consistent results with simulation. The estimated extremal index are $\hat{\theta}=1.000$ and $\hat{\theta}=0.246$ in these two examples. We know that $\theta=1$ corresponds to the independent sequence, and increase the local dependence of the random process, $\theta$ will be more close to 0. This means, our algorithm can accurately learn the distribution $\mathbb{P}\{ \max_{ 1\leq i \leq n} S_i\}$ for dependent sequence. Moreover, we don’t need to know beforehand whether $\{S_i\}$ are independent or dependent. The algorithm can adaptively and accurately learn this information from data, reflected in the estimated values of $\theta$.
Application to choice of threshold
----------------------------------
[**Scan statistics over graph.**]{} We consider the problem of community detection which has been studied in [@randomgraph14]. The problem is cast into detecting a dense subgraph in a random graph. The null hypothesis is that the random graph is an Erdos-Renyi Graph, i.e., edges between nodes are Bernoulli random variables with probability $p_0$ being one. Alternatively, there is a subgraph such that the edges are formed with higher probability $p_1 > p_0$. Let $W_{ij}$ denote the adjacency matrix of the random. The scan test detects a community when the statistic $\max_{\mathcal G} \sum_{(i, j) \in {\mathcal G}} W_{ij} > x$, where $\mathcal{G}$ denotes a subgraph contains the community and $x$ is the threshold. Let $N$ be the number of nodes. If we assume the size of the community is $k$, there are $N \choose k$ such $\mathcal{G}$. Since $N \choose k$ is usually a very large number, we randomly pick $n$ possible $\mathcal{G}$ when forming the scan statistics.
We consider the case where $N = 100$, $p_0 = 0.1$, $k = 10$, and $n = 5000$. The Monte Carlo results are obtained from 100 repetitions of the experiments. As shown in Table \[table1\], the threshold obtained via DTM is consistent with and higher that obtained from Monte Carlo simulation (in fact, the Monte Carlo results, in this case, are obtained from a relatively small number of repetitions; hence the estimated thresholds from Monte Carlo tend to be small).
$\alpha$ 0.1 0.05 0.03 0.01
------------- ------- ------- ------- -------
Monte Carlo 12.00 12.00 13.00 13.00
DTM 13.71 14.50 14.64 14.55
: Scan over random graph, threshold obtained via Monte Carlo simulation versus DTM.
\[table1\]
[**Change-point using MMD statistic.**]{} We show that DTM can aid change-point detection in the online setting. In this example, the objective is to detect the activity changes over the network by monitoring the adjacency matrix $W$. Still let $N=100$, and the observations are a snapshot of a realization of the adjacency matrix with dimension 100 by 100. Let $p_0=0.3$ before the change-point and $p_1=0.4$ after the change-point. The true change-point occurs at time $4000$. We introduce the [*maximum mean discrepancy*]{} (MMD) as the detection statistic and use the online sliding window search scheme to monitor potential changes. The experiment setting is the same as [@Li15]. We set the block size $B=50$ and only use one block. Every time, MMD is formed by the to-be-test data $X$ within the sliding block with and the reference data $Y$ with the same size by $\text{MMD}^2[X, Y] = \frac{1}{B(B-1)} \sum_{i, j=1, i \neq j}^B h(x_i, x_j, y_i, y_j),$ where $h(x_i, x_j, y_i, y_j)=k(x_i, x_j)+k(y_i, y_j)-k(x_i, y_j)-k(x_j, y_i)$ and $k(\cdot)$ is the kernel function (we are using Gaussian kernel). It is well-known that MMD is a nonparametric statistic and the asymptotic distribution is an infinite summation of weighted chi-squares, which has no closed-form. When applying MMD to the online change-point detection, we need to characterize the tail probability of the maximal value of MMD over time, and from which get the threshold indicating when to stop the algorithm and make a decision that there is a change-point.
DTM provides a cheap and accurate approach to getting the threshold. As shown in Figure \[MMD\]. We first apply the detection algorithm on the raw data (realizations of adjacency matrix), and get a sequence of $\{ \text{MMD}_t\}$, which are our $\{S_t\}$ in DTM. Given $n=2000$ samples of $\{ {\text{MMD} }_{t}\}$, the estimated parameters are $(\hat{\sigma}, \hat{\xi}, \hat{\mu}, \hat{\theta})=(0.647, 0.000, 5.717, 0.306)$. Let ARL range from $.5\times 10^4$ to $1.5 \times 10^4$. The relation of threshold $x$ and ARL can be computed by $\mathbb{P}_0\{\max_{1\leqslant i \leqslant n} S_i > x\} = 1-e^{-n/{\rm ARL}}$. We compare this result with Monte Carlo, as shown on the right panel of Figure \[MMD\], and get the consistent results. For $\text{ARL}=.5\times 10^4$, we get the approximated threshold $x=5.54$ (this value for Monte Carlo is 5.35). We mark this threshold as the red line (as shown in the left panel of Figure \[MMD\]) and sequentially monitor the change-point. It shows that the threshold successfully detects the change-point occurring at time 4000.
Note that DTM works on the detection statistics directly, whereas for Monte Carlo method one needs to generate realizations of the adjacency matrix and form the MMD to get an approximation because the closed-form of MMD is unknown, which would be rather computationally expensive.
[**Max K-armed bandit.**]{} In the max K-armed bandit (or called extreme bandit) setting, the objective is to find the best arm defined as the arm with the heaviest tail. Consider the following setting and algorithm. There are $K$ arms, each with underlying distribution $P_k$. At any time $t$, the the policy only choose one arm $i$ to pull and get one observation of the reward defined as $R_{i, s}, s= 1, 2, \dots$. Let $N_i(t)$ denote the number of times that arm $i$ has been sampled up to time $t$. For any arm $i$, given the $N_i(t)$ observations till now, we define the upper confidence bound (UCB) $x_i^{\text{up}}$ to be: $P\left\{ \max_{ 1\leq s \leq N_i(t)} R_{i,s} > x^{\text{up}}_{i} \right\} =\delta$ and define the lower confidence bound (LCB) $x_i^{\text{low}}$ to be: $P\left\{ \max_{ 1\leq s \leq N_i(t)} R_{i,s} < x^{\text{low}}_{i} \right\} =\delta$, where $\delta$ is the confidence parameter. UCB and LCB play a crucial role in identifying the best arms in many online algorithm, such as Action Elimination (AE) algorithm, Upper Confidence Bound (UCB) algorithm, and LUCB algorithm [@Jamieson14]. For example, in the UCB algorithm, every time we choose to pull the arm with the highest UCB to get the reward. And stop the algorithm whenever the LCB for the best arm till now is higher than the UCBs for any other arms.
Our DTM provides a data-driven approach to estimating the UCB and LCB based on the real observed rewards to date, and can adaptively update the estimations given new observations. As a illustration, we consider the following example. We let $K=2$, and consider the Pareto distribution $P_k(x) =1-x^{-\alpha_k}$, where $\alpha = [3.5, 4.0]$. Fix $\delta=.005$. For the first example, rewards are generated from the underlying distributions. We first sample the two arms a fixed number of times. In the experiment, this number is 500. Then we can adopt DTM to estimate (LCB, UCB) from the history observations. The estimation results are (4.086, 8.001) for arm one, and (3.914, 7.172) for arm two. Then if we use the UCB algorithm, we will pull arm one in the next step. We find that the algorithm will stick to pulling arm one, which is the best arm. And we also demonstrate the adaptive UCB and LCB for arm one in Figure \[reward\].
Next, we consider the case where the observed rewards are stationary and temporal dependent for each arm. The setting is the same, however, the rewards for each arm is generated as a moving average of the first example. The introduced moving window would induce the dependence of the observations. Set the window to be 10. After 500 observations, the initial estimation for (LCB, UCB) is (2.439, 2.638) for arm one, and (1.614, 2.085) for arm two. We also demonstrate the adaptive UCB and LCB for arm one if we continue to pull it, as displayed in Figure \[reward\_dep\].
Conclusion
==========
We present a novel distribution-free approach, the data-driven threshold machine (DTM), to choose threshold such that the extreme values are bounded by a pre-specified level. DTM only requires one sample path for a reliable estimate of the threshold. Numerical examples demonstrate the robustness of the method. As of future work, our approach can be extended using the general Khintchine’s theorem to find a lower threshold for the lower tail.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'The X-Ray Telescope (XRT) on the Japanese/USA/UK [*Hinode (Solar-B)*]{} spacecraft has detected emission from a quiescent active region core that is consistent with nanoflare heating. The fluxes from 10 broadband X-ray filters and filter combinations were used to constructed Differential Emission Measure (DEM) curves. In addition to the expected active region peak at Log T $=$ 6.3-6.5, we find a high-temperature component with significant emission measure at Log T $>$ 7.0. This emission measure is weak compared to the main peak – the DEM is down by almost three orders of magnitude – which accounts of the fact that it has not been observed with earlier instruments. It is also consistent with spectra of quiescent active regions: no Fe XIX lines are observed in a CHIANTI synthetic spectrum generated using the XRT DEM distribution. The DEM result is successfully reproduced with a simple two-component nanoflare model.'
author:
- 'J.T. Schmelz, S.H. Saar, E.E. DeLuca, L. Golub, V.L. Kashyap, M.A. Weber, J.A. Klimchuk'
title: 'Hinode X-Ray Telescope Detection of Hot Emission from Quiescent Active Regions: A Nanoflare Signature?'
---
Introduction
============
If nanoflares heat the corona (Parker 1983, 1988; Klimchuk 2006), then we would expect an ever-present hot component with a temperature of 10 MK or higher, even in quiescent active regions. A glimpse of this hot material may have been seen by, for example, the Hard X-Ray Imaging Spectrometer on Solar Maximum Mission (Martens, van den Oord & Hoyng 1985) or the Mg XII imager onboard CORONAS-F (Grechnev et al. 2006). For the most part, however, it has only been possible to put an observational upper limit on the amount of this hot material because high temperature spectral lines like Ca XIX (Lemen, Bentley & Sylwester 1986; Fludra & Schmelz 1999) and Fe XIX (Schmelz 1993; Wang, Innes & Solanki 2006; Teriaca et al. 2006) were not seen except during solar flares. Even observations of the pervasive S XV line by the Yohkoh Bragg Crystal Spectrometer (Culhane et al. 1991) could be explained with cooler (3-4 MK) plasma (Watanabe et al. 1995; Sterling 1997).
Broadband imagers like the Yohkoh Soft X-Ray Telescope (Tsuneta et al. 1991) were designed to investigate hot, flaring plasma, but results tended to rely on filter ratios and an isothermal approximation to estimate plasma parameters. In this letter, we use data from the X-Ray Telescope (XRT; Golub et al. 2007) on Hinode (Kosugi et al. 2007). XRT is a broadband instrument similar to the Soft X-Ray Telescope, but with better spatial resolution, sensitivity, and temperature coverage. A detailed 3 MK thermal map of an active region observed by XRT has already been published by Reale et al. (2007) using an innovative combined filter ratio method. The properties of the instrument allow us to go one step further, however, and perform a true multithermal analysis. XRT should be able to observe the high temperature plasma predicted by nanoflare models, or at least determine a strict upper limit for the amount of emitting material.
Analysis
========
XRT observed AR 10955 (S09W30) on 2007 May 13 at 1800 UT. The data set includes full Sun, full resolution images in all filters as well as two filter combinations. The data were processed with standard XRT software available in SolarSoft (xrt\_prep.pro). This includes subtraction of a model dark/read-noise frame, correction for vignetting, removal of high frequency pattern noise, and normalization by exposure time. Spacecraft jitter was removed (xrt\_jitter.pro), and long and short exposures (see Table 1) were co-aligned and combined to increase image dynamic range. The data from the thickest channel (Be\_thick) were subject to additional Fourier filtering to remove some low-level, residual, longer-wavelength noise patterns. We have used updated filter calibrations (Narukage et al. 2009) and 1640 Å of diethylhexyl phthalate, a time-dependent contamination layer on the CCD, which can now be modeled using the make\_xrt\_wave\_resp.pro and make\_xrt\_temp\_resp.pro routines available in SolarSoft.
Figure 1 shows the XRT Ti\_poly image of AR 10955, which produced a small GOES A7 flare at 18:12 UT. Our observations indicate that the effects of this flare were limited to the left side of the active region, so we focused our attention on the outlined area. Even here, well away from the flaring material, there are significant counts in the Be\_thick channel. Since we are specifically searching for hot plasma, we restricted our analysis to those pixels within the outlined area having significant counts in Be\_thick. The noise in the Be\_thick image (on the disk but outside the active region and bright point areas) is: $-$0.00039$\pm$0.00245 Data Numbers s$^{-1}$ and we chose a threshold of $>$ 0.03 Data Numbers s$^{-1}$. Given the known response of XRT to optically thin thermal plasma, we can generate the possible Differential Emission Measure (DEM) curves that can reproduce the observed fluxes in 10 different XRT filters and filter combinations. We chose uncertainties of 3%, which result from comparing the Monte-Carlo realizations of the full area outlined in Figure 1 with the individual pixel-by-pixel results: the 1 $\sigma$ spread in the two DEM distributions is similar. (This result will be examined in greater detail in a future paper.)
The orange (heavy dashed) curve in Figure 2a shows the [*optimal median*]{} of the Monte-Carlo iterations generated by the Markov-Chain Monte Carlo (MCMC) based reconstruction algorithm (Kashyap & Drake 1998). The optimal median is the iteration best matching the median DEM solution in each temperature bin, after discarding non-convergent solutions. The routine fits a locally smoothed DEM curve to the data by comparing the predicted to the observed fluxes and modifying the solution randomly to obtain new realizations. The smoothing scale varies with temperature to account for changes in the information content available as codified in the filter response curves. The new realizations are kept or discarded according to the Metropolis criterion based on changes in the $\chi^2$ values at each step, resulting in a Markov-Chain that explores the parameter space efficiently. This results in a sample of solutions that is representative of the actual probability distribution of the DEM (see e.g., Smith & Roberts 1993; Casella 1996). Many of the details and advantages of this method, including DEM dynamic range, uncertainties, and smoothing, were discussed by Schmelz, Kashyap & Weber (2007) and references therein. The predicted-to-observed flux ratio for each filter is listed in Table 1.
The yellow (heavy solid) curve in Figure 2a is generated with the same XRT data, but using xrt\_dem\_iterative2.pro. Testing and validation of the method with synthetic data leading up to the launch of Hinode were presented by Golub et al. (2004) and Weber et al. (2004). The routine employs a forward-fitting approach where a DEM is guessed and folded through each response to generate predicted fluxes. This process is iterated to reduce the $\chi^2$ between the predicted and observed fluxes. The DEM function is interpolated using several spline points, which are directly manipulated by mpfit.pro, a well-known and much-tested IDL routine that performs a Levenberg-Marquardt least-squares minimization (Levenberg 1944; Marquardt 1963). There are N$_i$ - 1 splines, representing the degrees of freedom for N$_i$ observations. This routine uses Monte-Carlo iterations to estimate errors on the DEM solution (grey dashes in Figure 2a). For each iteration, the observed flux in each filter was varied randomly and the program was run again with the new values. The distribution of these variations was Gaussian with a centroid equal to the observed flux and a width equal to the uncertainty (3%). The predicted-to-observed flux ratios are also listed in Table 1. We are pleased to find that the two methods are in excellent agreement: the xrt\_dem\_iterative2.pro results are within 1$\sigma$ of the MCMC results for five of the XRT filters and within 2$\sigma$ for the remaining filters.
To evaluate the quality of DEM reconstruction that can be achieved with XRT, we have compared the results of xrt\_dem\_iterative2.pro with a known input model. In each case, a nominal observation is calculated for a given DEM model, and then the procedure samples the observation 100 times by including random photon noise at the 3% level. The distribution of calculated DEM curves (relative to the known DEM model) indicates the accuracy and robustness of the analysis method. Figure 3 (top) shows the results for the isothermal cases: Log T $=$ 6.1 and 6.4. Each plot shows the input delta function (solid line), the 100 Monte-Carlo realizations (grayscale), and the median of all realizations for each temperature bin (diamonds). The results use all nine filters and demonstrate that the routine can determine the correct result by placing most of the power in the temperature bin corresponding to the delta function model, with a smaller amount placed in the adjacent bins. Recall that the fitted DEM is a spline curve through several knots; although this is not the best function for fitting a delta function, our results are still quite good. For a multi-thermal example, we analyzed the active region model from CHIANTI (v.5.2.1; Dere et al. 1997; Landi et al. 2006). The results suggest how well the input DEM can be reconstructed as a function of the number of observing channels used. Figure 3 (bottom) shows the model AR DEM (solid line with two humps). The first panel uses only the four thickset XRT filters. It can determine the presence of the hotter peak, as indicated by the convergence of the median fit to the model DEM curve, but fails to detect the cooler material. The last panel uses eight filters. The cooler model component is detected, but not accurately represented, which is consistent with the results from Figure 2a where the Monte-Carlo iterations begin to diverge at low temperatures.
The DEM realizations seen in Figure 2a are constrained by 10 XRT filters and filter combinations available for the active region core. The resulting curves show the expected component at Log T $=$ 6.3-6.5, but there is also a high-temperature component associated with this emission that is down by almost three orders of magnitude from the main peak. This component is exactly the type of feature that we would expect if the active region were heated by nanoflares. The lower emission measure explains why it has not been detected by other instruments – the [*dynamic range*]{} of a typical DEM curve is approximately two orders of magnitude. Any features beyond this range would, in general, not be trusted. With XRT and its high sensitivity to hot plasma, however, this restriction does not apply. The peak filter responses span almost four orders of magnitude, so the DEM dynamic range is substantially greater than previous instruments. One important point, however, is that the observations do not require that the DEM curve be bimodal. There is a possibility (see grey Monte-Carlo iterations) that the DEM decreases monotonically beyond Log T $=$ 6.4. This has implications for the possible distribution of nanoflare energies (see below). We also note that all DEM results depend heavily on instrument calibration. A lot of effort has gone into determining the XRT filter thicknesses (Narukage et al. 2009), but problems may still remain. One of the advantages of DEM over traditional filter ratio methods is that these results are not as sensitive to an uncertainty in one particular filter. For example, both DEM methods used here indicate that the predicted flux for the Thick\_Al filter is too high, a problem that could be related to the adopted thickness of the filter. Small changes to the filter thickness change the amount of predicted flux (and therefore the goodness of fit) but not the DEM shape.
One obvious concern is that the high temperature plasma revealed in these DEM curves might result from the GOES A7 flare from the left side of AR 10955. The thickest filters, however, were observed well before the flare (see Table 1), thus reducing the likelihood of significant contamination. To address any residual concern, we have generated synthetic spectra using the ch\_ss.pro routine available in CHIANTI. We have chosen the wavelength range of 13-14 Å because it contains several high temperature iron lines that can be used as a discriminator between flaring and quiescent plasma. The yellow DEM curve from Figure 2a (the orange curve gets similar results) was used to produce the dotted spectrum seen in Figure 4. The Ne IX triplet lines (13.45, 13.55, 13.70 Å) with a peak formation temperature of Log T $=$ 6.2 as well as an Fe XVII line (13.82 Å; Log T $=$ 6.6) are clearly visible. These lines are observed routinely from quiescent active regions (Schmelz et al. 1996), but there is no hint of the hotter Fe XIX line (13.52 Å; Log T $=$ 6.9). The solid spectrum was generated with a DEM from the decay phase of a GOES B8 flare (Reeves et al. 2009). The strongest Fe XIX line is now obvious, as are several smaller Fe XIX and Fe XX flare lines which are all missing from the active region core spectrum. These results indicate that not only is the hot DEM component observed by XRT consistent with earlier observations from spectrometers, but more importantly, it is unlikely to have resulted from the flare. (Note: direct comparison of the AR core and A7 flare spectra would have been optimal, but lack of the thick filter observations during the flare prevented accurate flare DEM analysis; see Table 1 for the exact times.)
We have modeled the observed DEM curve with nanoflares using the EBTEL hydrodynamic simulation code (Klimchuk, Patsourakos & Cargill 2008). Our goal at this time is not to accurately reproduce the details of the curve, but rather to demonstrate that impulsive heating can explain the overall shape. We therefore make a number of simplifying assumptions. First, we assume the observed volume outlined in Figure 1 is comprised of unresolved loop strands with a half length of 60 Mm. Second, we assume that there are just two types of nanoflares. Most of the strands are heated by weak nanoflares that repeat every $4.0\times10^3$ s, while 0.3% of the strands are heated by much stronger nanoflares that repeat every $3.0\times10^5$ s. The model DEM curve is obtained by time averaging the weak and strong nanoflare simulations over these durations and combining in the appropriate proportion. The nanoflares have a triangular heating profile with a duration of 500 s and amplitude that corresponds to a time-averaged energy flux through the footpoints of $5.625\times10^6$ erg cm$^{-2}$ s$^{-1}$ and $3.0\times10^7$ erg cm$^{-2}$ s$^{-1}$ for the weak and strong events, respectively. This time average includes the long interval between successive nanoflares. Note that the canonical observed radiative loss rate from active regions is $10^7$ erg cm$^{-2}$ s$^{-1}$ (Withbroe & Noyes 1977). In addition to the nanoflares, there is a very weak steady background heating that corresponds to a static equilibrium with a temperature of 0.45 MK. All strands cool fully (i.e., to 0.45 MK) before being reheated by the next nanoflare, but the plasma drains slowly enough that the density in the weak nanoflare strands never drops below 15 times the equilibrium density.
The DEM distribution predicted by the model is given in Figure 2b. The solid curve is from the coronal part of the strands, while the dashed curve includes both the corona and the transition region footpoints (moss). The agreement with the observed curve in Figure 2a is remarkably good given the simplicity of the model. It is clear that nanoflare heating is consistent with the observations. Note that XRT does not provide significant constraints for plasma cooler than about 1 MK (for our filter set), so disagreement with this part of the curve is not meaningful.
Nonequilibrium ionization will reduce the high temperature DEM derived from observations relative to the actual values. However, we do not expect the effect to be large for 500 s nanoflares like those in our simulations (Bradshaw & Cargill 2006; Reale & Orlando 2008). The temperature cooling timescale in the strong nanoflare simulation is always greater than 500 s.
One can ask whether steady heating is also consistent with the observations. It seems likely that a suitable distribution of steady heating could reproduce the DEM curve. However, the extreme level of steady heating required to maintain the hottest stands is highly implausible. For example, an energy flux of several times $10^9$ erg cm$^{-2}$ s$^{-1}$ is required to produce a static equilibrium at 20 MK. This is more than two orders of magnitude greater than the canonical value. If the energy for the heating is provided by the stressing of the magnetic field by footpoint motions, then a steady photospheric velocity of order 100 km s$^{-1}$ is required. Such a flow is completely unsupported by observations. Wave heating is also implausible. Even with a wave speed as large as 1000 km s$^{-1}$, the fluctuation velocity would need to be of order 1000 km s$^{-1}$ to provide the necessary energy flux (assuming that such waves could dissipate fully in the low corona, which is unlikely). The corresponding displacement amplitude is totally unrealistic for wave periods $>$ 10 s. Thus, we conclude that while the cooler plasma in the DEM distribution could be due to either impulsive or steady heating, the hotter plasma can only be a result of nanoflares.
Conclusions
===========
The DEM distribution generated for a quiescent active region core shows a high-temperature component with significant hot (Log T $>$ 7.0) emission measure. XRT is sensitive to this plasma because it has a greater DEM dynamic range (almost four orders of magnitude) than any earlier solar X-ray imager or spectrometer. The hot plasma result is consistent with observations from spectrometers in that the synthetic spectrum generated using the XRT active region core DEM shows no Fe XIX lines, which would indicate flaring plasma. The EBTEL code was used to successfully model the result. For these reasons, we suggest that this might be the first convincing evidence of a nanoflare signature.
This work was inspired by a talk given by Fabio Reale at the September 2008 Hinode-2 Meeting in Boulder, CO. We would like to thank Fabio, as well as Paola Testa, John Raymond, and Nancy Brickhouse for helpful discussions. Hinode is a Japanese mission developed and launched by ISAS/JAXA, with NAOJ as domestic partner and NASA and STFC (UK) as international partners. It is operated by these agencies in co-operation with ESA and the NSC (Norway). Solar physics research at the University of Memphis is supported by a Hinode subcontract from NASA/SAO as well as NSF ATM-0402729.
Bradshaw, S. J. & Cargill, P. J. 2006, , 458, 987 Casella, G. 1996, Test, 5, 249 Culhane, J.L et al. 1991, , 136, 89 Dere, K.P., Landi, E., Mason, H.E., Monsignori Fossi, B.C., Young, P.R. 1997, , 125, 149 Fludra, A. & Schmelz, J.T. 1999, , 348, 286 Golub, L. et al. 2007, , 2007, 243, 63 Golub, L., Deluca, E.E., Sette, A., Weber, M. 2004, ASP Conference Series, Tokyo, Japan. Eds. T. Sakurai and T. Sekii, 325, 217 Kashyap, V. & Drake, J.J. 1998, , 503, 450 Klimchuk, J.A. 2006, , 234, 41 Klimchuk, J.A., Patsourakos, S., Cargill, P.J. 2008, , 682, 1351 Kosugi, T. et al. 2007, , 243, 3 Landi, E., Del Zanna, G., Young, P.R., Dere, K.P., Mason, H.E., Landini, M. 2006, , 162, 261 Lemen, J.R., Bentley, R.D., Sylwester, J. 1986, Adv. Spa. Res, 6, 245 Levenberg, K. 1944, The Quarterly of Applied Mathematics 2, 164 Marquardt, D. 1963, SIAM Journal on Applied Mathematics 11, 431 Narukage, N. et al. 2009, in prep. Parker, E.N. 1983, , 264, 642 Parker, E.N. 1988, 330, 474 Reale, F. & Orlando, S. 2008, , 284, 715 Reale, F. et al. 2007, Science, 318, 1582 Reeves, K. et al. 2009, in prep. Schmelz, J.T. 1993, , 408, 373 Schmelz, J.T., Kashyap, V. L. & Weber, M.A. 2007, , 660, L157 Schmelz, J.T., Saba, J.L.R., Ghosh, D. & Strong, K.T. 1996, , 473, 519 Smith, A.F.M. & Roberts, G.O. 1993, Journal of the Royal Statistical Society, Series B, 55, 3 Sterling, A. 1997, Geo Res. Let., 24, 2263 Teriaca, L., Falchi, A., Falciani, R., Cauzzi, G., Maltagliati, L. 2006, , 455, 1123 Tsuneta, S. et al. 1991, , 136, 37 Wang, T.J. Innes, D.E., Solanki, S.K. 2006, , 455, 1105 Watanabe, T et al. 1995, , 157, 169 Weber, M.A., DeLuca, E.E., Golub, L. & Sette, A.L. 2004, Proc. IAU Symp. 223, Multi- Wavelength Investigations of Solar Activity, 321 Withbroe, G.L.& Noyes, R.W. 1977, , 15, 363
[lcccccc]{} Al\_mesh &18:04:43 &4.10 sec &18:03:47 &0.18 sec &1.38$\pm$.04 &1.33$\pm$.04\
Al\_poly &18:17:06 &5.80 sec &18:16:38 &0.51 sec &——$^1$ &——$^1$\
C\_poly &18:10:00 &8.20 sec &18:09:01 &0.51 sec &1.17$\pm$.03 &1.12$\pm$.03\
Ti\_poly &18:03:27 &8.20 sec &18:03:17 &0.51 sec &0.86$\pm$.02 &0.82$\pm$.02\
Al\_poly-Ti &18:06:33 &16.4 sec &18:05:23 &1.45 sec &0.86$\pm$.02 &0.81$\pm$.02\
C-Ti &18:08:17 &16.4 sec &18:07:17 &1.03 sec &0.80$\pm$.02 &0.75$\pm$.02\
Be\_thin &18:12:31 &23.1 sec &18:11:47 &1.03 sec &1.47$\pm$.04 &1.37$\pm$.04\
Be\_med &18:14:06 &46.3 sec &18:13:13 &2.05 sec &1.16$\pm$.03 &1.09$\pm$.03\
Al\_med &18:15:39 &46.3 sec &18:15:04 &2.05 sec &1.26$\pm$.03 &1.18$\pm$.03\
Al\_thick &18:02:19 &46.3 sec &18:01:54 &16.4 sec &1.99$\pm$.07 &1.87$\pm$.07\
Be\_thick &18:00:37 &65.5 sec & —— & —— &1.09$\pm$.03 &1.03$\pm$.03\
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'We define the quadratic algebra $\su(2)_\alpha$ which is a one-parameter deformation of the Lie algebra $\su(2)$ extended by a parity operator. The odd-dimensional representations of $\su(2)$ (with representation label $j$, a positive integer) can be extended to representations of $\su(2)_\alpha$. We investigate a model of the finite one-dimensional harmonic oscillator based upon this algebra $\su(2)_\alpha$. It turns out that in this model the spectrum of the position and momentum operator can be computed explicitly, and that the corresponding (discrete) wavefunctions can be determined in terms of Hahn polynomials. The operation mapping position wavefunctions into momentum wavefunctions is studied, and this so-called discrete Hahn-Fourier transform is computed explicitly. The matrix of this discrete Hahn-Fourier transform has many interesting properties, similar to those of the traditional discrete Fourier transform.'
---
[**The $\su(2)_\alpha$ Hahn oscillator and a discrete Hahn-Fourier transform**]{}\
[**E.I. Jafarov[^1], N.I. Stoilova[^2] and J. Van der Jeugt**]{}\
Department of Applied Mathematics and Computer Science, Ghent University,\
Krijgslaan 281-S9, B-9000 Gent, Belgium\
E-mail: [email protected], [email protected] and [email protected]
Short title: $\su(2)_\alpha$ Hahn oscillator
PACS numbers: 03.67.Hk, 02.30.Gp
Introduction
============
In standard theory of quantum mechanics, position and momentum operators are (essentially) self-adjoint operators in some infinite-dimensional Hilbert space, satisfying the canonical commutation relations. Quantum mechanics in finite dimensions has attracted much attention in recent years [@Vourdas]. In a finite-dimensional Hilbert space, the canonical commutation relations no longer hold. Despite this, finite-dimensional quantum mechanics has been useful in areas such as quantum computing and quantum optics [@Braunstein; @Miranowicz]. The defining relations for a quantum mechanical oscillator in a finite-dimensional Hilbert space are not unique [@Santhanam], and several type of models have been proposed. Our interest comes mainly from models related to some Lie algebra (or a deformation thereof). The finite oscillator model that has been studied most extensively is based on the Lie algebra $\su(2)$ or $\so(3)$ in the one-dimensional case [@Atak2005; @Atak2001; @Atak2001b]. In the case of a two-dimensional oscillator this has been generalized by the same authors to $\so(4)$. Such finite oscillator models are of particular importance in optical image processing [@Atak2005], and more generally in models where only a finite number of eigenmodes can exist. For example signal analysis dealing with a finite number of discrete sensors or data points led to physical models realizing a one-dimensional finite oscillator [@Atak1994; @Atak1997; @Atak1999b]. In a previous paper [@JSV2011], a new model for the finite one-dimensional harmonic oscillator was proposed based on the algebra $\uu(2)_\alpha$, a one-parameter deformation of the Lie algebra $\uu(2)$. This $\uu(2)_\alpha$ model offers an alternative position and momentum spectrum compared to the $\su(2)$ model. Furthermore, the position wavefunctions have simple expressions in terms of Hahn polynomials, with interesting properties related to those of a parabose oscillator [@JSV2011].
For a one-dimensional finite oscillator, one considers three (essentially self-adjoint) operators: a position operator $\hat q$, its corresponding momentum operator $\hat p$ and a (pseudo-) Hamiltonian $\hat H$ which is the generator of time evolution. These operators should satisfy the Hamiltonian-Lie equations (or the compatibility of Hamilton’s equations with the Heisenberg equations): $$[\hat H, \hat q] = -i \hat p, \qquad [\hat H,\hat p] = i \hat q,
\label{Hqp}$$ in units with mass and frequency both equal to 1, and $\hbar=1$. Furthermore, one requires [@Atak2001]:
- all operators $\hat q$, $\hat p$, $\hat H$ belong to some (Lie) algebra (or superalgebra) $\cal A$;
- the spectrum of $\hat H$ in (unitary) representations of $\cal A$ is equidistant.
The case with ${\cal A}= \su(2)$ (or its enveloping algebra) has been treated considerably in a number of papers [@Atak2001; @Atak2001b; @Atak2005]. The relevant representations are the common $\su(2)$ representations labelled by an integer or half-integer $j$. In such a representation, the Hamiltonian is taken as $\hat H=J_0+j+\frac12$, where $J_0=J_z$ is the diagonal $\su(2)$ operator. Thus the spectrum of $\hat H$ is $n+\frac12$ ($n=0,1,\ldots,2j$). With $\hat q = \frac12(J_++J_-)=J_x$ and $\hat p = \frac{i}{2}(J_+-J_-)=-J_y$, the relations are satisfied. Clearly, $\hat q$ and $\hat p$ have a finite spectrum given by $\{-j,-j+1,\ldots,+j\}$ [@Atak2001]. More important, the position wavefunctions have been constructed, and are given by Krawtchouk functions (normalized symmetric Krawtchouk polynomials). These discrete wavefunctions have interesting properties, and their shape is reminiscent of those of the canonical oscillator [@Atak2001]. This is explained by the fact that under the limit $j\rightarrow \infty$ the discrete wavefunctions coincide with the canonical wavefunctions in terms of Hermite polynomials [@Atak2001; @Atak2003].
In [@JSV2011], the case with ${\cal A}=\uu(2)_\alpha$ was investigated ($\alpha>-1$). For this one-parameter deformation of $\uu(2)$, only the representations labelled by a half-integer $j$ survive as representations of $\uu(2)_\alpha$ (so only the even-dimensional representations). This led to an alternative model of the finite oscillator, with the spectrum of $\hat H$ again given by $n+\frac12$ ($n=0,1,\ldots,2j$). For the operators $\hat q$ and $\hat p$, the spectrum is again finite and equidistant in steps of one unit, except that there is a gap of size $2\alpha+2$ in the middle; explicitly, it is given by $$-\alpha-j-\frac12, -\alpha-j+\frac12, \ldots, -\alpha-1; \alpha+1,\alpha+2, \ldots,\alpha+j+\frac12.
\label{q-spectrum-even}$$ An interesting result in [@JSV2011] was that the position wavefunctions could be constructed and they turned out to be normalized Hahn (or dual Hahn) polynomials. An investigation of these discrete wavefunctions gave rise to remarkable plots, suggesting in fact a relation with the parabose oscillator. Indeed it was shown that under the limit $j\rightarrow \infty$ the discrete wavefunctions coincide with the parabose wavefunctions in terms of Laguerre polynomials [@JSV2011].
Despite the novel results presented in [@JSV2011], it remains somehow intriguing that the odd-dimensional $\uu(2)$ representations can not be deformed as $\uu(2)_\alpha$ representations. In the present paper, we present the solution to this problem. It turns out that one should consider a different one-parameter deformation of $\su(2)$, involving only the three $\su(2)$ operators and a parity operator (and no central element). This new deformation, denoted by $\su(2)_\alpha$, is defined and it is shown that in this case the odd-dimensional $\su(2)$ representations, labelled by an integer value $j$, can be deformed as $\su(2)_\alpha$ representations. Using $\su(2)_\alpha$ as a model for the finite oscillator, yields the same spectrum of $\hat H$ given by $n+\frac12$ ($n=0,1,\ldots,2j$). The finite spectrum of the position and momentum operators $\hat q$ and $\hat p$ is however quite different, and given by $$0, \quad \pm \sqrt{k(2\alpha+k+1)}, \qquad (k=1,\ldots,j).$$ The position (and momentum) wavefunctions are constructed, and turn out to be again Hahn polynomials (in this case with parameters $(\alpha,\alpha)$ or $(\alpha+1,\alpha+1)$). These wavefunctions are once more discrete versions of the parabose wavefunctions, as a limit computation shows. A fascinating question in this context is the relation between the position wavefunctions and the momentum wavefunctions. In the canonical case, these functions are related through the Fourier transform. Here, we show that these wavefunctions are related through a discrete version of the Fourier transform, which we refer to as the discrete Hahn-Fourier transform. For readers primarily interested in special functions, the computation of this discrete Hahn-Fourier transform is probably the most attractive part. It involves a special case of a bilinear generating function (the Poisson kernel) for dual Hahn polynomials.
The contents of the remaining sections is as follows: in section 2 the deformed algebra $\su(2)_\alpha$ and its representations are constructed. In section 3 we use $\su(2)_\alpha$ as a model for the finite oscillator, and determine in particular the spectrum of the position and momentum operators, and their eigenvectors. The structure of these eigenvectors is studied in section 4, yielding position and momentum wavefunctions. In section 5 we determine the operation that transforms position wavefunction into momentum wavefunctions, the so-called discrete Hahn-Fourier transform $F$, and its properties. The somewhat technical but interesting proof of the ${}_4F_3(1)$ form of the matrix elements of $F$ is given in section 6. Finally we give a summary of the results in section 7.
The algebra $\su(2)_\alpha$ and its representations
===================================================
The Lie algebra $\su(2)$ [@Wybourne; @Humphreys] can be defined by its basis elements $J_0$, $J_+$, $J_-$ with commutators $[J_0,J_\pm]=\pm J_\pm$ and $[J_+,J_-]=2J_0$. The non-trivial unitary representations of $\su(2)$, corresponding to the star relations $J_0^\dagger=J_0$, $J_\pm^\dagger=J_\mp$, are labelled [@Wybourne; @Humphreys] by a positive integer or half-integer $j$ and have dimension $2j+1$. The action on a standard basis vectors $|j,m\rangle$ (with $m=-j,-j+1,\ldots,+j$) is given by $$J_0 |j,m\rangle = m\;|j,m\rangle,\qquad
J_\pm |j,m\rangle = \sqrt{(j\mp m)(j\pm m +1)}\;|j,m\pm 1\rangle.$$ The universal enveloping algebra of $\su(2)$ can be extended by a parity operator $P$ with action $P |j,m\rangle = (-1)^{j+m}\;|j,m\rangle$. This means that $P$ commutes with $J_0$, anticommutes with $J_+$ and $J_-$, and $P^2=1$. This extended algebra can be deformed by a parameter $\alpha$, leading to the definition of $\su(2)_\alpha$.
Let $\alpha$ be a parameter. The algebra $\su(2)_\alpha$ is a unital algebra with basis elements $J_0$, $J_+$, $J_-$ and $P$ subject to the following relations:
- $P$ is a parity operator satisfying $P^2=1$ and $$[P,J_0]=PJ_0-J_0P=0, \qquad \{P,J_\pm\}=PJ_\pm + J_\pm P= 0.
\label{P}$$
- The $\su(2)$ relations are deformed as follows: $$\begin{aligned}
& [J_0, J_\pm] = \pm J_\pm, \label{J0J+} \\
& [J_+, J_-] = 2 J_0 + 2(2\alpha+1)J_0P.
\label{J+J-}\end{aligned}$$
Note that this is different from the $\uu(2)_\alpha$ deformation introduced in [@JSV2011]; in particular it does not involve a central element. For the deformation $\uu(2)_\alpha$, the even-dimensional $\uu(2)$ representations ($j$ half-integer) could be deformed. The current $\su(2)_\alpha$ is “complementary” in the sense that now the odd-dimensional $\su(2)$ representations ($j$ integer) can be deformed:
Let $j$ be an integer (i.e. $2j$ is even), and consider the space $W_j$ with basis vectors $|j,-j\rangle$, $|j,-j+1\rangle$, $\ldots$, $|j,j\rangle$. Assume that $\alpha>-1$. Then the following action turns $W_j$ into an irreducible representation space of $\su(2)_\alpha$. $$\begin{aligned}
& P |j,m\rangle = (-1)^{j+m}\;|j,m\rangle,\label{act-P}\\
& J_0 |j,m\rangle = m\;|j,m\rangle,\label{act-J0}\\
& J_+ |j,m\rangle =
\begin{cases}
\sqrt{(j-m)(j+m +2\alpha +2)}\;|j,m+1\rangle, & \text{if $j+m$ is even;}\\
\sqrt{(j-m+2\alpha+1)(j+m+1)}\;|j,m+1\rangle, & \text{if $j+m$ is odd,}
\end{cases} \label{act-J+}\\
& J_- |j,m\rangle =
\begin{cases}
\sqrt{(j+m +2\alpha +1)(j-m+1)}\;|j,m-1\rangle, & \text{if $j+m$ is odd;}\\
\sqrt{(j+m)(j-m+2\alpha+2)}\;|j,m-1\rangle, & \text{if $j+m$ is even.}
\end{cases} \label{act-J-}\end{aligned}$$
The proof is essentially by direct computation, the same as [@JSV2011 Proposition 2]. Note that the representation given in this proposition is unitary under the star conditions $P^\dagger=P$, $J_0^\dagger=J_0$, $J_\pm^\dagger=J_\mp$. Also note that for $\alpha=-\frac12$, the deformation is trivial (both for the algebra and the representations).
Using $\su(2)_\alpha$ as a model for the one-dimensional oscillator
===================================================================
Quite similar as in the non-deformed case [@Atak2005] or in the $\uu(2)_\alpha$ deformed case [@JSV2011], let us choose the position, momentum and Hamiltonian (representation dependent) operators as follows: $$\hat q = \frac12 (J_++J_-), \qquad
\hat p = \frac{i}{2}(J_+-J_-), \qquad
\hat H = J_0+j+\frac12 .$$ These operators satisfy . In the representation space $W_j$, $\hat H |j,m\rangle = (m+j+\frac12)|j,m\rangle$, therefore the spectrum of $\hat H$ is linear and given by $$n+\frac12 \qquad (n=0,1,\ldots,2j).$$ From the actions -, one finds $$2\hat q |j,m\rangle = \sqrt{(j+m)(j-m +2\alpha +2)}\;|j,m-1\rangle +
\sqrt{(j-m)(j+m+2\alpha+2)}\;|j,m+1\rangle,$$ if $j+m$ is even, and $$2\hat q |j,m\rangle = \sqrt{(j-m+1)(j+m+2\alpha+1)}\;|j,m-1\rangle +
\sqrt{(j+m+1)(j-m +2\alpha +1)}\;|j,m+1\rangle,$$ if $j+m$ is odd. The action of $2i\hat p$ is similar. In the (ordered) basis $\{ |j,-j\rangle, |j,-j+1\rangle, \ldots, |j,j-1\rangle, |j,j\rangle \}$ of $W_j$, the operators $2\hat q$ and $2i\hat p$ take the matrix form $$\begin{aligned}
2\hat q&=\left(
\begin{array}{ccccc}
0 & M_0& 0 & \cdots & 0 \\
M_0 & 0 & M_1 & \cdots & 0\\
0 & M_1 & 0 & \ddots & \\
\vdots & \vdots & \ddots & \ddots& M_{2j-1}\\
0 & 0 & & M_{2j-1} & 0
\end{array}
\right)\equiv M^q,
\label{Mq} \\
2i\hat p&=\left(
\begin{array}{ccccc}
0 & M_0& 0 & \cdots & 0 \\
-M_0 & 0 & M_1 & \cdots & 0\\
0 & -M_1 & 0 & \ddots & \\
\vdots & \vdots & \ddots & \ddots& M_{2j-1}\\
0 & 0 & & -M_{2j-1} & 0
\end{array}
\right)\equiv M^p,
\label{Mp}\end{aligned}$$ with $$M_k= \begin{cases}
\sqrt{(k+1)(2j+2\alpha -k+1)}, & \text{if $k$ is odd;}\\
\sqrt{(k+2\alpha+2)(2j-k)}, & \text{if $k$ is even.}
\end{cases}
\label{Ma}$$ For these matrices the eigenvalues and eigenvectors can be constructed in terms of Hahn polynomials. Hahn polynomials $Q_n(x;\alpha, \beta, N)$ [@Koekoek; @Ismail] of degree $n$ ($n=0,1,\ldots,N$) in the variable $x$, with parameters $\alpha>-1$ and $\beta>-1$ are defined by [@Koekoek; @Ismail]: $$Q_n(x;\alpha,\beta,N) = {\;}_3F_2 \left( {\genfrac{}{}{0pt}{}{-n,n+\alpha+\beta+1,-x}{\alpha+1,-N}} ; 1 \right),
\label{defQ}$$ in terms of the generalized hypergeometric series $_3F_2$ of unit argument [@Bailey; @Slater]. Their (discrete) orthogonality relation reads [@Koekoek; @Ismail]: $$\sum_{x=0}^N w(x;\alpha, \beta,N) Q_l(x;\alpha, \beta, N) Q_n(x;\alpha,\beta,N) = h(n;\alpha,\beta,N)\, \delta_{ln},
\label{orth-Q}$$ where $$\begin{aligned}
& w(x;\alpha, \beta,N) = \binom{\alpha+x}{x} \binom{N+\beta-x}{N-x} \qquad (x=0,1,\ldots,N); \\
& h(n;\alpha,\beta,N)= \frac{(n+\alpha+\beta+1)_{N+1}(\beta+1)_n n!}{(2n+\alpha+\beta+1)(\alpha+1)_n(N-n+1)_n N!}.\end{aligned}$$ We have used here the common notation for Pochhammer symbols [@Bailey; @Slater] $(a)_k=a(a+1)\cdots(a+k-1)$ for $k=1,2,\ldots$ and $(a)_0=1$. Orthonormal Hahn functions $\tilde Q$ are determined by: $$\tilde Q_n(x;\alpha,\beta,N) \equiv \frac{\sqrt{w(x;\alpha,\beta,N)}\, Q_n(x;\alpha,\beta,N)}{\sqrt{h(n;\alpha,\beta,N)}}.
\label{Q-tilde}$$ Recall that dual Hahn polynomials have a similar expression. In fact, for $x\in\{0,1,\ldots,N\}$ the right hand side of is the dual Hahn polynomial $R_x(\lambda(n);\alpha,\beta,N)$ of degree $x$ in the variable $\lambda(n)=n(n+\alpha+\beta+1)$; see [@Koekoek; @Ismail] for their orthogonality relations.
In terms of these, the eigenvalues and (orthonormal) eigenvectors of $M^q$ can be constructed:
Let $M^q\equiv 2\hat q$ be the tridiagonal $(2j+1)\times(2j+1)$-matrix and let $U=(U_{kl})_{0\leq k,l\leq 2j}$ be the $(2j+1)\times(2j+1)$-matrix with matrix elements: $$\begin{aligned}
& U_{2i,j} =(-1)^i \tilde Q_0(i;\alpha,\alpha,j), \;i\in\{0,1,\ldots,j\};\;
U_{2i+1,j} = 0, \; i\in\{0,\ldots,j-1\}; \label{Uj}\\
& U_{2i,j-k} = U_{2i,j+k} = \frac{(-1)^i}{\sqrt{2}} \tilde Q_k(i;\alpha,\alpha,j)
, \;i\in\{0,1,\ldots,j\}, \; k\in\{1,\ldots,j\}; \label{Ueven}\\
& U_{2i+1,j-k} = -U_{2i+1,j+k} = -\frac{(-1)^i}{\sqrt{2}} \tilde Q_{k-1}(i;\alpha+1,\alpha +1,j-1), \;
i\in\{0,1,\ldots,j-1\}, \nn \\
& \hskip 10.5 cm k \in\{1,\ldots,j\}\label{Uodd}.\end{aligned}$$ Then $U$ is an orthogonal matrix: $$U U^T = U^TU=I,$$ the columns of $U$ are the eigenvectors of $M^q$, i.e. $$M^q U = U D^q,
\label{MUUD}$$ where $D^q= \diag (\epsilon_0,\epsilon_1,\ldots,\epsilon_{2j})$ is a diagonal matrix containing the eigenvalues $\epsilon_k$ of $M^q$: $$\epsilon_{j-k}=-2\sqrt{k(2\alpha +k+1)}, \quad \epsilon_{j}=0, \quad \epsilon_{j+k}=2\sqrt{k(2\alpha+k+1)}, \label{epsilon}
\quad (k=1,\ldots,j).$$ \[propU\]
[**Proof.**]{} Using the orthogonality of the Hahn polynomials, and the explicit expressions -, a simple computation shows that $(U^TU)_{kl}=\delta_{kl}$. Thus $U^TU=I$, the identity matrix, and hence $UU^T=I$ holds as well.
Now it remains to verify and . By the form of $M^q$: $$\big(M^qU\big)_{ik}= \sum_{l=0}^{2j}M_{il}^q U_{lk}=M_{i-1}U_{i-1,k}+M_{i}U_{i+1,k}.
\label{MU}$$ We have to consider in six distinct cases, according to $i$ even or odd, and to $k$ belonging to $\{0,1,\ldots,j-1\}$, to $\{j+1,j+2,\ldots,2j\}$ or $k=j$. Let us consider the case that $i$ is odd and $k\in\{0,1,\ldots,j-1\}$. Then, relabelling the indices appropriately, , and yield: $$\begin{aligned}
&(M^qU)_{2i+1,j-k}=M_{2i}U_{2i,j-k}+M_{2i+1}U_{2i+2,j-k} \\
& =(-1)^i\sqrt{2}\sqrt{(\alpha+i+1)(j-i)}\tilde Q_k(i;\alpha,\alpha,j)+
(-1)^{i+1}\sqrt{2}\sqrt{(i+1)(j+\alpha-i)}\tilde Q_k(i+1;\alpha,\alpha,j)\\
&= (-1)^i \sqrt{2}\sqrt{\frac{(\alpha+1)_{i+1}(\alpha+1)_{j-i}}{i!(j-i-1)!h(k;\alpha,\alpha,j)}}[ Q_k(i;\alpha,\alpha,j)- Q_k(i+1;\alpha,\alpha,j)].\end{aligned}$$ At this point, the forward shift operator formula for Hahn polynomials [@Koekoek (9.5.6)] can be applied and yields $$\begin{aligned}
(M^qU)_{2i+1,j-k}& =(-1)^i \sqrt{2} \sqrt{\frac{(\alpha+1)_{i+1}(\alpha+1)_{j-i}}{i!(j-i-1)!h(k;\alpha,\alpha,j)}} \frac{k(k+2\alpha+1)}{(\alpha+1)j} Q_{k-1}(i;\alpha+1,\alpha+1,j-1)\\
&= -2\sqrt{k(k+2\alpha+1)} U_{2i+1,j-k} = \epsilon_{j-k}U_{2i+1,j-k}=\big(UD^q\big)_{2i+1,j-k}.\end{aligned}$$ For the other five cases, the computations are similar and for some of them the backward shift operator formula for Hahn polynomials [@Koekoek (9.5.8)] is applied.
Note that in the case of $\uu(2)_\alpha$ [@JSV2011], the equation corresponding to was related to two new difference equations for Hahn polynomials [@SV2011; @Gorin]. Here, the equation is just corresponding to known forward or backward shift operator formulas. The above proposition yields in particular the spectrum of the position operator:
The $2j+1$ eigenvalues $q$ of the position operator $\hat q$ in the representation $W_j$ are given by $$-\sqrt{j(2\alpha+j+1)}, -\sqrt{(j-1)(2\alpha+j)}, \ldots, -\sqrt{2\alpha+2}; 0 ;
\sqrt{2\alpha+2}, \ldots, \sqrt{j(2\alpha+j+1)}.$$ It will be appropriate to label these $\hat q$-eigenvalues as $q_k$, where $k=-j,-j+1,\ldots,+j$, so $$q_{\pm k} = \pm\sqrt{k(2\alpha+k+1)}, \qquad k=0, 1, \ldots, j.$$ \[prop4\]
One can compare the spectrum of $\hat q$ with that in related models, see Figure \[fig1\]. In the non-deformed case ($\alpha=-\frac12$, or the $\su(2)$ model), the spectrum is just equidistant as already mentioned in the introduction. In the case of $\uu(2)_\alpha$ (and $j$ half-integer), the spectrum is given by : apart from an extra gap of size $2\alpha+2$ in the middle, it is again equidistant. In the current case of $\su(2)_\alpha$ (and $j$ integer), the spectrum given by Proposition \[prop4\] is not equidistant.
Essentially the eigenvectors of have components proportional to Hahn polynomials with parameters $(\alpha,\alpha)$ when the component has even index and with parameters ($\alpha+1,\alpha+1)$ when the component has odd index. It is convenient to introduce a notation for these eigenvectors: the orthonormal eigenvector of the position operator $\hat q$ in $W_j$ for the eigenvalue $q_k$, denoted by $|j,q_k)$, is given in terms of the standard basis by $$|j,q_k) = \sum_{m=-j}^j U_{j+m,j+k} |j,m\rangle.$$
Let us now turn our attention to the momentum operator $\hat p$. Up to signs, the matrix $M^p$ is the same as the matrix $M^q$. The analysis of its eigenvalues and eigenvectors is thus very similar. We present just the final result here:
Let $M^p\equiv 2i\hat p$ be the tridiagonal $(2j+1)\times(2j+1)$-matrix and let $V=(V_{kl})_{0\leq k,l\leq 2j}$ be the $(2j+1)\times(2j+1)$-matrix with matrix elements $$V_{k,l}= - i^{k+1} U_{kl},
\label{V}$$ where $U$ is the matrix determined by -. Then $V$ is a unitary matrix, $V V^\dagger = V^\dagger V=I$. The columns of $V$ are the eigenvectors of $M^p$, i.e. $$M^p V = V D^p,
\label{MVVD}$$ where $D^p$ is a diagonal matrix containing the eigenvalues $\varepsilon_k$ of $M^p$: $$\begin{aligned}
& D^p= \diag (\varepsilon_0,\varepsilon_1,\ldots,\varepsilon_{2j}), \nn\\[1mm]
& \varepsilon_{j-k}=-2i\sqrt{k(2\alpha +k+1)}, \quad \epsilon_{j}=0, \quad \epsilon_{j+k}=2i\sqrt{k(2\alpha+k+1)}
\quad (k=1,\ldots,j). \label{varepsilon}\end{aligned}$$
Hence the $2j+1$ eigenvalues $p$ of the momentum operator $\hat p$ in the representation $W_j$ are given by $$-\sqrt{j(2\alpha+j+1)}, -\sqrt{(j-1)(2\alpha+j)}, \ldots, -\sqrt{2\alpha+2}; 0 ;
\sqrt{2\alpha+2}, \ldots, \sqrt{j(2\alpha+j+1)},$$ in other words, the momentum operator $\hat p$ has the same spectrum as the position operator $\hat q$. We shall denote these $\hat p$-eigenvalues by $p_k$, where $k=-j,-j+1,\ldots,+j$, so $$p_{\pm k} = \pm\sqrt{k(2\alpha+k+1)}, \qquad k=0, 1, \ldots, j.$$ The normalized eigenvector of the momentum operator $\hat p$ in $W_j$ for the eigenvalue $p_k$, denoted by $|j,p_k)$, is then given by $$|j,p_k) = \sum_{m=-j}^j V_{j+m,j+k} |j,m\rangle.$$
Finally, it is worth mentioning another property of the matrix elements of $V$, that follows from the explicit expressions , - and the orthogonality properties of the Hahn polynomials: $$V^T V = \left( \begin{array}{cccc}
0 & \cdots & 0 & -1 \\ 0 & \cdots & -1 & 0 \\ \vdots & \qdots & \vdots &\vdots\\ -1 & \cdots & 0 & 0 \\ \end{array}\right).
\label{VTV}$$ Note also that from it follows that $$V={\cal J}U \qquad\hbox{where}\qquad
{\cal J}=\diag(-i,1,i,-1,\ldots),
\label{VJU}$$ and the sequence $(-i,1,i,-1)$ is repeated, ending with $i$ or $-i$ (since the matrices are odd-dimensional). One can also write ${\cal J}= -i \diag(i^0,i^1,i^2,i^2,\ldots,i^{2j})$.
Position and momentum wavefunctions and their properties
========================================================
The position (resp. momentum) wavefunctions of the $\su(2)_\alpha$ finite oscillator are the overlaps between the $\hat q$-eigenvectors (resp. $\hat p$-eigenvectors) and the $\hat H$-eigenvectors (or equivalently, the $J_0$-eigenvectors $|j,m\rangle$). Let us denote them by $\Phi^{(\alpha)}_{j+m}(q)$ (resp. $\Psi^{(\alpha)}_{j+m}(p)$ ), where $m=-j,-j+1,\ldots,+j$, and where $q$ (resp. $p$) assumes one of the discrete values $q_k$ (resp. $p_k$) $(k=-j,-j+1,\ldots,+j)$. Therefore, in the notation of the previous section: $$\begin{aligned}
& \Phi^{(\alpha)}_{j+m}(q_k)= \langle j,m | j,q_k ) = U_{j+m,j+k}, \label{UPhi}\\
& \Psi^{(\alpha)}_{j+m}(p_k)= \langle j,m | j,p_k ) = V_{j+m,j+k}. \label{VPsi}\end{aligned}$$ Let us consider the explicit form of these wavefunctions, first for the position variable. For $j+m$ even, $j+m=2n$, and for positive $q$-values one has $$\Phi^{(\alpha)}_{2n} (q_{k}) = \frac{(-1)^n}{\sqrt{2}} \tilde Q_{k}(n;\alpha,\alpha,j), \qquad n=0, 1, \ldots, j,
\qquad k=1, \ldots, j,$$ or equivalently: $$\begin{aligned}
\Phi^{(\alpha)}_{2n} (q_{k}) &= \frac{(-1)^n}{\sqrt{2}}
\sqrt{\frac{w(n;\alpha,\alpha,j)}{h(k;\alpha,\alpha,j)}}
{\ }_3F_2 \left( {\genfrac{}{}{0pt}{}{-k,k+2\alpha+1,-n}{\alpha+1,-j}} ; 1 \right)\nn \\
& =\frac{(-1)^n}{\sqrt{2}}
\sqrt{\frac{w(n;\alpha,\alpha,j)}{h(k;\alpha,\alpha,j)}}
R_n \left( q_k^2; \alpha, \alpha, j\right), \qquad q_{k}^2 = k(2\alpha+k+1),
\label{Phi-even}\end{aligned}$$ where $R_n \left( \lambda_k; \alpha, \alpha, j\right)$ is a dual Hahn polynomial [@Koekoek; @Ismail] of degree $n$ in the variable $\lambda(k)\equiv q_k^2=k(2\alpha+k+1)$. In a similar way, one finds for $j+m$ odd, $j+m=2n+1$, and for positive $q$-values, that $$\begin{aligned}
\Phi^{(\alpha)}_{2n+1} (q_{k}) & = \frac{(-1)^n}{\sqrt{2}}
\sqrt{\frac{w(n;\alpha+1,\alpha+1,j-1)}{h(k-1;\alpha+1,\alpha+1,j-1)}}
{\ }_3F_2 \left( {\genfrac{}{}{0pt}{}{-k+1,k+2\alpha+2,-n}{\alpha+2,-j+1}} ; 1 \right)\nn \\
&=\sqrt{\frac{w(n;\alpha+1,\alpha+1,j-1)}{h(k-1;\alpha+1,\alpha+1,j-1)}}
R_n \left( q_k^2-2(\alpha+1);\alpha+1,\alpha+1,j-1 \right).
\label{Phi-odd}\end{aligned}$$ For $q=0$ or negative $q$-values, the expressions are of course analogous, according to -.
Before studying some special properties of these position wavefunctions, and determining the momentum wavefunctions, let us consider plots of these functions for some $\alpha$-values. We choose a fixed value of $j$, say $j=30$, and plot some of the wavefunctions $\Phi^{(\alpha)}_n(q)$ for certain values of $\alpha$. Since $\alpha=-\frac12$ is a special case (where $\su(2)_\alpha$ reduces to $\su(2)$), there are three cases to be considered: $-1<\alpha<-\frac12$, $\alpha=-\frac12$ and $\alpha>-\frac12$. In Figure \[fig2\] we take $\alpha=-\frac12$, $\alpha=-0.7$ and $\alpha=2$ respectively. We also plot in each case the ground state $\Phi^{(\alpha)}_0(q)$, some low energy states $\Phi^{(\alpha)}_1(q)$ and $\Phi^{(\alpha)}_2(q)$, and the highest energy state. The plots are similar as in the $\uu(2)_\alpha$ case (where $j$ is half-integer). The most obvious difference is that 0 is part of the spectrum now. Another difference, but more difficult to see in the plots, is that the spectrum is not equidistant for $\alpha\ne-\frac12$. For $\alpha=-\frac12$, these plots coincide with the ones given in the $\su(2)$ model [@Atak2005; @Atak2001]. The wavefunctions $\Phi^{(-1/2)}_n(q)$ are Krawtchouk functions. This follows also from the following transformations of ${}_3F_2$ series to ${}_2F_1$ series when $\alpha=-\frac12$, according to [@Atak2005 (48)] $$\begin{aligned}
& {\;}_3F_2 \left( {\genfrac{}{}{0pt}{}{-k,k,-n}{1/2,-j}} ; 1 \right)= (-1)^n \frac{\binom{2j}{2n}}{\binom{j}{n}}
{\;}_2F_1 \left( {\genfrac{}{}{0pt}{}{-2n,-j-k}{-2j}} ; 2 \right),\\
& {\;}_3F_2 \left( {\genfrac{}{}{0pt}{}{-k+1,k+1,-n}{3/2,-j+1}} ; 1 \right)= -\frac{(-1)^n}{2k} \frac{\binom{2j}{2n+1}}{\binom{j-1}{n}}
{\;}_2F_1 \left( {\genfrac{}{}{0pt}{}{-2n-1,-j-k}{-2j}} ; 2 \right). \end{aligned}$$
For $\alpha\ne-\frac12$, the plots are comparable with the parabose wavefunctions [@JSV2011]. One can indeed again study the behaviour of the discrete wavefunctions $\Phi^{(\alpha)}_n(q)$ when the representation parameter $j$ tends to infinity. In this process, one should pass from a discrete position variable $q$ to a continuous variable $x$. This can be done by putting $q=j^{1/2} x$ and then compute the limit of $j^{1/4} \Phi^{(\alpha)}_n(q)$ for $j\rightarrow \infty$. The actual computation is similar to the limit computation performed in [@JSV2011 §4], so we shall not give any details. Note that, due to , $$q^2 = (k+\alpha+\frac12)^2-(\alpha+\frac12)^2,\hbox{ or }
k=-\alpha-\frac12 \pm \sqrt{q^2+(\alpha+\frac12)^2}.
\label{kq}$$ Using this last expression for $k$ in the ${}_3F_2$ expression of , and replacing herein $q$ by $j^{1/2} x$, the limit can be computed and yields: $$\lim_{j\rightarrow\infty} j^{1/4} \Phi^{(\alpha)}_{2n}(j^{1/2} x ) =
(-1)^n \sqrt{\frac{n!}{\Gamma(\alpha+n+1)}}\; |x|^{\alpha+1/2} e^{-x^2/2} L_n^{(\alpha)}(x^2),
\label{psi-even}$$ in terms of Laguerre polynomials. Similarly, one finds from : $$\lim_{j\rightarrow\infty} j^{1/4} \Phi^{(\alpha)}_{2n+1}(j^{1/2} x ) =
(-1)^n \sqrt{\frac{n!}{\Gamma(\alpha+n+2)}}\; x |x|^{\alpha+1/2} e^{-x^2/2} L_n^{(\alpha+1)}(x^2).
\label{psi-odd}$$ The functions in the right hand side of - are known: they are the wavefunctions of the parabose oscillator [@Mukunda; @Ohnuki; @JSV2008] with parameter $a=\alpha+1>0$, see the appendix of [@JSV2011].
It remains here to consider the momentum wavefunctions. Due to the fact that the spectrum of $\hat q$ and $\hat p$ is the same, and due to the similarity of the matrix of eigenvectors $V$ (compared to $U$), the expressions are analogous and we give only the final result here: $$\begin{aligned}
\Psi^{(\alpha)}_{2n} (p_{k}) & = -\frac{i}{\sqrt{2}}
\sqrt{\frac{w(n;\alpha,\alpha,j)}{h(k;\alpha,\alpha,j)}}
R_n \left( p_k^2; \alpha, \alpha, j\right), \qquad p_{k}^2 = k(2\alpha+k+1),
\label{Psi-even} \\
\Psi^{(\alpha)}_{2n+1} (p_{k}) & =
\sqrt{\frac{w(n;\alpha+1,\alpha+1,j-1)}{h(k-1;\alpha+1,\alpha+1,j-1)}}
R_n \left( p_k^2-2(\alpha+1);\alpha+1,\alpha+1,j-1 \right).
\label{Psi-odd}\end{aligned}$$
The discrete Hahn-Fourier transform
===================================
In canonical quantum mechanics, the momentum wavefunction (in $L^2(\R)$) is given by the Fourier transform of the position wavefunction (and vice versa): $$\Psi(p)= \frac{1}{\sqrt{2\pi}} \int e^{-ipq}\Phi(q)dq.$$ In the current case, we are dealing with discrete wavefunctions, and we should look for an analogue of this. In terms of the notation of the previous section, let $$\Phi (q_k)=\left(
\begin{array}{c}
\Phi_0(q_k) \\
\Phi_1 (q_k) \\
\vdots \\
\Phi_{2j}(q_k)
\end{array}
\right), \qquad
\Psi (p_k)=\left(
\begin{array}{c}
\Psi_0(p_k) \\
\Psi_1 (p_k) \\
\vdots \\
\Psi_{2j}(p_k)
\end{array}
\right)\qquad (k=-j,\ldots,+j).
\label{PhiV-PsiV}$$ So it is natural to define the discrete Fourier transform in this case as the matrix $F=(F_{lk})_{-j\leq l,k\leq +j}$ relating these two wavefunctions. In other words: $$\Psi(p_l)=\sum_{k=-j}^j F_{kl}\;\Phi(q_k).
\label{F}$$ As this generalized discrete Fourier transform maps Hahn polynomials into Hahn polynomials, we shall refer to it as the discrete Hahn-Fourier transform. By -, the columns of $V$ consist of the column vectors $\Psi(p_k)$ ($k=-j..,+j$) and similarly for the matrix $U$. So actually means that $V = U F$, or: $$F=U^T V.
\label{FVU}$$ Using the explicit matrix elements from $U$ and $V$, - and , this leads to following form of the matrix elements of $F$: $$\begin{aligned}
&F_{j-k,j\mp l}=F_{j+k,j\pm l}=-\frac{i}{2}\sum_{n=0}^j(-1)^n \tilde Q_k(n;\alpha,\alpha,j) \tilde Q_l(n;\alpha,\alpha,j)
\label{F1}\\
&\pm \frac12 \sum_{n=0}^{j-1}(-1)^n \tilde Q_{k-1}(n;\alpha+1,\alpha+1,j-1) \tilde Q_{l-1}(n;\alpha+1,\alpha+1,j-1),
\quad k,l=1,\ldots,j;\nn\\
&F_{j\mp k,j}=F_{j,j\mp k}=-\frac{i}{\sqrt{2}}\sum_{n=0}^j(-1)^n \tilde Q_k(n;\alpha,\alpha,j) \tilde Q_0(n;\alpha,\alpha,j), \quad k=1,\ldots,j;\label{F2}\\
& F_{jj}=-i\sum_{n=0}^j(-1)^n \tilde Q_0^2(n;\alpha,\alpha,j). \label{F3}\end{aligned}$$
In the following section, we shall determine explicit expressions for the above matrix elements. But before that, we can already summarize some properties of the discrete Hahn-Fourier transform matrix $F$.
The $(2j+1)\times(2j+1)$-matrix $F$ is symmetric, $F^T=F$, and unitary, $F^\dagger F=F F^\dagger =I$. Furthermore, it satisfies $F^4=I$, so its eigenvalues are $\pm 1, \pm i$. A set of orthonormal eigenvectors of $F$ is given by the rows of $U$, determined in Proposition \[propU\]. The multiplicity of the eigenvalues depends on the parity of $j$. When $j=2n$ is even, then the multiplicity of $-i,1,i,-1$ is $n+1,n,n,n$ respectively. When $j=2n+1$ is odd, then the multiplicity of $-i,1,i,-1$ is $n+1,n+1,n+1,n$ respectively.
[**Proof.**]{} The symmetry of $F$ is easily seen from the expressions -. The unitarity of $F$ follows from , the orthogonality of the real matrix $U$ and the unitarity of $V$. Again using and the orthogonality of $U$, one finds $F^2=F^TF= V^TUU^TV=V^TV$. But the explicit form of $V^TV$ is known, see . Since $(V^TV)^2=I$, the result $F^4=I$ follows. So the eigenvalues can only be $\pm 1, \pm i$. Using again and , one sees that $F=U^TV=U^T{\cal J}U$, or $$F U^T = U^T {\cal J}.$$ In other words, the columns of $U^T$ (or the rows of $U$) form a set of orthonormal eigenvectors of $F$, and the eigenvalues of $F$ are found in the diagonal matrix ${\cal J}$. From the explicit form of ${\cal J}$, the statement of the multiplicities follows.
Note that these properties are similar to those of the common discrete Fourier transform [@McClennan; @Atak1997]. For the case $\alpha=-\frac12$, the matrix $F$ was already studied in [@Atak1997]. In that special case, the position and momentum wavefunctions are in terms of Krawtchouk polynomials, and the corresponding discrete Fourier transform can be referred to as the discrete Krawtchouk-Fourier transform. The matrix elements of $F$ are in that special case again Krawtchouk functions [@Atak1997].
One of our main results is the explicit computation of the elements of $F$ for general $\alpha$. Note from - that all these matrix elements are of the form $\sum_{n=0}^j(-1)^n \tilde Q_k(n;\alpha,\alpha,j) \tilde Q_l(n;\alpha,\alpha,j)$, with $0\leq k,l \leq j$. So apart from a factor $(h(k;\alpha,\alpha,j)h(l;\alpha,\alpha,j))^{-1/2}$, this expression is equal to: $$\begin{aligned}
S(k,l,\alpha,j) &= \sum_{n=0}^j (-1)^n w(n;\alpha,\alpha,j) Q_k(n;\alpha,\alpha,j)Q_l(n;\alpha,\alpha,j) \label{S1}\\
&= \sum_{n=0}^j (-1)^n \binom{\alpha+n}{n}\binom{\alpha+j-n}{j-n} Q_k(n;\alpha,\alpha,j)Q_l(n;\alpha,\alpha,j) \label{S2}\\
&= \sum_{n=0}^j (-1)^n \binom{\alpha+n}{n}\binom{\alpha+j-n}{j-n} R_n(\lambda(k);\alpha,\alpha,j)R_n(\lambda(l);\alpha,\alpha,j).
\label{S3}\end{aligned}$$ In the form , the right hand side is just like the orthogonality relation of Hahn polynomials, except for the extra factor $(-1)^n$. In the form , one can see that $S(k,l,\alpha,j)$ is a special case of the Poisson kernel (or a bilinear generating function) for dual Hahn polynomials: $$\sum_{n=0}^j t^n \binom{\alpha+n}{n}\binom{\beta+j-n}{j-n} R_n(\lambda(k);\alpha,\beta,j)R_n(\lambda(l);\alpha,\beta,j).$$ However, as far as we know a closed form expression for this Poisson kernel is not known. In fact, the best one can do is re-express the product $R_n(\lambda(k);\alpha,\beta,j)R_n(\lambda(l);\alpha,\beta,j)$ into a ${}_8F_7$ hypergeometric series (following the approach of [@Gasper §8.3], where the $q$-analogue is given). Here we will show that this Poisson kernel does have a simple expression when $\beta=\alpha$ and $t=-1$.
We shall prove the following result, yielding the explicit expression for the discrete Hahn-Fourier transform matrix $F$:
The special Poisson kernel for dual Hahn polynomials $S(k,l,\alpha,j)$, where $k$ and $l$ are integers with $0\leq k,l \leq j$, satisfies $$S(k,l,\alpha,j)=(-1)^{k+l+j} S(k,l,\alpha,j),
\label{S0}$$ so it is 0 whenever $k+l+j$ is odd. For the other cases, it is given by: $$\begin{aligned}
S(2K,2L,\alpha,2J)=& \frac{2^{2J}(\frac12)_{J-K}(\frac12)_{J-L}(\alpha+1)_J(\alpha+J+1)_K(\alpha+J+1)_L}{(2J)!(\frac12)_J} \nn\\
& \times {\ }_4F_3 \left( {\genfrac{}{}{0pt}{}{-K,K+\alpha +\frac12,-L,L+\alpha+\frac12}{\alpha+J+1,\alpha +1,-J}} ; 1 \right)\label{KL1} \\
S(2K+1,2L+1,\alpha,2J)=
&\frac{2^{2J}(\frac12)_{J-K}(\frac12)_{J-L}(\alpha+1)_{J+1}(\alpha+J+2)_K(\alpha+J+2)_L}{(2J)!J(\frac12)_J} \nn\\
&\times {\ }_4F_3 \left( {\genfrac{}{}{0pt}{}{-K,K+\alpha +\frac32,-L,L+\alpha+\frac32}{\alpha+J+2,\alpha +1,-J+1}} ; 1 \right)\label{KL2} \\
S(2K,2L+1,\alpha,2J+1)=
&\frac{2^{2J+1}(\frac12)_{J-K+1}(\frac12)_{J-L}(\alpha+1)_{J+1}(\alpha+J+2)_K(\alpha+J+2)_L}{(2J+1)!(\frac12)_{J+1}}\nn\\
&\times {\ }_4F_3 \left( {\genfrac{}{}{0pt}{}{-K,K+\alpha +\frac12,-L,L+\alpha+\frac32}{\alpha+J+2,\alpha +1,-J}} ; 1 \right)\label{KL3}\end{aligned}$$ and finally $S(2K+1,2L,\alpha,2J+1)$ is given by replacing $K$ and $L$ in the right hand side of . \[maintheo\]
First of all, note that $S(k,l,\alpha,j)=0$ for $k+l+j$ odd implies that in expression only one of the two parts survive (either the real part or else the imaginary part). Together with Theorem \[maintheo\] this implies that each element $F_{lk}$ is, up to a factor, equal to a terminating Saalsch" utzian ${}_4F_3(1)$ series. In other words, up to a factor each $F_{lk}$ is also a Racah polynomial [@Koekoek]. The unitarity of the matrix $F$ does not lead to any new relations for Racah polynomials, but just follows from their discrete orthogonality relations.
Before discussing the proof of Theorem \[maintheo\], let us examine what happens to this discrete Hahn-Fourier transform in the limit when $j\rightarrow\infty$. Suppose that $j=2J$ is even, and let us first consider the limit of the imaginary matrix elements of $F$. When $k$ and $l$ are even ($j=2J$, $k=2K$ and $l=2L$), then according to : $$F_{j+k,j\pm l} = -\frac{i}{2} (h(k;\alpha,\alpha,j)h(l;\alpha,\alpha,j))^{-1/2} S(2K,2L,\alpha,2J),$$ with $S(2K,2L,\alpha,2J)$ given by . Just as for the limit computation of the wavefunctions, see , it is necessary to make the replacements $$k=\sqrt{q_k^2+(\alpha+\frac12)^2}-(\alpha+\frac12),\qquad l=\sqrt{p_l^2+(\alpha+\frac12)^2}-(\alpha+\frac12),$$ and put $q_k=j^{1/2}x$ and $p_l=j^{1/2}p$ in the above expression. After doing this, the limit of the ${}_4F_3$ series appearing in $S(2K,2L,\alpha,2J)$ is fairly easy to determine by termwise computation: $$\lim_{j\rightarrow \infty} {\ }_4F_3 \left( {\genfrac{}{}{0pt}{}{-\frac{k}{2},\frac{k}{2}+\alpha +\frac12,-\frac{l}{2},
\frac{l}{2}+\alpha+\frac12}{\alpha+\frac{j}{2}+1,\alpha +1,-\frac{j}{2} }} ; 1 \right) =
{\;}_0F_1\left( {\genfrac{}{}{0pt}{}{ - }{\alpha+1}}; -\frac{1}{4}x^2p^2\right).$$ The last expression is proportional to $J_\alpha(xp)$, where $J_\alpha$ is the Bessel function of the first kind [@Temme]. The limit of the factors in front of the ${}_4F_3$ series are elementary but a bit more tedious to compute, and we shall not give the details of this computation. Similarly, one has to determine the limit of the real matrix elements of $F$. Adding both contributions, one finds: $$\lim_{j\rightarrow \infty} j^{1/2} F_{j+k,j\pm l} = \frac12 \left(
|xp|^{1/2} J_\alpha(|xp|) + ixp |xp|^{-1/2} J_{\alpha+1}(|xp|)
\right).$$ The function in the right hand side is known: it is the kernel of the generalized Fourier transform, studied by Mukunda [*et al*]{} in the context of the parabose oscillator [@Mukunda]. So one can conclude that our discrete Hahn-Fourier transform tends to the generalized Fourier transform of [@Mukunda] for large values of $j$.
Computation of the discrete Hahn-Fourier transform matrix elements
==================================================================
The purpose of this section is the proof of Theorem \[maintheo\]. In fact, we shall show and ; the proof of and is analogous.
Let us first collect some known transformation formulas for hypergeometric series. The first is Thomae’s transformation of a terminating Saalschützian ${}_4F_3(1)$ series [@Slater (2.4.1.7)], [@Bailey (7.2.1)]: $${\;}_4F_3 \left( {\genfrac{}{}{0pt}{}{a,b,c,-N}{e,f,g}} ; 1 \right)=
\frac{(f-c)_N(e+f-a-b)_N}{(f)_N(e+f-a-b-c)_N} {\;}_4F_3 \left( {\genfrac{}{}{0pt}{}{e-a,e-b,c,-N}{e,e+f-a-b,e+g-a-b}} ; 1 \right)
\label{Thomae}$$ where $e+f+g=1+a+b+c-N$. For the Hahn polynomials appearing in , the following transformation formula holds [@Gasper1974 p. 186]: $$Q_k(x;\alpha,\alpha,N)= (-1)^k Q_k(N-x;\alpha,\alpha,N).
\label{Q-symm}$$ In the same paper [@Gasper1974 (3.18)-(3.19)], one can find the following ${}_4F_3$ expressions for Hahn polynomials with parameter $\beta=\alpha$: $$\begin{aligned}
Q_{2k}(x;\alpha,\alpha,N)&=
{\;}_4F_3 \left( {\genfrac{}{}{0pt}{}{-k,k+\alpha+\frac12,-x,x-N}{\alpha+1,-N/2,(1-N)/2}} ; 1 \right), \label{Qeven}\\
Q_{2k+1}(x;\alpha,\alpha,N)&= \frac{N-2x}{N}
{\;}_4F_3 \left( {\genfrac{}{}{0pt}{}{-k,k+\alpha+\frac32,-x,x-N}{\alpha+1,(1-N)/2,(2-N)/2}} ; 1 \right). \label{Qodd}\end{aligned}$$
Next, we need an equation that goes back to a classical identity for $6j$-coefficients. For $6j$-coefficients of $\su(2)$, the following holds [@Varshalovich (9.8.4)]: $$\sum_x (-1)^{p+q+x}(2x+1)
{\left\{ {\genfrac{}{}{0pt}{}{a}{c}} {\genfrac{}{}{0pt}{}{b}{d}} {\genfrac{}{}{0pt}{}{x}{p}}\right\} } {\left\{ {\genfrac{}{}{0pt}{}{a}{d}} {\genfrac{}{}{0pt}{}{b}{c}} {\genfrac{}{}{0pt}{}{x}{q}}\right\} } = {\left\{ {\genfrac{}{}{0pt}{}{a}{b}} {\genfrac{}{}{0pt}{}{c}{d}} {\genfrac{}{}{0pt}{}{q}{p}}\right\} }.
\label{su2-6j}$$ It turns out that we need the corresponding identity for $6j$-coefficients of positive discrete series representations of $\su(1,1)$, which reads: $$\sum_{j_{23}} (-1)^{j_{13}+j_{23}-j'}\; U^{k_1,k_2,k_{12}}_{k_3,k,k_{23}} U^{k_1,k_3,k_{13}}_{k_2,k,k_{23}} =
U^{k_2,k_1,k_{12}}_{k_3,k,k_{13}}.
\label{su11-6j}$$ Herein, $U^{k_1,k_2,k_{12}}_{k_3,k,k_{23}}$ is the standard notation for the Racah coefficient of $\su(1,1)$ [@LNM]. In , $k_1$, $k_2$ and $k_3$ are $\su(1,1)$ representation labels (i.e. they are positive real numbers), and in a common notation [@LNM] $$k_{12}=k_1+k_2+j_{12}, \quad k=k_{12}+k+3+j', \quad k_{13}=k_1+k_3+j_{13},\quad k_{23}=k_2+k_3+j_{23},$$ with $j_{12}$, $j'$, $j_{13}$ and $j_{23}$ all nonnegative integers. The summation in the right hand side of runs over all $j_{23}$ with $0\leq j_{23} \leq j'+j_{12}$. Using the explicit expression in terms of a ${}_4F_3(1)$ series for the Racah coefficients in , given by [@LNM (4.41)], and making the following replacements: $$(j',j_{12},j_{13}) \longrightarrow (p,q,r),\qquad
(\frac{k_1}{2}, \frac{k_2}{2},\frac{k_3}{2}) \longrightarrow (a,b,c), \qquad
j_{23} \longrightarrow n,$$ one arrives at the following identity between terminating Saalschützian ${}_4F_3(1)$ series: $$\begin{aligned}
& \sum_{n=0}^{p+q} \binom{p+q}{n} \frac{(b+c+2n-1)}{(b+c+n-1)}
\frac{(b+c)_n(a+b+c+p+q-1)_n}{(b+c+p+q)_n(1-a-p-q)_n} \nn \\
& \times {\;}_4F_3 \left( {\genfrac{}{}{0pt}{}{-n,n+b+c-1,-q,q+a+b-1}{b,a+b+c+p+q-1,-p-q}} ; 1 \right)
{\;}_4F_3 \left( {\genfrac{}{}{0pt}{}{-n,n+b+c-1,-r,r+a+c-1}{c,a+b+c+p+q-1,-p-q}} ; 1 \right) \nn\\
&= (-1)^{p-r} \frac{(b+c)_{p+q}(a)_r(a)_q}{(a)_{p+q}(b)_q(c)_r}
{\;}_4F_3 \left( {\genfrac{}{}{0pt}{}{-q,q+a+b-1,-r,r+a+c-1}{a,a+b+c+p+q-1,-p-q}} ; 1 \right).
\label{id1}\end{aligned}$$
This is not yet in the form needed for our proof of Theorem \[maintheo\]. But after performing Thomae’s transformation on the 2nd and 3rd ${}_4F_3(1)$ series appearing in , and replacing $r$ by $p+q-r$, we have the following:
Let $p$, $q$ and $r$ be nonnegative integers, and $a$, $b$ and $c$ arbitrary parameters, then $$\begin{aligned}
& \sum_{n=0}^{p+q} \binom{p+q}{n} \frac{(b+c+2n-1)}{(b+c+n-1)}
\frac{(b+c)_n(b)_n}{(c)_n(b+c+p+q)_n} \nn \\
& \times {\;}_4F_3 \left( {\genfrac{}{}{0pt}{}{-n,n+b+c-1,-q,q+a+b-1}{b,a+b+c+p+q-1,-p-q}} ; 1 \right)\nn\\
&\times {\;}_4F_3 \left( {\genfrac{}{}{0pt}{}{-n,n+b+c-1,-r,r+1-a-c-2p-2q}{b,1-a-p-q,-p-q}} ; 1 \right) \nn\\
= &(-1)^{r} \frac{(b+c)_{p+q}(1-c-p-q)_r}{(c)_p(a+b+c+p+q-1)_{q}(1-a-p-q)_r} \nn\\
&\times {\;}_4F_3 \left( {\genfrac{}{}{0pt}{}{-q,q+a+b-1,-r,r+1-a-c-2p-2q}{b,1-c-p-q,-p-q}} ; 1 \right).
\label{id2}\end{aligned}$$
Note that the identities - show at first sight some similarity with the expansion formulas constructed in [@Lievens] (or some $q$-analogues in [@Gasper2000]); they turn out to be quite different however.
We now come to the final part of this section:
[**Proof of Theorem \[maintheo\].**]{} First of all, from , note the symmetry $S(l,k,\alpha,j)=S(k,l,\alpha,j)$. Starting from , using for $Q_k$ and $Q_l$, and then reversing the order of summation (i.e. replace $n$ by $n-j$), one finds . This implies that $S(k,l,\alpha,j)=0$ whenever $k+l+j$ is odd. So we need to simplify the expression only when $k+l+j$ is even. We shall do this explicitly in the case that $k$, $l$ and $j$ are even: $k=2K$, $l=2L$ and $j=2J$, with $K$, $L$ and $J$ nonnegative integers (the other three cases are similar). So we need to compute $$S(2K,2L,\alpha,2J)
= \sum_{n=0}^{2J} (-1)^n \binom{\alpha+n}{n}\binom{\alpha+2J-n}{2J-n} Q_{2K}(n;\alpha,\alpha,2J)Q_{2L}(n;\alpha,\alpha,2J).$$ Using , this can be written as $S(2K,2L,\alpha,2J) = \sum_{n=0}^{2J} T_n(K,L)$, where $$\begin{aligned}
T_n(K,L)& = (-1)^n \binom{\alpha+n}{n}\binom{\alpha+2J-n}{2J-n}
{\;}_4F_3 \left( {\genfrac{}{}{0pt}{}{-K,K+\alpha+\frac12,-n,n-2J}{\alpha+1,-J,-J+\frac12}} ; 1 \right)\nn\\
& \qquad\times{\;}_4F_3 \left( {\genfrac{}{}{0pt}{}{-L,L+\alpha+\frac12,-n,n-2J}{\alpha+1,-J,-J+\frac12}} ; 1 \right).
\label{TnKL}\end{aligned}$$ Making an appropriate replacement of the summation variable, it is easy to see that $$\sum_{n=J}^{2J} T_n(K,L) = \sum_{n=0}^J T_n(K,L),$$ hence we can split the total sum $\sum_{n=0}^{2J} T_n(K,L)$ in two equal parts; taking care of the overlapping middle term, there comes $$S(2K,2L,\alpha,2J) = 2 \left(\sum_{n=0}^{J-1} T_n(K,L) +\frac12 T_J(K,L) \right).
\label{S2S}$$ In order to perform the summation in the right hand side of , one can make the following substitution in : $$p=J-K,\ q=K,\ r=L,\ a=\frac12,\ b=\alpha+t,\ c=-2J-\alpha, \label{subs}$$ and then take the limit $t\rightarrow 1$ (in fact, for all terms in one can immediately take $t=1$, except for the term with $n=p+q=J$ the limit process is actually necessary). Consider, after the substitution , the $n$th term in the left hand side of , with $0\leq n<p+q=J$: the coefficient in front of the two ${}_4F_3(1)$’s becomes: $$\frac{(2J)!}{(\alpha+1)_{2J}} (-1)^n \binom{\alpha+n}{n}\binom{\alpha+2J-n}{2J-n}.$$ Consider similarly the term with $n=p+q=J$; here the coefficient becomes, when $t\rightarrow 1$: $$\frac12 \frac{(2J)!}{(\alpha+1)_{2J}} (-1)^J \binom{\alpha+J}{J}\binom{\alpha+J}{J}.$$ Furthermore, it is easy to see that under this substitution and $t\rightarrow 1$ the two ${}_4F_3(1)$’s in the left hand side of become the ${}_4F_3(1)$ expressions of . Thus, up to the coefficient $\frac{(2J)!}{(\alpha+1)_{2J}}$, the right hand side of can be summed according to , and one finds, after simplifications: $$\begin{aligned}
S(2K,2L,\alpha,2J)& = \frac{2^{2J}(\frac12)_{J-K}(\frac12)_{J-L}(\alpha+1)_J(\alpha+J+1)_K(\alpha+J+1)_L}{(2J)!(\frac12)_J} \\
& \times {\ }_4F_3 \left( {\genfrac{}{}{0pt}{}{-K,K+\alpha +\frac12,-L,L+\alpha+\frac12}{\alpha+J+1,\alpha +1,-J}} ; 1 \right),\end{aligned}$$ proving .
Summary
=======
The most popular finite oscillator model, especially for applications in quantum optics, is based on the Lie algebra $\su(2)$ or $\so(3)$ [@Atak2005; @Atak2001; @Atak2001b]. The dimension of the model depends on the $\su(2)$ representation label $j$, which is an integer or half-integer positive number. Its mathematical properties have been studied well, in particular the properties of the wavefunctions given by Krawtchouk functions. These wavefunctions have interesting plots, and in the limit $j\rightarrow \infty$ these discrete wavefunctions tend to the continuous canonical oscillator wavefunctions. The discrete Fourier transform turning position wavefunctions into momentum wavefunctions has also been investigated in this case [@Atak1997].
A first type of deformation of this $\su(2)$ model was offered by its $q$-deformation. The $su_q(2)$ model for the finite oscillator was investigated in [@Ballesteros; @AKW]. The main properties are: the position operator has a discrete anharmonic spectrum, and the wavefunctions are given in terms of dual $q$-Krawtchouk polynomials [@AKW].
We have now considered two different type of deformations of the $\su(2)$ model, closely related to each other. The first deformation $\uu(2)_\alpha$ was given in [@JSV2011] and allows a deformation of the even-dimensional representations only ($j$ half-integer). The second deformation $\su(2)_\alpha$ was the topic of this paper, and allows a deformation of the odd-dimensional representations only ($j$ integer). Both cases have a deformation parameter $\alpha>-1$ and for $\alpha=-\frac12$ they reduce to the nondeformed $\su(2)$ model. The (discrete) spectrum of the position operator can be constructed explicitly in the deformed algebras. In the case of $\uu(2)_\alpha$ it is equidistant except for an extra shift in the middle of the spectrum. In the case of $\su(2)_\alpha$ it is no longer equidistant but has the simple form $\pm \sqrt{k(k+2\alpha+1)}$ ($k=0,1,\ldots,j$). For $k$ sufficiently large, this behaves like $\pm(k+\alpha+\frac12)$, so it tends to an equidistant distribution for large $k$.
The position and momentum wavefunctions have been constructed explicitly for the deformed models. They are given in terms of Hahn polynomials. For $\uu(2)_\alpha$, the even wavefunctions are normalized Hahn polynomials with parameters $(\alpha,\alpha+1)$, and the odd wavefunctions with parameters $(\alpha+1,\alpha)$. For $\su(2)_\alpha$, the even wavefunctions are normalized Hahn polynomials with parameters $(\alpha,\alpha)$, and the odd wavefunctions with parameters $(\alpha+1,\alpha+1)$. The plots of these discrete wavefunctions have nice properties, and in both cases they tend to the parabose wavefunctions when $j$ is large. For $\alpha\rightarrow -\frac12$ they tend to the Krawtchouk wavefunctions of the $\su(2)$ model; and of course for $\alpha\rightarrow -\frac12$ and $j\rightarrow \infty$ they tend to the canonical oscillator wavefunctions in terms of Hermite polynomials.
An interesting extra object studied in this paper is the operation that transforms position wavefunctions into momentum wavefunctions, i.e. the discrete analogue of the Fourier transform. For the case of $\su(2)_\alpha$, this discrete Hahn-Fourier transform has been constructed explicitly, and is determined by a matrix $F$. This matrix shares many classical properties with the common discrete Fourier transform matrix. In fact, it has the extra interesting feature that there is a natural basis of eigenvectors of $F$. The main computational result of the paper is the proof that the matrix elements of $F$ have a simple form in terms of terminating Saalschützian ${\,}_4F_3(1)$ series, i.e. in terms of Racah polynomials. Note that for $\uu(2)_\alpha$ this Hahn-Fourier matrix was not determined in [@JSV2011], but knowing the results for $\su(2)_\alpha$ it should be a routine computation to do this.
Acknowledgments {#acknowledgments .unnumbered}
===============
E.I. Jafarov was supported by a postdoc fellowship from the Azerbaijan National Academy of Sciences. N.I. Stoilova was supported by project P6/02 of the Interuniversity Attraction Poles Programme (Belgian State – Belgian Science Policy).
[99]{}
A. Vourdas, Rep. Progr. Phys. [**67**]{}, 267-320 (2004).
S.L. Braunstein, V. Buzek and M. Hillery, Phys. Rev. A [**63**]{}, 052313 (2001).
A. Miranowicz, W. Leonski and N. Imoto, in [*Modern nonlinear optics*]{}, Adv. Chem. Phys. [**119**]{} 155-193 (2001).
T.S. Santhanam and B. Santhanam, J. Phys. A [**42**]{}, 205303 (2009).
N.M. Atakishiyev, G.S. Pososyan and K.B. Wolf, Phys. Part. Nuclei [**36**]{}, 247-265 (2005).
N.M. Atakishiyev, G.S. Pososyan, L.E. Vicent and K.B. Wolf, J. Phys. A [**34**]{}, 9381-9398 (2001).
N.M. Atakishiyev, G.S. Pososyan, L.E. Vicent and K.B. Wolf, J. Phys. A [**34**]{}, 9399-9415 (2001).
N.M. Atakishiyev and K.B. Wolf, Rev. Mex. Fis. [**40**]{}, 366-377 (1994).
N.M. Atakishiyev and K.B. Wolf, J. Opt. Soc. Am. A. [**14**]{}, 1467-1477 (1997).
N.M. Atakishiyev, L.E. Vicent and K.B. Wolf, J. Comp. Appl. Math. [**107**]{}, 73-95 (1999).
E.I. Jafarov, N.I. Stoilova and J. Van der Jeugt, J. Phys. A [**44**]{}, 265203 (2011).
N.M. Atakishiyev, G.S. Pososyan and K.B. Wolf, Int. J. Mod. Phys. A [**18**]{}, 317-327 (2003).
B.G. Wybourne, [*Classical Groups for Physicists*]{}, (John Wiley & Sons, London, 1974).
J.E. Humphreys, [*Introduction to Lie Algebras and Representation Theory*]{}, (Springer Verlag, Berlin, 1972).
R. Koekoek, P.A. Lesky and R.F. Swarttouw, [*Hypergeometric orthogonal polynomials and their $q$-analogues*]{} (Springer-Verlag, Berlin, 2010).
M.E.H. Ismail, [*Classical and quantum orthogonal polynomials in one variable*]{} (Cambridge University Press, Cambridge, 2005).
W.N. Bailey, [*Generalized hypergeometric series*]{} (Cambridge University Press, Cambridge, 1964).
L.J. Slater, [*Generalized hypergeometric functions*]{} (Cambridge University Press, Cambridge, 1966).
N.I. Stoilova and J. Van der Jeugt, SIGMA [**7**]{}, 033 (2011).
V.E. Gorin, Funct. Anal. Appl. [**42**]{}, 180-197 (2008).
N. Mukunda, E.C.G. Sudarshan, J.K. Sharma and C.L. Mehta, J. Math. Phys. [**21**]{}, 2386-2394 (1980).
Y. Ohnuki and S. Kamefuchi, [*Quantum Field Theory and Parastatistics*]{} (Springer-Verslag, New-York, 1982).
E. Jafarov, S. Lievens and J. Van der Jeugt, J. Phys. A [**41**]{}, 235301 (2008).
J.H. McClellan and T.W. Parks, IEEE Trans. Audio Electroacoust. [**20**]{} 66–74 (1972).
G. Gasper and M. Rahman, [*Basic hypergeometric series*]{} (Cambridge University Press, Cambridge, 2004).
N.M. Temme, [*Special functions: an introduction to the classical functions of mathematical physics*]{} (Wiley, New York, 1996).
G. Gasper, J. Math. Anal. Appl. [**45**]{}, 176-198 (1974).
D.A. Varshalovich, A.N. Moskalev and V.K. and Khersonski[ĭ]{}, [*Quantum theory of angular momentum*]{}, (World Scientific Publishing Co. Inc., Teaneck, NJ, 1988).
J. Van der Jeugt, Lecture Notes Math. [**1817**]{}, 25-92 (2003). S. Lievens, J. Comp. Appl. Math. [**169**]{}, 419-430 (2004).
G. Gasper, Contemp. Math. [**254**]{}, 187–198 (2000).
A. Ballesteros and S.M. Chumakov, J. Phys. A [**32**]{}, 6261-6269 (1999).
N.M. Atakishiyev, A.U. Klimyk and K.B. Wolf, J. Phys. A [**37**]{}, 5569-5587 (2004).
------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
(a) ![Plots of a typical spectrum of the position operator, (a) in the case of the $\su(2)$ model, (b) in the case of the $\uu(2)_\alpha$ model, and (c) in the case of the $\su(2)_\alpha$ model.[]{data-label="fig1"}](spectrum0.eps "fig:")
\[5mm\] (b) ![Plots of a typical spectrum of the position operator, (a) in the case of the $\su(2)$ model, (b) in the case of the $\uu(2)_\alpha$ model, and (c) in the case of the $\su(2)_\alpha$ model.[]{data-label="fig1"}](spectrum2.eps "fig:")
\[5mm\] (c) ![Plots of a typical spectrum of the position operator, (a) in the case of the $\su(2)$ model, (b) in the case of the $\uu(2)_\alpha$ model, and (c) in the case of the $\su(2)_\alpha$ model.[]{data-label="fig1"}](spectrum1.eps "fig:")
------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
[cc]{}\
![Plots of the discrete wavefunctions $\Phi^{(\alpha)}_n(q)$ in the representation with $j=30$. The four top figures are for $\alpha=-1/2$, the middle figures for $\alpha=-0.7$, and the four bottom figures for $\alpha=2$. In each case, we plot the wavefunctions for $n=0,1,2,60$.[]{data-label="fig2"}](fig-a05-n0.eps "fig:") & ![Plots of the discrete wavefunctions $\Phi^{(\alpha)}_n(q)$ in the representation with $j=30$. The four top figures are for $\alpha=-1/2$, the middle figures for $\alpha=-0.7$, and the four bottom figures for $\alpha=2$. In each case, we plot the wavefunctions for $n=0,1,2,60$.[]{data-label="fig2"}](fig-a05-n1.eps "fig:")\
![Plots of the discrete wavefunctions $\Phi^{(\alpha)}_n(q)$ in the representation with $j=30$. The four top figures are for $\alpha=-1/2$, the middle figures for $\alpha=-0.7$, and the four bottom figures for $\alpha=2$. In each case, we plot the wavefunctions for $n=0,1,2,60$.[]{data-label="fig2"}](fig-a05-n2.eps "fig:") & ![Plots of the discrete wavefunctions $\Phi^{(\alpha)}_n(q)$ in the representation with $j=30$. The four top figures are for $\alpha=-1/2$, the middle figures for $\alpha=-0.7$, and the four bottom figures for $\alpha=2$. In each case, we plot the wavefunctions for $n=0,1,2,60$.[]{data-label="fig2"}](fig-a05-n60.eps "fig:")\
\
![Plots of the discrete wavefunctions $\Phi^{(\alpha)}_n(q)$ in the representation with $j=30$. The four top figures are for $\alpha=-1/2$, the middle figures for $\alpha=-0.7$, and the four bottom figures for $\alpha=2$. In each case, we plot the wavefunctions for $n=0,1,2,60$.[]{data-label="fig2"}](fig-a07-n0.eps "fig:") & ![Plots of the discrete wavefunctions $\Phi^{(\alpha)}_n(q)$ in the representation with $j=30$. The four top figures are for $\alpha=-1/2$, the middle figures for $\alpha=-0.7$, and the four bottom figures for $\alpha=2$. In each case, we plot the wavefunctions for $n=0,1,2,60$.[]{data-label="fig2"}](fig-a07-n1.eps "fig:")\
![Plots of the discrete wavefunctions $\Phi^{(\alpha)}_n(q)$ in the representation with $j=30$. The four top figures are for $\alpha=-1/2$, the middle figures for $\alpha=-0.7$, and the four bottom figures for $\alpha=2$. In each case, we plot the wavefunctions for $n=0,1,2,60$.[]{data-label="fig2"}](fig-a07-n2.eps "fig:") & ![Plots of the discrete wavefunctions $\Phi^{(\alpha)}_n(q)$ in the representation with $j=30$. The four top figures are for $\alpha=-1/2$, the middle figures for $\alpha=-0.7$, and the four bottom figures for $\alpha=2$. In each case, we plot the wavefunctions for $n=0,1,2,60$.[]{data-label="fig2"}](fig-a07-n60.eps "fig:")\
\
![Plots of the discrete wavefunctions $\Phi^{(\alpha)}_n(q)$ in the representation with $j=30$. The four top figures are for $\alpha=-1/2$, the middle figures for $\alpha=-0.7$, and the four bottom figures for $\alpha=2$. In each case, we plot the wavefunctions for $n=0,1,2,60$.[]{data-label="fig2"}](fig-a2-n0.eps "fig:") & ![Plots of the discrete wavefunctions $\Phi^{(\alpha)}_n(q)$ in the representation with $j=30$. The four top figures are for $\alpha=-1/2$, the middle figures for $\alpha=-0.7$, and the four bottom figures for $\alpha=2$. In each case, we plot the wavefunctions for $n=0,1,2,60$.[]{data-label="fig2"}](fig-a2-n1.eps "fig:")\
![Plots of the discrete wavefunctions $\Phi^{(\alpha)}_n(q)$ in the representation with $j=30$. The four top figures are for $\alpha=-1/2$, the middle figures for $\alpha=-0.7$, and the four bottom figures for $\alpha=2$. In each case, we plot the wavefunctions for $n=0,1,2,60$.[]{data-label="fig2"}](fig-a2-n2.eps "fig:") & ![Plots of the discrete wavefunctions $\Phi^{(\alpha)}_n(q)$ in the representation with $j=30$. The four top figures are for $\alpha=-1/2$, the middle figures for $\alpha=-0.7$, and the four bottom figures for $\alpha=2$. In each case, we plot the wavefunctions for $n=0,1,2,60$.[]{data-label="fig2"}](fig-a2-n60.eps "fig:")\
[^1]: Permanent address: Institute of Physics, Azerbaijan National Academy of Sciences, Javid av. 33, AZ-1143 Baku, Azerbaijan
[^2]: Permanent address: Institute for Nuclear Research and Nuclear Energy, Boul. Tsarigradsko Chaussee 72, 1784 Sofia, Bulgaria
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'This paper considers the secure state estimation problem for noisy systems in the presence of sparse sensor integrity attacks. We show a fundamental limitation: that is, $2\rho$-detectability is necessary for achieving bounded estimation errors, where $\rho$ is the number of attacks. This condition is weaker than the $2\rho$-observability condition typically assumed in the literature. Conversely, we propose a real-time state estimator that achieves the fundamental limitation. The proposed state estimator is inspired by robust control and FDI: that is, it consists of local Luenberger estimators, local residual detectors, and a global fusion process. We show its performance guarantees for $\mathcal H_2$, $\mathcal H_\infty$, and $\ell_1$ systems. Finally, numerical examples show that it has relatively low estimation errors among existing algorithms and average computation time for systems with a sufficiently small number of compromised sensors.'
author:
- 'Yorie Nakahira, and Yilin Mo, [^1][^2][^3]'
bibliography:
- 'security.bib'
- 'bib.bib'
title: 'Attack-Resilient $\mathcal H_2$, $\mathcal H_\infty$, and $\ell_1$ State Estimator'
---
Secure state estimation, fault tolerance, sparse sensor integrity attacks, Luenberger observer, robust control
Introduction
============
Fault tolerance in Cyber-physical Systems (CPSs) is of great importance [@securecontrol2008; @cardenas2009challenges; @rajkumar2010cyber; @sridhar2012cyber; @datta2016accountability]. For example in the power system, false data injection can introduce errors in state estimation and provide financial gains for attackers [@henrik2010; @Xie2011; @liu2009]. In flights, autonomous vehicles, and the Internet of Things, manipulations in software and sensing can cause human injury and economic damage [@farwell2011stuxnet; @maurer2016autonomous; @nobles2016cyber]. Motivated by these security issues, this paper studies the secure estimation problem for noisy systems in the presence of sensor integrity attacks.
For the secure estimation problem in static systems, robust estimators are extensively studied in the literature. Common robust estimators include the M-estimator, L-estimator, and R-estimator [@Kassam1985; @robust2006; @robust2009], and they are used to account for sensor integrity attacks in [@mo2015secure]. For the secure estimation problem in dynamical systems, robust control provides tools to deal with noise in estimation and control [@dahleh1994control; @zhou1996robust]. Although robust control typically assumes that system disturbances are bounded or follow well-defined distributions, such assumptions may not be valid for sensor faults caused by intelligent attackers [@mo2015secure; @datta2016accountability]. Fault detection and isolation (FDI) also provides methods for identifying and pinpointing faults in sensors [@gertler1998fault; @venkatasubramanian2003review; @isermann2006fault; @chen2012robust]. One common approach of FDI for linear dynamical systems under sensor integrity attacks is to construct *residuals* that take non-zero values only in the presence of faults (see [@patton2013issues] and references therein). The generation of such residuals is possible only when a fault is separable from normal disturbances and modeling uncertainties, which requires certain kinds of system observability.
When attackers can change the measurements of a limited number of sensors in large-scale systems, sensor attacks can be modeled as *sparse* but unbounded disturbances. For sparse sensor integrity attacks, recent literature has studied the fundamental limitation and achievable performance to identify the attacks and estimate the system states. Fawzi et al. show that if $\rho$ sensors are compromised, then $2 \rho$-observability (, the system remains observable after removing any set of $2 \rho$ sensors) is necessary to guarantee perfect attack identification and accurate state estimation for noiseless systems [@Fawzi2014]. The authors further propose to solve a $\ell_0$ problem to achieve accurate state estimation under the assumption of $2 \rho$-observability. This work is generalized to noisy systems by Pajic et al. [@pajic2017design; @pajic2017attack]. Shoukry et al. propose to use the Satisfiability Modulo Theory (SMT) solver to harness the complexity in secure estimation [@shoukry2017secure]. However, the worst case complexity for the $\ell_0$ optimization and that of the SMT solver are combinatorial. Moreover, these estimators also have delays, which may cause performance degradation when used for real-time control. To transform the problem into a convex program, Fawzi et al. and Mo et al. propose to use optimization based methods [@Fawzi2014; @mo2016secure]. To address the estimation delays, various Luenberger-like observers are proposed [@nakahira2015dynamic; @shoukry2016event; @mo2016secure; @lu2017secure; @chong2015observability; @shoukry2017secure]. It is worth noticing that the estimators proposed in [@pajic2017design; @pajic2017attack; @shoukry2016event; @mo2016secure; @lu2017secure; @chong2015observability; @shoukry2017secure] require the assumption of $2 \rho$-observability or stronger to guarantee accurate attack identification and secure state estimation.
In this paper, we consider the fundamental limitation and achievable performance to achieve fault-tolerant estimation. By fault-tolerant estimation, we refer to achieving bounded estimation errors. Compared with fault identification, fault-tolerant estimation requires relaxed assumptions and accounts for potentially non-detectable and non-identifiable attacks in noisy systems. We prove that a necessary condition to achieve fault-tolerant estimation under $\rho$ compromised sensors is that the system is $2 \rho$-detectable (the system needs to remain detectable after removing any set of $2 \rho$ sensors). This necessary condition suggests that, if a system has many stable modes, then the number of sensors required to achieve fault-tolerant estimation is much smaller than that to achieve fault identification. Conversely, we propose a secure state estimator that guarantees bounded estimation error under the assumption of $2\rho$-detectability. The proposed state estimator is inspired by robust control and FDI: that is, it consists of the local Luenberger estimators, the local residual detectors, and a global fusion process. A preliminary version of this paper was presented at the 2015 IEEE Conference on Decision and Control, deriving the worst-case estimation errors in the $\ell_1$ system [@nakahira2015dynamic]. This paper extends the result of [@nakahira2015dynamic] to the $\mathcal H_2$ system and the $\mathcal H_\infty$ system. To the best of our knowledge, our paper is the first to show that a mixture of two-norm bounded and sparse-unbounded input can produce *two-norm* bounded output. Finally, numerical examples show that the proposed state estimator has relatively low estimation errors among existing algorithms and average computation time for systems with a sufficiently small number of compromised sensors.
Preliminary
===========
Notations
---------
The set of natural numbers is denoted $\mathbb N$, the set of non-negative integers is denoted $\mathbb Z_+$, the set of real numbers is denoted $\mathbb R$, the set of non-negative real numbers is denoted $\mathbb R_+$, and the set of complex numbers is denoted $\mathbb C$. The cardinality of a set $S$ is denoted $| S |$. A sequence $\{x(t)\}_{t \in Z_+}$ is abbreviated by the lower case letter $x$, and the truncated sequence from $t_1$ to $t_2$ is denoted $x(t_1:t_2)$. Let $\|x\|_0 = | \{ i: \exists t \text{ s.t. } x_i(t) \neq 0 \}|$ denote the number of entries in $x$ that take non-zero values for some time. The infinity-norm of a sequence $x \in \mathbb R^n$ is defined as $\|x\|_\infty \triangleq \sup_{t} \max_{i} |x_i(t)|,$ and the two-norm of a sequence $x$ is defined as $\|x\|_2 \triangleq (\sum_{t=0}^\infty \sum_{i = 1}^n |x_i(t)|^2 )^{1/2}.$ Similarly, the norms of a truncated sequence $x(0:T)$ are defined as $
\|x(t_1:t_2)\|_\infty \triangleq \max_{t_1 \leq t \leq t_2}\max_i |x_i(t)| $ and $
\|x(t_1:t_2)\|_2 \triangleq (\sum_{t=t_1}^{t_2}\sum_{i = 1}^n |x_i(t)|^2 )^{1/2}.
$ Let $\ell_\infty$ be the space of sequences with bounded infinity-norm, and $\ell_2$ be the space of sequences with bounded two-norm.
LTI Systems and System Norms
----------------------------
Let $\plant$ be the following discrete-time linear time-invariant (LTI) system: $$\begin{aligned}
\label{eq:LTI}
&x(t+1) = A x(t) + B w(t) , & y(t) = C x(t) + D w(t),\end{aligned}$$ with the initial condition $x(0) = 0$, system state $x(t) \in \mathbb R^n$, system input $w(t) \in \mathbb R^l$, and system output $y(t) \in \mathbb R^m$. The transfer matrix of the system is = . The transfer function of the system is $\hat \plant(z) = B ( z I - A )^{-1} C + D$. For $\ell_p$ system input and $\ell_q$ system output, the system norm (namely, *induced norm*) is given by $$\begin{aligned}
\label{eq:induced_norm}
\|\plant\|_{p \rightarrow q} \triangleq \sup_{\| w\|_p \neq 0} \frac{ \|y \|_q}{\| w\|_p}.
\end{aligned}$$ In particular, the induced-norms for $(p,q) = (2, 2), (2, \infty), (\infty, \infty)$ are given by $\mathcal H_\infty$, $\mathcal H_2$, and $\mathcal L_1$ norms, respectively. These induced-norms are bounded when $A$ is strictly stable (, all the eigenvalues of $A$ is in the open unit circle). See [@dahleh1994control; @zhou1996robust] for further details.
To construct a linear state estimator with bounded estimation errors, the LTI system is required to be detectable, , there exist some matrix $K$ such that $A+KC$ is stable. Given the matrix $K$ such that $A+KC$ is strictly stable, we can construct a linear estimator \[eq:LTIest\] x(t+1) = A x(t) - K ( y(t) - Cx(t)),x(0)=0. We define its estimation error $e$ and residual vector $r$ as $$\begin{aligned}
&e(t) \triangleq x(t) - \hat x(t), &r(t) \triangleq y(t) - C\hat x(t),
\label{eq:deferrorandresidue}\end{aligned}$$ respectively. The signals $e$ and $r$ satisfy the following dynamics $$\begin{aligned}
& e(t+1) = (A + K C) e(t) + (B + KD) w(t), & e(0) = 0\\
& r(t) = C e(t) + D w(t),\end{aligned}$$ Thus, the LTI system from $w$ to $e$, $E(K)$, and the LTI system from $w$ to $r$, $\pwe(K)$, are respectively given by \[eq:def\_Ek\] &E(K) =\
\[eq:def\_Gk\] &(K) = . Because $A + K C$ is strictly stable, both $E(K) $ and $\pwe(K)$ have bounded induced-norms. The induced-norms upper-bound the values of $\| e \|_q$ and $\| r \|_q$ as follows. \[lemma:nec\_original\] If $\| w \|_p \leq 1$, then the estimation error $e$ and residual vector $r$ satisfy $$\begin{aligned}
&&\| e \|_q \leq \| E(K) \|_{p \rightarrow q},
&&\| r \|_q \leq \| \pwe(K)\|_{p \rightarrow q}.
$$
Problem Formulation {#sec:problem}
===================
We study the secure state estimation problem in the presence of sensor attacks. Consider the discrete-time LTI system: $$\label{eq:system}
\begin{aligned}
&x(t+1) = A x(t) + B w(t) , & x(0)=0\\
&y(t) = C x(t) + D w(t) + a(t),
\end{aligned}$$ where $x(t) \in \inr^{n}$ is the system state, $w(t) \in \inr^l$ is the input disturbance, $y(t) \in \inr^{m}$ is the output measurement, and $a(t) \in \inr^{m}$ is the bias injected by the adversary (we call $a(t)$ the *attack*). This system is illustrated in Fig. \[fig:robustestimationwithattack\]. The time indices $t \in \mathbb Z_+$ are non-negative integers and start from zero. Without loss of generality, we assume that the disturbance matrix $B$ has full row rank (otherwise we can perform the Kalman decomposition and work on the controllable space of $(A,B)$). Each sensor is indexed by $i \in \{ 1, \cdots, m\}$ and produces measurement $y_i(t)$, which jointly comprises the measurement vector $y(t) = [y_1(t),\dots,y_m(t)]^T$. Sensor $i$ is said to be *compromised* if $a_i(t) \neq 0$ at some time $t \in \mathbb Z_+$ and is said to be *benign* otherwise. The maximum number of sensors that the attacker can compromise is $\gam$, , $$\begin{aligned}
\label{eq:attack}
\|a\|_0 \leq \gam .\end{aligned}$$ If $a(t)$ satisfies , then we say that it is $\gam$*-sparse*. Let $\mathcal S \triangleq \{1,\dots,m\}$ denote the set of all sensors, $\mathcal C\subset \mathcal S$ denote the set of compromised sensors, and $\G \triangleq \mathcal S\backslash \mathcal C$ denote the set of benign sensors. The set $\mathcal C$ is assumed to be unknown.[^4] A causal state estimator is an infinite sequence of functions $\{f_t\}$ where $f_t$ is a mapping from all output measurements to a state estimate: $$\label{def:state_estimator}
\hat x(t) = f_t( y(0:t-1) ).$$ The estimation error of is defined as the difference between the system state and the state estimate:[^5] $$e(t) \triangleq x(t) - \hat x(t) .$$
We consider the input containing a mixture of a $p$-norm bounded disturbance and a $\gam$-sparse attack and study the following worst-case estimation error in $q$-norm: $$\label{eq:defworstcaseerror}
\sup_{\|w\|_p\leq 1,\,\|a\|_0\leq \gam} \|e\|_q,$$ where $$\begin{aligned}
(p, q) = (2,2),(2,\infty), (\infty,\infty). \end{aligned}$$ We consider instead of attack isolation because the attack on a noisy system may not be correctable in the sense defined in [@Fawzi2014; @pajic2014robustness].
An causal state estimator $\{f_t\}$ is said to be $\epsilon$-resilient to attack if its worst-case estimation error satisfies $\sup_{\|w\|_p\leq 1,\,\|a\|_0\leq \gam} \|e\|_q < \epsilon$, where $\epsilon$ is a positive and finite scalar.
When the estimator is $\epsilon$-resilient for some finite $\epsilon >0$, then we say the state estimator is resilient to attack. The goal of this paper is to study the design problem of a resilient state estimator $\{ f_t\}$. Towards that end, we first show a fundamental limitation for the existence of a resilient estimator (Section \[sec:limitations\]), and we then propose a resilient estimator (Section \[sec:design\]) and analyze the estimation errors (Section \[sec:est\_analysis\]).
Necessary and Sufficient Conditions for Resilience to Attack {#sec:main}
============================================================
In this section, we first provide a necessary condition for the existence of a resilient estimator and then, assuming the necessary condition, propose a resilient estimator. We first define some notation that will be used later.
Let $e_i$ be the $i_{th}$ canonical basis vector of the space $\mathbb R^m$ and $\mathcal I = \{i_1,\dots,i_{m'}\}\subseteq \mathcal S$ be an index set with carnality $m' (\leq m)$. We define the projection map $P_{\I}: \mathbb R^m \rightarrow \mathbb R^{m'}$ as $$\begin{aligned}
\label{eq:proj_matrix}
P_\I = \begin{bmatrix}
e_{i_1}&\dots&e_{i_{m'}}
\end{bmatrix}^T\in \mathbb R^{m' \times m}.
\end{aligned}$$
Using $P_{\mathcal I}$ in , the measurements of the set of sensors $\mathcal I \subset \mathcal S$ can be written as $$\begin{aligned}
y_{\I}(t) \triangleq P_{\mathcal I} y(t) \in \mathbb R^{m'}.
\end{aligned}$$ Similarly, the measurement matrix and the sensor noise matrix corresponding to the set of sensors $\mathcal I $ can be respectively written as $$\begin{aligned}
&C_{\mathcal I} \triangleq P_\I C, &D_{\mathcal I} \triangleq P_\I D.\end{aligned}$$
Necessary Condition for Resilience to Attack {#sec:limitations}
--------------------------------------------
In this section, we give a fundamental limitation for achieving bounded worst-case estimation errors.
The system is said to be $\chi$-detectable if $(A, C_\K)$ is detectable for any set of sensors $\K \subset \mathcal S $ with cardinality $|\K| = m - \chi$.
\[thm:nonexistence\] If system is not $2 \rho$-detectable, then there is no state estimator $\{ f_t\}$ that is $\epsilon$-resilient to attack for any finite $\epsilon > 0$.
Theorem \[thm:nonexistence\] implies that the following (denote as Condition A) is necessary for the existence of a resilient state estimator:
- The system is $2 \rho$-detectable.
The Proposed Estimator {#sec:design}
----------------------
Assuming condition A, we now propose a resilient state estimator. The proposed estimator constitutes two procedures: 1) local estimation and 2) global fusion. The local estimators are defined by groups of $m-\gam$ sensors for all combinations $$\mathcal V \triangleq \{\I\subset \mathcal S:|\I|=m-\gam\}.$$ The number of such groups (local estimators) is $|\mathcal V | = \left( \begin{matrix} m \\ \gam \end{matrix} \right)$. Each local estimator $\mathcal I$ generates a state estimation $\hat x^\I$ separately based on the measurements of its sensors $y_{\mathcal I}$. In the global fusion process, the state estimate $\hat x$ is generated using the estimates from all local estimators $\mathcal I \in \mathcal V$. With slight overlap of notation, we use $\mathcal I \in \mathcal V$ to refer to a set of sensors as well as to the estimator that uses these sensors. Next, we outline these procedures and formally state the estimator in Algorithm 1.
### Local Estimations
From Assumption A, for any set of sensors $\I \in \comset$, there exists a matrix $K^\I \in \mathbb R^{(m-\gam) \times n}$ such that $A + K^\I C_\I$ is strictly stable (has all eigenvectors in the open unit circle).[^6] Using this matrix $K^\I$, we construct a local estimator that only uses measurements from the set of sensors $\I$ to produce a *local state estimate* $\hat x^{\I}$:[^7] $$\begin{aligned}
\label{eq:est}
\hat x^{\I}(t+1) = A \hat x^\I (t) - K^\I ( y_\I(t) - C_\I \hat x^\I(t))\end{aligned}$$ with the initial condition $\hat x^\I (0) = 0$. The estimation error and residual vector of is respectively defined as $$\begin{aligned}
\label{eq:e_def}
&e^\I(t) \triangleq x(t) - \hat x^\I(t)\\
\label{eq:r_def}
&r^\I(t) \triangleq y_\I(t) - C_\I\hat x^\I(t) \end{aligned}$$ The LTI system from $w$ to $e^\I$ is $E^\I(K^\I)$ defined in , whereas the LTI system from $w$ to $r^\I$ is $\pwe^\I(K^\I)$ defined in .
When the set $\mathcal I$ does not contain any compromised sensors, *i.e.*, $a_\I = 0$, the residual vector $r^\I(t)$ is determined by disturbance $w$ alone and is bounded by \[eq:localdetection\] && r\^\_q \^(K\^)\_[p q]{}. Condition can only be violated when the set $\mathcal I$ contains compromised sensors, so is a necessary condition for all the sensors in set $\I$ to be benign. The local estimator at time $t$ uses the necessary condition to determine the validity of its estimate and label local estimator $\I$ to be *invalid* upon observing $\| r^\I(0:t) \|_q > \| \pwe^\I(K^\I)\|_{p \rightarrow q} .
$
### Global Fusion
From above, the set of valid local estimators $\I \in \comset(t)$ is characterized as $$\begin{aligned}
\comset(t) \triangleq \left\{\I\in \mathcal S:\| r^\I(0:t) \|_q \leq \| \pwe^\I(K^\I)\|_{p \rightarrow q} \right\}.
\label{eq:comset}\end{aligned}$$ Using $\comset(t)$, we compute the *global state estimate* as follows: $\hat x(t) = [ \hat x_1(t) , \hat x_2(t) , \cdots , \hat x_n(t) ]$, where $$\label{eq:global_fusion}
\hat x_i(t) =
\begin{dcases}
\frac{1}{2} \left( \min_{\I \in \comset(t)} \hat x^\I_i (t) + \max_{\J \in \comset(t)} \hat x^\J_i (t) \right) & q = \infty\\
\frac{1}{ |\comset(t)| } \sum_{\I(t) \in \comset(t)} \hat x^\I_i(t) & q = 2.
\end{dcases}$$
\[alg:estimation\_algorithm\]
Initialize $\comset(0) \gets \comset$ and $\hat x^\I (0) \gets 0, \I \in \comset(0)$ (Local Estimation) Initialize $\comset (t) \gets \emptyset$ Determine $\hat x^\I(t)$ from and $r^\I(t)$ from $\comset (t) \gets \{ \comset (t), \I \}$ Obtain estimate $\hat x(t)$ from (Global Fusion)
Resilience of the Proposed Estimator {#sec:est_analysis}
------------------------------------
Previous works have shown that there exist estimators that can detect the attacks and recover the exact state for noiseless systems if the system is $2 \rho$-observable [@Fawzi2014 Proposition 2][@chong2015observability Theorem 1][@shoukry2016event Theorem 3.2]. In this section, we show that the proposed estimator is resilient to attack when the system is $2 \gam$-detectable.
\[thm:error\_bound\_infinity\] The estimator in Algorithm 1 has a bounded estimation error. In particular, the estimation error is upper-bounded by $$\begin{aligned}
\begin{array}{lr}
\hspace{-2mm}\underset{\I \in \comset}{\max} \|E^{\I}(K^{\I})\|_{\infty} + \underset{\I,\J \in \comset}{\max}\sqrt{ \frac{1}{2} \log | \comset | } \; \part^{\I,\J}_{{2,2}} \hspace{-8mm}&\text{if }(p,q) = (2, 2)\\
\hspace{-2mm}\underset{\I,\J\in \comset }{\max}\Big( \|E^{\I}(K^{\I})\|_{2} + {1 \over 2} \part^{\I,\J}_{{2, \infty}}\Big) &\text{if }(p,q) = (2, \infty) \\
\hspace{-2mm} \underset{\I,\J\in \comset }{\max} \Big( \|E^{\I}(K^{\I})\|_{1} + {1 \over 2} \part^{\I,\J}_{{\infty, \infty}}\Big) &\text{if }(p,q) = (\infty, \infty) .
\end{array}\end{aligned}$$ In the above formula, the term $\part^{\I,\J}_{{p,q}}$ is defined as $$\begin{aligned}
&\part^{\I,\J}_{{p,q}} = \alpha_{{p,q}}^{\I \cap\J}(\beta^{\I, \I \cap\J} _{{p,q}}+\beta^{\J, \I \cap\J}_{{p,q}})\\
& \alpha^\K_{{p,q}} \triangleq \inf_{K:A+KC_\K\text{ strictly stable}}\left\| \left[
\begin{array}{c|c}
A+K C_\K & \begin{bmatrix}
I&K
\end{bmatrix}\\
\hline
I & 0
\end{array}
\right]\right\|_{p \rightarrow q} \\
& \beta^{\I,\K}_{{p,q}} \triangleq \left\|\begin{bmatrix}
-K^\I\\
P_{\K,\I}
\end{bmatrix}\right\|_{p \rightarrow q} \|r^\I(0:T)\|_p,
\end{aligned}$$ where $P_{\K,\I}\in \mathbb R^{|\K|\times|\I|}$ is the unique solution of $P_\K = P_{\K,\I} P_\I$, and $\| \cdot \|_{p \rightarrow q}$ is an induced norm on matrix.
An immediate consequence of Theorem \[thm:error\_bound\_infinity\] is that condition A is a necessary and sufficient condition for the construction of a resilient state estimator, and that Algorithm 1 resilient to attack. The estimation error upper-bound in Theorem \[thm:error\_bound\_infinity\] decomposes into two terms: $\| E^\I(K^\I) \|_{p \rightarrow q}$ and the remaining. The first term $\| E^\I(K^\I) \|_{p \rightarrow q}$ characterize the error between a local estimator and the true state. That is, if the local estimator $\I$ is used for a system with no attack ($a \equiv 0$), then its estimation error is bounded by $\| E^\I(K^\I) \|_{p \rightarrow q}$. The second term exists due to the attack in an unknown set of sensors. When $(p,q) = (2, 2)$, the error upper-bound grows at the order $o(\sqrt{ \rho \log m} )$ for $m \rightarrow \infty$. Hence, the error can be kept small even for systems with large $m$. Moreover, it shall be noted that an increase in the tolerable number of compromised sensors $\gam$ may result in an increase in both terms, thus increasing the worst-case estimation error $\sup_{\| w \|_p \leq 1, \|a\|_0 = 0 } \| e\|_q $.
A necessary and sufficient condition for the existence of an $\epsilon$-resilient estimator for some finite $\epsilon >0$ is that $(A,C_\K)$ is detectable for any index set $\K\subset \mathcal S$ with cardinality $m - 2\gam$.
\[thm:esterror\] Consider system (\[eq:system\]) with $\gam$-sparse attack. The state estimator in Algorithm 1 is $\epsilon$-resilient to attack for some finite $\epsilon > 0$.
Proof for Resilience of the Proposed Estimator
==============================================
We highlight important parts of the proof of Theorem \[thm:error\_bound\_infinity\] in this section and present the complete proof in the extended version of this paper [@sup]. The proof of Theorem \[thm:error\_bound\_infinity\] has two procedures: 1) bounding local estimation errors, and 2) bounding global fusion errors. Specifically, from the triangular inequality, at any time $t \in \mathbb Z_+$, the estimation error satisfies $$\begin{aligned}
\nonumber
\|e \|_q
&= \left \| \big( x - \hat x^{\I} \big) + \big( \hat x^{\I} - \hat x \big) \right\|_q \\
\label{eq:globalerror_1}
&\leq\left \| x - \hat x^{\I} \right \|_q + \left\| \hat x^{\I} - \hat x \right \|_q .
\end{aligned}$$ where $\I \in \mathcal B \subset \comset$ is a set that only contains benign sensors (denote $\I$ as the *benign estimator*). The benign estimator $\I$ exists from assumption . The first term $ \|x - \hat x^{\I}\|_q$ can be bounded using Lemma \[lemma:nec\_original\] by $$\begin{aligned}
\label{eq:error_firstterm}
\|x - \hat x^{\I}\|_q \leq \|E^{\I}(K^{\I})\|_{p \rightarrow q} .\end{aligned}$$ Now it only remains to show that the second term is bounded.
To bound the second term, we first bound the difference between the estimates of any two valid local estimators $\mi,\,\mj \in \comset(T)$ up to time $T \in \mathbb Z_+$, which is given in Lemma \[lemma:difference\]. We then use Lemma \[lemma:difference\] to show that the difference between the estimates of the benign estimator $\I \in \comset$ and the global estimator is finite. This is shown in Lemma \[lemma:part2\_1\] for $(p, q) = (2,2)$ and in Lemma \[lemma:part2\_2\] for $(p, q) = (2,\infty),(\infty,\infty)$. In these lemmas, each set of sensors in $\comset$ are labeled into $$\begin{aligned}
\J_1, \J_2, \cdots , \J_{ |\comset| }.\end{aligned}$$ \[lemma:difference\] Assume that Condition A holds. Let $\mi,\,\mj\in \mathcal V(T)$ be two sets of sensors that are valid at time $T$. The divergence between the local estimator $\mi$ and $\mj$ up to time $T$ satisfies $$\begin{aligned}
\label{eq:local_divergence_cond}
\| \hat x^{\mi}(0:T) - \hat x^{\mj}(0:T) \|_q \leq \part^{\mi,\mj}_{{p,q}} ,
\end{aligned}$$ where right hand side is finite, , $\part^{\mi,\mj}_{{p,q}} < \infty$ .
\[lemma:part2\_1\] If condition holds for $(p,q) = (2,2)$ at all time $T \in \mathbb Z_+$, then the divergence between the benign estimator $\I$ and the global estimator satisfies $$\begin{aligned}
\|\hat x^{\I} - \hat x \|_2 \leq \max_{\mi, \mj \in \mathcal V } \sqrt{ \frac{1}{2} \log |\mathcal V| }\; \part^{\mi,\mj}_{p,q} .
\end{aligned}$$
\[lemma:part2\_2\] If condition holds for $(p,q) = (2,\infty), (\infty,\infty)$ at all time $T \in \mathbb Z_+$, then the divergence between the benign estimator $\I$ and the global estimator satisfies $$\begin{aligned}
\|\hat x^{\I} - \hat x \|_\infty \leq {1 \over 2} \max_{\J \in \comset} \part^{\I,\J}_{{p, \infty}} .
\end{aligned}$$
Taking supremum over all $t \in \mathbb Z_+$ and maximizing over all sensor sets $\I$ in Lemma \[lemma:part2\_1\], we obtain $$\begin{aligned}
\sup_{\substack{ \|w\|_2 \leq 1 \\ \|a\|_0\leq \gam}} \|e\|_2 \leq
\underset{\I \in \comset}{\max} \|E^{\I}(K^{\I})\|_{\infty} + \underset{\mi, \mj \in \comset}{\max}\sqrt{ \frac{1}{2} \log | \comset | } \; \part^{\I,\J}_{{2,2}} .\end{aligned}$$ Applying similar argument for the case of $q = \infty$, we obtain $$\begin{aligned}
\|e\|_\infty \leq \max_{\I, \J \in \comset} \left( \|E^{\I}(K^{\I})\|_{ p \rightarrow \infty} + \frac{1}{2} \part^{\I,\J}_{{p,\infty}} \right),\end{aligned}$$ where $p = 2, \infty$.
Proof of Lemma \[lemma:difference\] {#sec:proof1}
-----------------------------------
We first present a lemma, using which we prove Lemma \[lemma:difference\].
\[lemma:y0\] Consider system (\[eq:LTI\]) where $(A,C)$ is detectable and $\|w \|_p \leq 1$. If $y(t) = 0$ for all $t = 0, 1, \cdots, T$, then \[eq:finitestatenorm\] x(0:T) \_q \_[K:A+KC]{} E(K) \_[p q]{}, where $E(K)$ is given in .
As $(A,C)$ is detectable, $A+KC$ is strictly stable for some matrix $K$. For such stabilizing $K$, we can construct the state estimator . Since $y(0:T) = 0$, the state estimator produces zero estimate $\hat x(0:T) = 0$. From Lemma \[lemma:nec\_original\], we obtain $$\begin{split}
\| x(0:T) \|_q
= \|e(0:T)\|_q
\leq \| E(K) \|_{p \rightarrow q}.
\end{split}$$ Taking infimum over all $K$ such that $A + KC$ is strictly stable, we obtain .
Let $\mi,\,\mj \in \comset(T)$. We first compute the dynamics of the local estimates $\hat x^\mall(t)$, $i = 1,2$. From and , $$\label{eq:suberror}
\begin{aligned}
&\hat x^\mall(t+1) = A\hat x^\mall(t) -K^\mall r^\mall(t), \hat x^\mall(0) = 0 \\
&y_\mall(t) = C_\mall \hat x^\mall(t) + r^\mall(t)
\end{aligned}$$ for $t \leq T$. We define the sequences $\phi^\mall(t)$ and $\varphi^\mall(t)$ by $$\begin{aligned}
&\phi^\mall(t) \triangleq -K^\mall r^\mall(t) ,
&\varphi^\mall(t) \triangleq P_{{\K_{1,2}},\mall} r^\mall(t),\end{aligned}$$ where $P_{{\K_{1,2}},\mall}\in \mathbb R^{| {{\K_{1,2}}} |\times|\mall|}$ is the unique solution of $P_{{\K_{1,2}}} = P_{{\K_{1,2}},\mall} P_\mall$. Let $\mathcal K_{1,2} = \mi \cap \mj$ the intersection between the two sets $\mi, \mj$. As the measurements from subset $\mathcal K_{1,2} \subset \mall$ also satisfies $y_{\K_{1,2}}(t) = C_{\K_{1,2}} \hat x^\mall(t) + P_{{\K_{1,2}},\mall} r^\mall(t),$ combining with yields $$\label{eq:suberror2}
\begin{aligned}
&\hat x^\mall(t+1) = A\hat x^\mall(t) + \begin{bmatrix}
I & 0
\end{bmatrix}\begin{bmatrix}
\phi^\mall(t)\\
\varphi^\mall(t)
\end{bmatrix}, \,\hat x^\mall(0) = 0\\
&y_{\K_{1,2}}(t) = C_{\K_{1,2}} \hat x^\mall(t) + \begin{bmatrix}
0 & I
\end{bmatrix}\begin{bmatrix}
\phi^\mall(t)\\
\varphi^\mall(t)
\end{bmatrix}.
\end{aligned}$$ Now, let $\Delta (t)$ be the difference between the local estimator $\mi$ and local estimator $\mj$, , $$\begin{aligned}
\Delta (t) \triangleq \hat x^\mi(t)-\hat x^\mj(t).\end{aligned}$$ Subtracting the equation for $\mi$ from equation for $\mj$, we obtain the dynamics of $\Delta $ as follows: $$\begin{aligned}
&\Delta (t+1) = A\Delta (t) + \begin{bmatrix}
I & 0
\end{bmatrix}\begin{bmatrix}
\phi^\mi(t)-\phi^\mj(t)\\
\varphi^\mi(t)-\varphi^\mj(t)
\end{bmatrix},\,\Delta(t) = 0,\nonumber\\
&0 = C_{\K_{1,2}} \Delta (t) + \begin{bmatrix}
0 & I
\end{bmatrix}\begin{bmatrix}
\phi^\mi(t)-\phi^\mj(t)\\
\varphi^\mi(t)-\varphi^\mj(t)
\end{bmatrix}.\end{aligned}$$
Because a valid set satisfies , the residual vectors of estimator $\mall$, $i = 1,2$, are bounded by $
\|r^\mall(0:T)\|_q \leq\|\pwe^\mall(K^\mall)\|_{p \rightarrow q},
$ which results in $$\begin{aligned}
\nonumber
\left\|\begin{bmatrix}
\phi^\mall(0:T)\\
\varphi^\mall(0:T)
\end{bmatrix}\right\|_p
&\leq \left\|\begin{bmatrix}
-K^\mall\\
P_{{\K_{1,2}},\mall}
\end{bmatrix}\right\|_{p \rightarrow p} \|r^\mall(0:T)\|_p \\
&= \beta^{\mall,\mi \cap \mj}_{p,q} \label{eq:phibound_I}.\end{aligned}$$ From the triangle inequality, we obtain $$\left\| \begin{bmatrix}
\phi^\mi(0:T) - \phi^\mj(0:T) \\
\varphi^\mi(0:T) - \varphi^\mj(0:T)
\end{bmatrix} \right \|_p
\leq \beta^{\mi,\mi \cap \mj}_{p,q} + \beta^{\mj,\mi \cap \mj}_{p,q} .$$ Substitute $\phi^\mi-\phi^\mj$ for $u$ in Lemma \[lemma:y0\] and $\Delta (t)$ for $x$, we obtain $$\| \hat x^{\mi}(0:T) - \hat x^{\mj}(0:T) \|_q \leq \alpha_{{p,q}}^{\mi \cap\mj}(\beta^{\mi, \mi \cap\mj} _{{p,q}}+\beta^{\mi, \mi \cap\mj}_{{p,q}}).$$
Proof of Lemma \[lemma:part2\_1\] {#sec:proof2}
---------------------------------
Define the following two optimization problems: $$\begin{aligned}
{3}
\op(n) := &\max_{ \s_k(i) \geq 0} &&\sum_{i=1}^{n}\left( \frac{1}{n-i-1} \right)^2\left( \sum_{k = i}^{n} \s_k(i)\right)^2\\
&\;\;\mathrm{s.t.}
&&\sum_{i=0}^{k} ( \s_k(i) )^2 \leq \1, \;\;\; k =1,2,\dots,n \\
\opd (n) := &\;\min_{\lambda_i > 0} &&\sum_{i=1}^{n} \lambda_i\\
&\;\;\mathrm{s.t.} &&\sum_{i=1}^j\frac{1}{\lambda_i}\leq (j+1)^2,\;\;\; j =1,2,\dots,n.
\end{aligned}$$ With the slight abuse of notation, we will also denote $\op(n)$, $\opd(n)$ as the optimal solutions of the optimization problem $\op(n)$, $\opd(n)$, respectively. We first show that $\op(N-1)$ with $$\begin{aligned}
\1 = \act\end{aligned}$$ is an upper-bound of $\left\| \hat x^{\I} - \hat x \right \|_2$ (Lemma \[lem:2norm\_1\]). The problem $\op(n)$ is then converted into its dual problem $\opd(n)$, between which the duality gap is zero (Lemma \[lem:2norm\_2\]). The dual problem $\opd (n)$ admits an analytical solution that can be upper-bounded by a simple formula (Lemma \[lem:2norm\_3\]).
\[lem:2norm\_1\] If condition holds for $(p,q) = (2,2)$, then the divergence between the benign estimator $\I$ and the global estimator satisfies $$\begin{aligned}
\label{eq:centerl2error}
\|\hat x^{\I} - \hat x \|_2^2 \leq \op (N-1),
\end{aligned}$$ where $N = |\mathcal V|$ and $\1 = \act$.
\[lem:2norm\_2\] The problems $\op (n)$, $\opd (n)$ have identical optimal values, , $\op (n) = \opd (n)$.
We use the following lemma to prove Lemma \[lem:2norm\_2\]. The proof of Lemma \[lem:matrix\_inequality\] is given in the extended version of this paper [@sup].
\[lem:matrix\_inequality\] The following two inequalities are equivalent $$\begin{aligned}
\label{eq:lem7_cond1}
& \Lambda= \diag( \lambda_1, \lambda_2 , \cdots, \lambda_n) \geq \mathbf 1\mathbf 1^T \\
\label{eq:lem7_cond2}
&\sum_{i=1}^n \frac{1}{\lambda_i} \leq 1 , \text{ and } \lambda_j > 0,\;\; j =1,\dots,n\end{aligned}$$
Let $v \in \mathbb R^{n(n-1)/2}$ be a vector that is composed of $\s_{i+1:n}(j) = \{\s_{i+1}(j),\s_{i+2}(j) , \cdots,\s_{n}(j) \} $ for all $j = 0, 1, \cdots, p$, , $$v \triangleq \begin{bmatrix}
\s_{1:n}(0) ,\s_{2:n}(1), \cdots,\s_{n}(n) \end{bmatrix} .$$ Let the following matrices be defined as $$\begin{aligned}
\nonumber& X = vv^T \geq 0 \\
\label{eq:def_X_F}
&F_0 = \text{diag}\left(\frac{1}{2^2},\cdots,\frac{\mathbf 1_{n-1}\mathbf 1^T_{n-1}}{(n)^2},\frac{\mathbf 1_{n}\mathbf 1^T_{n}}{(n+1)^2}\right) \\
\nonumber&F_i = \text{diag}\left( e_{n,i } ,e_{n-1,i } ,\cdots, e_{n-i+1,i } , 0_{n-i} , \cdots , 0_{1} \right)\end{aligned}$$ where $1_{k}$ is a $k$-dimensional vector with all elements being $1$; $e_{k,j}$ is a $k$-dimensional row vector with $j$-th entry being $1$ and other entries being $0$; and $0_k$ is a $k$-dimensional row vector with all elements being $0$. Using SDP relaxation [@low2013convex], the problem $\op(n)$ can be converted into $$\begin{aligned}
{3}
\op'(n) = & \max_{X \geq 0} \;\;&&\tr(F_0X)&\\
&\;\mathrm{s.t.} \;\;&& \tr(F_iX) \leq \1,\,\forall i = 1,\dots,n &\end{aligned}$$ Hence $\op(n) \leq \op'(n)$. This relaxation can be observed from the following relations: $$\begin{aligned}
& \sum_{i=0}^{n-1}\left( \frac{1}{n-i+1} \right)^2\left( \sum_{k = i+1}^{n} \s_k(i)\right)^2 = v^T F_0 v= \tr(F_0v v^T) \\
& \sum_{i=0}^{k} ( \s_k(i) )^2 = v^T F_i v = \tr(F_i v v^T) .\end{aligned}$$ We next show that the relaxation of the problem $\op(n)$ to the semidefinite problem $\op'(n)$ is also exact. Assume that $\op (n)$ is feasible and bounded. Let $X^*= \{ x^*_{ij}\}$ be the optimal solution of $\op'$. Define $X$ as $$X = \begin{bmatrix}
\sqrt{x_{11}^*}&\dots&\sqrt{x_{nn}^*}
\end{bmatrix}^T \begin{bmatrix}
\sqrt{x_{11}^*}&\dots&\sqrt{x_{nn}^*}
\end{bmatrix} .$$ From $x_{ii} = x^*_{ii}$ and , $X$ satisfies the contraints $$\begin{aligned}
\label{eq:rem2_1}
\tr(F_iX) = \tr(F_iX^*) = \1.\end{aligned}$$ Furthermore, because $X^*$ is the optimal solution and $x_{ij} = \sqrt{x_{ii}^*x_{jj}^*}\geq x_{ij}^*$ (due to $X^* \geq 0$), $$\begin{aligned}
\label{eq:rem2_2}
\tr(F_0X^*) \geq \tr(F_0X) \geq \tr(F_0X^*) . \end{aligned}$$ Therefore, and shows that $\op (n) = \op'(n)$.
Next, we consider the following dual problem of $\op'(n)$: $$\begin{aligned}
\label{eq:sdpdual}
\opd' (n) = \min_{\lambda} \; & \; \1 \sum_{i=1}^{n} \lambda_i \\
\mathrm{s.t.} \;\; &\sum_{i=1}^{n} \lambda_i F_i\geq F_0. \nonumber
\end{aligned}$$ Because there exists an strictly positive definite matrix $X > 0$ such that $\tr(F_i X) =\1$ for all $i = 1, 2, \cdots, n$, from Slater’s condition [@Boyd2004], strong duality holds between $\op'(n)$ and $\opd'(n)$. Let $ \Lambda_{1:i} = \diag(\lambda_1,\dots,\lambda_i)\in \mathbb R^{i\times i}$, and observe that $
\sum_{i=1}^{n} \lambda_i F_i = \diag(\Lambda_{1:1},\Lambda_{1:2},\dots,\Lambda_{1:n}).
$ Hence, the constraint $\sum_{i=1}^{n}\lambda_i F_i \geq F_0$ is equivalent to $$\begin{aligned}
& \diag(\lambda_1,\dots,\lambda_j) \geq \frac{1}{(j+1)^2} \mathbf 1\mathbf 1^T ,\,\forall j=1,\dots,n, \\
& \Longleftrightarrow \sum_{i=1}^j\frac{1}{\lambda_i}\leq (j+1)^2,\;\; \lambda_j > 0,\;\; j =1,\dots,n, \end{aligned}$$ where the second line is due to Lemma \[lem:matrix\_inequality\]. Therefore, the dual problem $\opd'(n)$ can be reformulated into $\opd(n)$.
\[lem:2norm\_3\] The solution of the problem $\opd (n)$ satisfies $$\opd(n) = \1 \left\{ \frac{1}{4} + \sum_{i = 2}^n \frac{1}{2 i + 1} \right \} \leq \frac{1}{2} \1 \log(n+1) .
\label{eq:oddharmonicseries}$$
Applying Lemma \[lemma:difference\] and Lemma \[lem:2norm\_1\], Lemma \[lem:2norm\_2\], and then Lemma \[lem:2norm\_3\] consecutively, we obtain $$\begin{aligned}
\|\hat x^{\I} - \hat x \|^2_2 \leq \op = \opd
\leq \frac{1}{2} \log( N ) \act. \end{aligned}$$
Numerical examples
==================
In this section, we study the proposed estimator numerically and compare it with existing algorithms from Shoukry et al. [@shoukry2017secure], Chong et al. [@chong2015observability], Pajic et al. [@pajic2014robustness], and Lu et al. [@lu2017secure]. We tested the IEEE 14-Bus system, the Unmanned Ground Vehicle (UGV), and the temperature monitor as follows.
- *IEEE 14-Bus system [@zimmerman2011matpower; @liu2009; @shoukry2017secure]:* The IEEE 14-Bus system is modeled as the system with $A$ and $C$ given in [@zimmerman2011matpower]. We additionally add process noise and sensor noise by setting $B = \begin{bmatrix} I_{10} & O_{10,35} \end{bmatrix}$ and $D = \begin{bmatrix} O_{10,35} & I_{35} \end{bmatrix}$.
- *Unmanned ground vehicle (UGV) [@pajic2014robustness; @shoukry2017secure]:* A UGV moving in a straight line has the dynamics $$\begin{aligned}
\begin{bmatrix} \dot p \\ \dot v \end{bmatrix} = \begin{bmatrix} 0 & 1 \\ 0 & -b/m \end{bmatrix} \begin{bmatrix} \dot x \\ \dot v \end{bmatrix} + \begin{bmatrix} 0 \\ 1/m \end{bmatrix} u + \begin{bmatrix} I_2 & O_{2,3} \end{bmatrix} w ,\end{aligned}$$ where $p$ is the position, $v$ is the velocity, $u$ is the force input, $w$ is the disturbance, $m$ is the mechanical mass, $b$ is the translational friction coefficient, $I_n$ is a $n$-dimensional identity matrix, and $O_{n,m}$ is a $m \times n$ zero matrix. We assume that the estimator can access the values of $u$. The UGV is equipped with a sensor measuring $x$ and two sensors measuring $v$, , $$\begin{aligned}
y = \begin{bmatrix} 1 & 0 \\ 0 & 1 \\ 0 & 1 \end{bmatrix} \begin{bmatrix}p \\ v \end{bmatrix} + \begin{bmatrix} O_{3,2} & I_2 \end{bmatrix} w .\end{aligned}$$ The system parameters are the same as [@shoukry2017secure]: $m = 0.8kg$, $b = 1$, and sampling interval $T_s = 0.1s$.
- *Temperature monitor [@mo2011sensor]:* The heat process in a plannar closed region $(z_1, z_2 ) \in [0, l ] \times [0, l ]$ can be expressed by $$\begin{aligned}
\label{eq:heat}
\frac{ \partial x } {\partial t } = \alpha \left( \frac{\partial^2 x } {\partial z^2_1 } + \frac{\partial^2 x } {\partial z^2_2 } \right) , \end{aligned}$$ where $\alpha$ is the speed of the diffusion process; and $x(z_1,z_2)$ is the temperature at position $(z_1, z_2 )$ subject to the boundary conditions $$\begin{aligned}
\frac{ \partial x } {\partial z_1 } \bigg|_{t , 0, z_2 }= \frac{ \partial x } {\partial z_1 } \bigg|_{t , l, z_2 }= \frac{ \partial x } {\partial z_1 } \bigg|_{t , z_1, 0 }= \frac{ \partial x } {\partial z_1 }\bigg|_{t , z_1, l } = 0 .\end{aligned}$$ We discretize the region using a $N \times N$ grid and the continuous-time with sampling interval $T_s$ to model into . We additionally add process noise and sensor noise by setting $w $, $B = \begin{bmatrix} I_{9} & O_{9,20} \end{bmatrix}$ and $D = \begin{bmatrix} O_{9,20} & I_{20} \end{bmatrix}$. We set $\alpha = 0.1m^2/s$, $l = 4m$, and $N = 5$ as in [@mo2011sensor].
The noise $w$ is generated from a uniform distribution between $[-1,1]$. The time horizon is set to be $T = 100$. The number of compromised sensors is set to be $1$; the compromised sensor is randomly chosen among non-critical sensors (, sensors that can be removed without losing observability). The attack signal is drawn from a Gaussian distribution with mean zero and variance $10^4$ and $1$. For the proposed algorithm, we used with $(p,q) = (2,2), (2, \infty)$, and $(\infty, \infty)$ simultaneously. We ran each example for $100$ times and recorded their average estimation errors $\| e \|_2$ and computation times. The code is written in Matlab (Windows) and runs on an Intel Core i5-4690 Processor (4x3.50GHz/6MB L3 Cache). We summarize the estimation errors and computation times in Table \[tab:errorA100\] and Table \[tab:compA100\]. Some entries are left as NA (not applicable) because the system (ii) does not satisfy the linear matrix inequality (LMI) assumption required by the algorithm proposed by [@lu2017secure]. The algorithms tested have different relative accuracies and computation times from example to example. Among all examples tested, the proposed algorithm has relatively low estimation errors and average computation times.
Proposed [@shoukry2017secure] [@chong2015observability] [@pajic2017design] [@lu2017secure]
------- ---------- ---------------------- --------------------------- -------------------- ----------------- --
(i) 11.1573 17.2948 13.7927 48.9880 17.5957
(ii) 6.9108 4.7246 7.2937 22.0884 NA
(iii) 6.7833 7.9424 6.8902 8.5448 18.5803
: Average computation time in second for time horizon $100$ when the attack variance is $10^4$.[]{data-label="tab:compA100"}
Proposed [@shoukry2017secure] [@chong2015observability] [@pajic2017design] [@lu2017secure]
------- ---------- ---------------------- --------------------------- -------------------- ----------------- --
(i) 0.0062 0.0108 0.0045 0.0006 0.0345
(ii) 0.0001 0.0026 0.0002 0.0003 NA
(iii) 0.0023 0.0096 0.0020 0.0005 0.0048
: Average computation time in second for time horizon $100$ when the attack variance is $10^4$.[]{data-label="tab:compA100"}
Conclusion
==========
In this paper, we propose a real-time state estimator for noisy systems that is resilient to sparse sensor integrity attacks. The proposed estimator is stable under relaxed assumptions. Its worst-case estimation errors are $O ( \log { {m}\choose{\rho} })$ in the $\mathcal H_2$ system, $O( 1 )$ in the $\mathcal H_\infty$ system, and $O( 1 )$ in the $\ell_1$ system. Its computational complexity is $O ( \log { {m}\choose{\rho} })$.
Acknowledgment {#acknowledgment .unnumbered}
==============
The authors would like to thank Professor John Doyle and Professor Richard Murray for insightful discussions.
[^1]: Y. Nakahira was with the Department of Computing and Mathematical Sciences, California Institute of Technology, Pasadena, CA, 91125 USA, e-mail: [email protected], website: http://users.cms.caltech.edu/\~ynakahir/.
[^2]: Y. Mo are with the Department of Electrical and Electronic Engineering, Nanyang Technological University, 639798 Singapore, e-mail: [email protected], website: http://yilinmo.github.io/
[^3]: Manuscript received in June 2017.
[^4]: Take the setting of [@pajic2014robustness] for example. When the system is noisy, the optimization problem $\min_{x_t \in \inr^n } \| [y(t)^T, \cdots , y( t+n-1)^T ]^T - \mathcal O x_t \|$ ($\mathcal O$ is the observability matrix) may not give correct set of compromised sensors $\{ i : \exists t, a_i (t) \neq 0\}$.
[^5]: Although abbreviate it as $e(t)$, the estimation error is also a function of disturbance $w$, attack $a$, and the estimator $\{f_t\}$.
[^6]: One way to find the matrix $K$ is via the Riccati equation, , $K^\I = P C_\I^T (C_\I PC_\I^T + D_\I D_\I^T)^{-1}$, where $P$ is unique stabilizing solution of the discrete-time algebraic Riccati equation $P = A ( P - PC_\I^T (C_\I PC_\I^T + D_\I D_\I^T)^{-1} C_\I P ) A^T + B B^T$. Sufficient conditions the existence of solution $P$ is that $(A, C_\I)$ is detectable and $(A, B B^T)$ is detectable.
[^7]: We use superscript notations for original vectors and matrices ( $K^\I$ and $x^\I$, respectively) and subscript for vectors and matrices projected by ( $y_\I$ and $C_\I$, respectively).
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: |
Let $G$ be a graph with vertex set $V(G)$ and edge set $E(G)$, and $f$ be a $0-1$ labeling of $E(G)$ so that the absolute difference in the number of edges labeled $1$ and $0$ is no more than one. Call such a labeling $f$ *edge-friendly*. The *edge-balanced index set* of the graph $G$, $EBI(G)$, is defined as the absolute difference between the number of vertices incident to more edges labeled $1$ and the number of vertices incident to more edges labeled $0$ over all edge-friendly labelings $f$. In 2009, Lee, Kong, and Wang [@LeeKongWang] found the $EBI(K_{l,n})$ for $l=1,2,3,4,5$ as well as $l=n$. We continue the investigation of the $EBI$ of complete bipartite graphs of other orders.\
2000 Mathematics Subject Classification: 05C78, 05C25\
Keywords: Edge-labeling, partial-coloring, edge-friendly labeling, friendly labeling, cordiality, friendly index set, edge-balance index set.
author:
- 'Elliot Krop [^1]'
- 'Keli Sikes [^2]'
title: |
On the Edge-balanced Index Sets\
of Complete Bipartite Graphs
---
Introduction
============
Definitions
-----------
For basic graph theoretic notation and definition see Diestel [@Diest]. The complete bipartite graph with $n$ and $m$ vertices partitioned into the first and second *parts*, respectively, where no pair of vertices in the same part are adjacent and all other pairs are adjacent, is denoted $K_{n,m}$. The *neighborhood* of a vertex $v$, denoted $N(v)$, is the set of all vertices adjacent to $v$. A *labeling* of a graph $G$ with $H \subseteq G$, is a function $f : H \rightarrow {\mathbb{Z}}_2$, where ${\mathbb{Z}}_2 = \{ 0,1 \}$. If $H=E(G)$ $(H=V(G))$ and $f$ is surjective, call the labeling an *edge labeling* (*vertex labeling*). Denote by $f(e)$ the *label* on edge $e$, and let $e_f(i) = \mbox{card}\{ e \in E : f(e)=i \}$. For a vertex $v$, let $N_0(v) = \{ u \in V: f(uv)=0 \}$ and $N_1(v) = \{ u \in V: f(uv)=1 \}$. A labeling $f$ is said to be *edge-friendly* if $|e_f(0)-e_f(1)| \leq 1$. An edge-friendly labeling $f : E \rightarrow {\mathbb{Z}}_2$ induces a partial vertex labeling $f^+ : V \rightarrow {\mathbb{Z}}_2$ defined by $f^+(v) = 0$ if the number of $0$-edges incident to $v$ is more than the number of $1$-edges incident to $v$ and $f^+(v) = 1$ if the number of $1$-edges incident to $v$ is more than the number of $0$-edges. If the number of $1$-edges incident to $v$ is equal to the number of $0$-edges then $f^+(v)$ is not defined, and we say that $v$ is *unlabeled*. For $i\in {\mathbb{Z}}_2$ let $v_f(i) = \mbox{card}\{v \in V : f^+(v) = i\}$. The *edge-balanced index set* of the graph $G$, $EBI(G)$, is defined as $\{ |v_f(0) - v_f(1)| : \mbox{the edge labeling $f$ is edge-friendly} \}$.
History and Motivation
----------------------
The study of balanced vertex labelings was introduced by S.M. Lee, A. Liu, and S.K. Tan [@LeeLiuTan] in 1992. Three years later, M. Kong and S.M. Lee [@KongLeeEBI] continued work in balanced edge labelings. In the second labeling problem, the authors attempted to classify the $EBI(G)$, which proved to be quite difficult. For more information, J. Gallian’s dynamic survey of graph labeling problems in the *Electronic Journal of Combinatorics* [@Gallian] provides an excellent overview of the subject.
There have been many attempts to classify $EBI(G)$ by looking at different families of graphs. A common problem in attempting to find the $EBI(G)$ for a particular graph family is in the search for the maximal element in the set. Often after such an element is found, the authors produce an algorithm for the lesser terms.
For complete bipartite graphs, Lee, Kong, and Wang [@LeeKongWang] found the $EBI(K_{l,n})$ for $l=1,2,3,4,5$ as well as $l=n$. In particular, they found $EBI(K_{n,n})=\{0,1,2,3,\dots, 2n-8\}$ for even $n$ and {0,2,4,…, 2n-4} for odd $n$. Little more is known for other graph families, as the edge-balance index set of regular graphs has been found for $2$-regular graphs, however, even for cubic graphs it is only known in some cases (Möbius Ladders). We continue this inquiry by attempting to determine the $EBI$ of unknown cases of $K_{n,m}$, following the methods of [@KropLeeRaridan].
Complete bipartite graphs with two odd parts
============================================
An Example
----------
We consider some special cases of the form $K_{n,n-2}$ where $n$ is small.
$EBI(K_{3,5})=\{0,2\}$
Parts differing by two
----------------------
We begin by showing our general method in the simplest case.
\[simplecase\] $EBI(K_{n,n-2})=\{0,2, \dots, 2n-10, 2n-8\}$ for odd $n>5$
We label edges of $G=K_{n,n-2}$ for maximum $EBI$. Vertices will be of two types, those which are incident to many more $0$-edges than $1$-edges, called *dense*, and those which are incident to marginally more $1$-edges than $0$-edges, called *sparse*. To find the maximum element of EBI, our goal is to minimize the number of dense vertices and maximize the number of sparse vertices. Call the vertices of the first part $A$ of $G$, $\{u, u_1, \dots, u_{n-3}\}$ and those of the second part $B$, $\{v,v', v_1, \dots, v_{n-2}\}$. Set $D=\{u,v,v'\}$ and $S=V(G)-D$.
The vertices $u,v,v'$ will be dense, so that all incident edges will be labeled zero, the rest of the vertices will be sparse. We will make half the sparse vertices in the larger part less sparse by one edge labeled $1$.
Define a cyclic order on the vertices in $S$ so that for positive $i < n-3$ and $u_i$, the *succeeding* vertex, $s(u_i)$, is $u_{i+1}$ and for $i=n-3$ the *succeeding* vertex is $u_1$. For positive $i < n-2$ and $v_i$, the *succeeding* vertex, $s(v_i)$, is $v_{i+1}$ and for $i=n-2$ the *succeeding* vertex is $v_1$. For positive $i \leq n-3$ and $v_i$, the *next* vertex is $u_{i}$, the *next* vertex of $v_{n-2}$ is $u_1$. For any positive integer $k$, the $next_k(v_i)=s(s(\dots s(u_i)))$ where the number of iterations of $s$ is $k$. For $i \in \{1, \dots, n-2\}$ define the set $U_i=\{next(v_i), next_1(v_i), \dots, next_{\frac{n-1}{2}-1}(v_i)\}$ so that the cardinality of $U_i$ is $\frac{n-1}{2}$. For $i \in \{1, \dots, n-2\}$ define the set $U'_i=\{next(v_i), next_1(v_i), \dots, next_{\frac{n-1}{2}}(v_i)\}$ so that the cardinality of $U'_i$ is $\frac{n+1}{2}$.
For $1\leq i \leq \frac{n-3}{2}$ label the edges in $S$ between the vertices in $U_i$ and $v_i$ by $1$ and the remaining edges incident to $v_i$ by $0$. For $\frac{n-1}{2}\leq i \leq n-2$ label the edges in $S$ between the vertices in $U'_i$ and $v_i$ by $1$ and the remaining edges incident to $v_i$ by $0$.
Notice that under this labeling, $$e(0)=2(n-3)+\frac{n-3}{2}\frac{n-5}{2}+\frac{n-1}{2}\frac{n-3}{2}$$ $$e(1)=\frac{n-3}{2}\frac{n+1}{2}+\frac{n-1}{2}\frac{n-1}{2}$$ and $e(0)-e(1)=-1$. Labeling the edges $uv$ and $uv'$ by $0$ produces the edge friendly labeling and the maximal EBI term.
To verify that larger values of EBI are impossible, we consider fewer dense vertices and count the maximum number of $1$-edges, to show that the graph cannot be edge-friendly. Suppose $G$ is labeled with $2$ dense vertices, say $u, v$. The maximum number of $0$-edges is attained by making the rest of the vertices sparse. Therefore, $$e(1)\geq \frac{n-1}{2}(n-1)$$ $$e(0)\leq n-3 + \frac{n-3}{2}(n-1)$$
The difference in the above quantities is greater than $1$ for $n\geq 5$, so any labeling with $2$ dense vertices is not edge-friendly.
To attain the smaller values of EBI we relabel the graph and switch $0$ and $1$ labels of pairs of edges incident to the same vertex.
The vertices $u,v,v', v''$ will be dense, so that all incident edges will be labeled zero, the rest of the vertices will be sparse.
Define a cyclic order on the vertices in $S$ so that for positive $i < n-3$ and $u_i$, the *succeeding* vertex, $s(u_i)$, is $u_{i+1}$ and for $i=n-3$ the *succeeding* vertex is $u_1$. For positive $i < n-3$ and $v_i$, the *succeeding* vertex, $s(v_i)$, is $v_{i+1}$ and for $i=n-3$ the *succeeding* vertex is $v_1$. For positive $i \leq n-3$ and $u_i$, the *next* vertex is $v_{i}$. For any positive integer $k$, the $next_k(u_i)=s(s(\dots s(v_i)))$ where the number of iterations of $s$ is $k$. For $i \in \{1, \dots, n-3\}$ define the set $V_i=\{next(u_i), next_1(u_i), \dots, next_{\frac{n-1}{2}}(u_i)\}$ so that the cardinality of $V_i$ is $\frac{n+1}{2}$.
For $1\leq i \leq n-3$ label the edges in $S$ between the vertices in $V_i$ and $u_i$ by $1$ and the remaining edges incident to $v_i$ by $0$.
Next we count the number of edges labeled $0$ and those labeled $1$ to verify that the labeling is edge-friendly.
$$e(0)=4(n-3)+\frac{n-7}{2}(n-3)=\frac{n+1}{2}(n-3)$$ $$e(1)=\frac{n+1}{2}(n-3)$$
To complete the labeling, let $(u,v)$ and $(u,v')$ be labeled $0$ and $(u,v'')$ be labeled $1$.
Next we switch edge-labels so that every pairwise switch decreases $|v(1)-v(0)|$ by $2$.
- $((v,u_2),(u_2,v_2)), \dots, ((v,u_{\frac{n-1}{2}}),(u_{\frac{n-1}{2}},v_{\frac{n-1}{2}}))$\
reducing the 1-degree on $v_2, \dots, v_{\frac{n-1}{2}}$ by one.
- $((v',u_1),(u_1,v_2)), ((v',u_2),(u_2,v_3)), \dots,$\
$((v',u_{\frac{n-3}{2}}),(u_{\frac{n-3}{2}},v_{\frac{n-1}{2}}))$ switching the labels on $v_2, \dots, v_{\frac{n-1}{2}}$ from $1$ to $0$.
- $((u,v_{\frac{n+1}{2}}),(v_{\frac{n+1}{2}},u_{\frac{n+1}{2}})), \dots, ((u,v_{n-4}),(v_{n-4},u_{n-4}))$ switching the labels on $u_{\frac{n+1}{2}},\dots, u_{n-4}$
Hence, we have produced the required EBI.
[ $\Box$\
]{}
General Case
------------
We use the method of the previous section to find the EBI of $K_{n,n-2a}$ for $a>1$.
$EBI(K_{n,n-2a})=\{0,2, \dots, 2n-2a-8, 2n-2a-6\}$ for $1\leq a \leq \frac{n-3}{4}$ and odd $n>5$
We follow the structure and notation of the proof of Theorem \[simplecase\]. For $a$ chosen as in the statement of the theorem and $2\leq c \leq 2a+1$, call the vertices of the first part $A$ of $G=K_{n,n-2a}$, $\{u, u_1, \dots, u_{n-2a-1}\}$ and those of the second part $B$, $\{v^1,v^2,\dots,v^c, v_1, \dots, v_{n-c}\}$. Set $D=\{u,v^1,v^2, \dots, v^c\}$ and $S=V(G)-D$. As before, the vertices of $D$ will be dense, so that all incident edges will be labeled $0$. The vertices of $S$ we will call sparse.
We verify that a smaller dense set where $c=1$ would not produce an edge-friendly labeling and hence that the EBI is no larger than $2n-2a-6$. We ignore the edge $(u,v)$, performing our edge-label count on the graph $G\backslash(u,v)$, and reintroduce it after the calculation. If $D=\{u,v\}$ with $u\in V(A)$ and $v\in V(B)$, then $$e(0)\leq n-2a-1 + \frac{n-2a-1}{2}(n-1)$$ $$e(1)\geq \frac{n-2a+1}{2}(n-1)$$ and $$e(1)-e(0) \geq 2a.$$ The edge $(u,v)$ can be labeled $0$, but nonetheless, the labeling is not edge-friendly for any $a>1$.
Define a cyclic order on the vertices in $S$ so that for positive $i < n-2a-1$ and $u_i$, the *succeeding* vertex, $s(u_i)$, is $u_{i+1}$ and for $i=n-2a-1$ the *succeeding* vertex is $u_1$. For positive $i < n-c$ and $v_i$, the *succeeding* vertex, $s(v_i)$, is $v_{i+1}$ and for $i=n-c$ the *succeeding* vertex is $v_1$. For positive $i \leq n-2a-1$ and $v_i$, the *next* vertex is $u_{i}$. For $i=n-2a-1+j$, $1 \leq j \leq 2a+1-c$, the *next* vertex of $v_i$ is $u_j$. For any positive integer $k$, the $next_k(v_i)=s(s(\dots s(u_i)))$ where the number of iterations of $s$ is $k$.
We consider the labeling where sparse vertices in $S\cap B$ are incident to exactly $\frac{n-2a+1}{2}$ $1$-edges. We define $K$ to be the maximum number of times we can relabel an $0$-edge incident to every vertex in $S\cap B$ (but not to $u$) as a $1$-edge, so that $e(1)\leq e(0)+1$. Note that if we ignore all edges between $D\cap A$ and $D\cap B$,
$$e(1)=(n-c)(\frac{n-2a+1}{2}+K)$$ $$e(0)=c(n-2a-1)+(n-c)(\frac{n-2a-1}{2}-K)$$
Notice that $K$ cannot exceed the number of available $1$-edges at each vertex in $A\cap S$, so $K\leq \frac{n-2a-3}{2}$. By applying this upper bound to the above expressions to calculate $e(1)-e(0)$, we see that there are enough $1$-edges so long as $a\leq \frac{n-3}{2}$.
If $c$ is odd, then $K$ is the maximum integer so that $N=2ac-n(c-1) +2K(n-c)\leq 1$, where $N$ is the difference $e(1)-e(0)$. Note that switching a label on an edge from $0$ to $1$ increases $N$ by $2$.
For $i \in \{1, \dots, n-c\}$ define the set $$U_i=\{next(v_i), next_1(v_i), \dots, next_{\frac{n-2a-1}{2}+K}(v_i)\}.$$ For $i \in \{1, \dots, n-c\}$ define the set $$U'_i=\{next(v_i), next_1(v_i), \dots, next_{\frac{n-2a+1}{2}+K}(v_i)\}.$$
*Step 1*: If $\left\lfloor \frac{-N}{2}\right\rfloor=0$, then continue to *Step 3*.
*Step 2*: For $1\leq i \leq \left\lfloor \frac{-N}{2}\right\rfloor$ label the edges in $S$ between vertices in $U'_i$ and $v_i$ by $1$ and the remaining edges incident to $v_i$ by $0$.
*Step 3*: For $\left\lfloor \frac{-N}{2}\right\rfloor+1 \leq i \leq n-c$ label the edges in $S$ between vertices in $U_i$ and $v_i$ by $1$ and the remaining edges incident to $v_i$ by $0$.
Concluding the labeling, we label the edges $(u,v^1), \dots, (u,v^{\frac{c-1}{2}})$ by $0$ and the edges $(u,v^{\frac{c+1}{2}}),\dots, (u,v^c)$ by $1$.
Notice that this labeling is edge-friendly by construction.
If $c$ is even, then we define $K,N,U_i, \mbox{ and } U'_i$ as above and repeat Steps 1-3. We conclude by labeling the edges of $(u,v^1), \dots, (u,v^{\frac{c}{2}})$ by $0$ and the edges $(u,v^{\frac{c+2}{2}}),\dots, (u,v^c)$ by $1$.
Again, this labeling is edge-friendly by construction.
For both the odd and even case, we justify that such a labeling is possible by counting the number of $1$-edges incident to any vertex $v_i$ in our labeling and showing that this does not exceed the number of vertices $u_j$ in the other part of $G$. Notice that it is enough to perform such a count for the case when $c=2a+1$, since the $1$-degree of $v_i$ is maximized. We must show $$n-2a-1\geq \frac{n-2a+1}{2}+a(2a+1-n)$$ which is true when $$a\leq \frac{n}{2} -\frac{3}{4}$$ and this inequality is satisfied by the range of $a$.
Thus, we have produced the following values of the EBI: $\{2n-6a-4,\dots, 2n-2a-6\}$.
To show the lesser values, we switch $0$ and $1$ labels of pairs of edges incident to the same vertex in the case where $c=2a+1$ so that every pairwise switch decreases $|v(1)-v(0)|$ by $2$. By this procedure, we switch the labels on $n-3a-2$ vertices in $B\cap S$ from $1$ to $0$.
For ease of notation we define a cyclic order on the vertices in $B\cap D$ so that for positive $i < 2a+1$ and $v^i$, the *succeeding* vertex, $s_1(v^i)$, is $v^{i+1}$ and for $i=2a+1$ the *succeeding* vertex, $s_1(v^i)$, is $v^1$. For any positive integer $k$, we define the $k^{th}$ succceding vertex as $s_k(v^i)=s_1(s_{k-1}(v^i))$.
Note that
1. $\deg_1(v_i)=\frac{n-2a+3}{2}+K \mbox{ for } 0\leq i \leq \left\lfloor \frac{-N}{2}\right\rfloor$\[a\]
2. $\deg_1(v_i)=\frac{n-2a+1}{2}+K \mbox{ for } \left\lfloor \frac{-N}{2}\right\rfloor +1 \leq i \leq n-2a-1$\[b\]
For \[a\]. each vertex in $B\cap S$ requires $K+2$ switches of labels on incident edges from $0$ to $1$ so that the label on the vertex becomes $0$. For \[b\]. each vertex in $B\cap S$ requires $K+1$ switches of labels on incident edges from $0$ to $1$ so that the label on the vertex becomes $0$.
If $n-3a-2\leq \lfloor \frac{-N}{2}\rfloor$, we switch the labels on the following edges from $0$ to $1$:
- Edges from vertices $(v^1, s_1(v^1), \dots, s_{K}(v^1))$ to $v_1$ reducing the 1-degree on $v_1$ by one.
- Edges from vertices $(s_{K+1}(v^1), \dots, s_{2K+1}(v^1))$ to $v_2$ reducing the 1-degree on $v_1$ by one.
and so on until we reach vertex $v_{n-3a-2}$.
If $n-3a-2> \lfloor \frac{-N}{2}\rfloor$, we switch the labels on the following edges from $0$ to $1$:
- Edges from vertices $(v^1, s_1(v^1), \dots, s_{K}(v^1))$ to $v_1$ reducing the 1-degree on $v_1$ by one.
- Edges from vertices $(s_{K+1}(v^1), \dots, s_{2K+1}(v^1))$ to $v_2$ reducing the 1-degree on $v_2$ by one.
<!-- -->
- Edges from vertices $(s_{(\lfloor\frac{-N}{2}\rfloor-1)(K+1)}(v^1), \dots, s_{(\lfloor\frac{-N}{2}\rfloor)(K+1)-1}(v^1))$ to $v_{\lfloor\frac{-N}{2}\rfloor}$ reducing the 1-degree on $v_{\lfloor\frac{-N}{2}\rfloor}$ by one.
- Edges from vertices $(s_{(\lfloor\frac{-N}{2}\rfloor)(K+1)}(v^1), \dots, s_{(\lfloor\frac{-N}{2}\rfloor+1)(K+1)}(v^1))$ to $v_{\lfloor\frac{-N}{2}\rfloor+1}$ reducing the 1-degree on $v_{\lfloor\frac{-N}{2}\rfloor+1}$ by one.
and continuing with sets of succeeding $K+2$ vertices of $B\cap D$ until we relabel $v_1, \dots, v_{n-3a-2}$.
Lastly, to show that such switches are possible, we calculate the number of edges incident with $B \cap D$ that we would need to switch, and show that this amount does not exceed the number of available edges. This is done by showing that $(K+2)(n-3a-2)\leq \frac{n-2a-1}{2}(2a+1)$ holds for $a\leq \frac{n-3}{4}$.
Hence, we have produced the required EBI.
[ $\Box$\
]{}
[MMM]{}
R. Diestel, *Graph Theory, Third Edition*, Springer-Verlag, Heidelberg Graduate Texts in Mathematics, Volume 173, New York, 2005
J.A. Gallian, A dynamic survey of graph labeling, *Electronic Journal of Combinatorics*, 17 (2010), \#DS6
M. Kong and S.-M. Lee, On edge-balanced graphs, *Graph Theory, Combinatorics and Algorithms*, 1 (1995), 711-722
E. Krop, S.-M. Lee, C. Raridan, On the edge-balanced index sets of product graphs, to appear in the Journal of the Indonesian Mathematical Society
S.-M. Lee, M. Kong, and Y.-C. Wang, On Edge-balance Index Sets of Some Complete k-partite Graphs, *Congressus Numerantium* 196, (2009) 71-94)
S.-M. Lee, A. Liu, S.K. Tan, On balanced graphs, *Cong. Number.* 87 (1992), pp. 59-64
[^1]: Department of Mathematics, Clayton State University, ([email protected])
[^2]: Department of Mathematics, Clayton State University, ([email protected])
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'In this paper, we focus on the task of extracting visual correspondences across videos. Given a query video clip from an action class, we aim to align it with training videos in space and time. Obtaining training data for such a fine-grained alignment task is challenging and often ambiguous. Hence, we propose a novel alignment procedure that learns such correspondence in space and time via cross video cycle-consistency. During training, given a pair of videos, we compute cycles that connect patches in a given frame in the first video by matching through frames in the second video. Cycles that connect overlapping patches together are encouraged to score higher than cycles that connect non-overlapping patches. Our experiments on the Penn Action and Pouring datasets demonstrate that the proposed method can successfully learn to correspond semantically similar patches across videos, and learns representations that are sensitive to object and action states.'
author:
- Senthil Purushwalkam
- Tian Ye
- Saurabh Gupta
- Abhinav Gupta
bibliography:
- 'biblioShort.bib'
- 'papers.bib'
title: Aligning Videos in Space and Time
---
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'The closed cone of flag vectors of Eulerian partially ordered sets is studied. It is completely determined up through rank seven. Half-Eulerian posets are defined. Certain limit posets of Billera and Hetyei are half-Eulerian; they give rise to extreme rays of the cone for Eulerian posets. A new family of linear inequalities valid for flag vectors of Eulerian posets is given.'
author:
- |
Margaret M. Bayer[^1] and Gábor Hetyei[^2]\
Department of Mathematics\
University of Kansas\
Lawrence KS 66045-2142\
title: Flag vectors of Eulerian partially ordered sets
---
Introduction
============
The study of Eulerian partially ordered sets (posets) originated with Stanley ([@Stanley-groups]). Examples of Eulerian posets are the posets of faces of regular CW spheres. These include face lattices of convex polytopes, the Bruhat order on finite Coxeter groups, and the lattices of regions of oriented matroids. (See [@Bjorner-CW] and [@BLVSWZ].)
The flag $f$-vector (or simply flag vector) of a poset is a standard parameter counting chains in the partially ordered set by ranks. In the last twenty years there has grown a body of work on numerical conditions on flag vectors of posets and complexes, especially those arising in geometric contexts. Early contributions are from Stanley on balanced Cohen-Macaulay complexes ([@Stanley-balanced]) and Bayer and Billera on the linear equations on flag vectors of Eulerian posets ([@Bayer-Billera]). A major recent contribution is the determination of the closed cone of flag vectors of all graded posets by Billera and Hetyei ([@Billera-Hetyei]). Results on flag vectors and other invariants of Eulerian posets and special classes of them are surveyed in [@Stanley-Eulerian].
Our goal has been to describe the closed cone ${\cal C}^{n+1}_{{\cal E}}$ of flag $f$-vectors of Eulerian partially ordered sets. This problem was posed explicitly in [@Billera-Liu]. The ideal description would give explicitly both the facets (i.e., crucial inequalities on flag vectors) and posets that generate the extreme rays. We have a complete solution only for rank at most seven. For arbitrary ranks we give some of the facets and extreme rays. The extreme rays of the general graded cone ([@Billera-Hetyei]) play an important role. We introduce half-Eulerian partially ordered sets in order to incorporate these limit posets in this work.
The remainder of this section provides definitions and other background, and the definition of the flag ${\mbox{$L$}}$-vector, which simplifies the calculations. Section \[half-Eul\] describes the extreme rays of the general graded cone, defines half-Eulerian posets, identifies which limit posets are half-Eulerian, and computes the corresponding $cd$-indices. Section \[sec-inequalities\] gives two general classes of inequalities on Eulerian flag vectors. Section \[sec-cone\] shows that the half-Eulerian limit posets all give extremes of the Eulerian cone, identifies some inequalities in all ranks as facet-inducing, and describes completely the cone for rank at most 7.
Background {#ss_gpo}
----------
A [*graded poset*]{} $P$ is a finite partially ordered set with a unique minimum element ${\hat{0}}$, a unique maximum element ${\hat{1}}$, and a [*rank function*]{} ${\rho}:
P\longrightarrow {\mbox{\bf N}}$ satisfying ${\rho}({\hat{0}})=0$, and ${\rho}(y)-{\rho}(x)=1$ whenever $y\in P$ covers $x\in P$. The [*rank ${\rho}(P)$ of a graded poset $P$*]{} is the rank of its maximum element. Given a graded poset $P$ of rank $n+1$ and a subset $S$ of $\{1,2,\ldots,n\}$ (which we abbreviate as $[1,n]$), define the [*$S$–rank–selected subposet of $P$*]{} to be the poset $$P_{S} {:=}\{ x \in P\::\: {\rho}(x) \in S\} \cup \{ {{\hat{0}}},{{\hat{1}}}\}.$$ Denote by $f_S (P)$ the number of maximal chains of $P_S$. Equivalently, $f_S(P)$ is the number of chains $x_1<\cdots<x_{|S|}$ in $P$ such that $\{{\rho}(x_1),\ldots,{\rho}(x_{|S|})\}=S$. The vector $\left(f_S (P)\::\: S\subseteq [1,n]\right) $ is called the [*flag $f$-vector*]{} of $P$. Whenever it does not cause confusion, we write $f_{s_1\,\ldots\, s_k}$ rather than $f_{\{s_1,\ldots,s_k\}}$; in particular, $f_{\{m\}}$ is always denoted $f_m$.
Various properties of the flag $f$-vector are more easily seen in different bases. An often used equivalent encoding is the [*flag $h$-vector*]{} given by the formula $$h_S (P){:=}\sum _{T\subseteq S} (-1)^{|S\setminus T|} f_T(P),$$ or, equivalently, $$f_S (P)= \sum _{T\subseteq S} h_T(P).$$ The [*$ab$-index*]{} $\Psi_P(a,b)$ of $P$ is a generating function for the flag $h$-vector. It is the following polynomial in the noncommuting variables $a$ and $b$: $$\label{E_ab}
\Psi_P(a,b)
=
\sum_{S \subseteq [1,n]} h_S (P) u_S,$$ where $u_S$ is the monomial $u_1 u_2\cdots u_n$ with $u_i=a$ if $i\not\in S$, and $u_i=b$ if $i\in S$.
The [*Möbius function*]{} of a graded poset $P$ is defined recursively for any subinterval of $P$ by the formula $$\mu ([x,y])=
\left\{
\begin{array}{cl}
1 & \mbox{if $x=y$},\\
-\sum_{x\le z<y} \mu([x,z]) &\mbox{otherwise}.\\
\end{array}
\right.$$ Equivalently, by Philip Hall’s theorem, the Möbius function of a graded poset $P$ of rank $n+1$ is the [*reduced Euler characteristic*]{} of the order complex, i.e., it is given by the formula $$\label{E_MEc}
\mu (P)=\sum_{S\subseteq [1,n]} (-1)^{|S|+1} f_S(P).$$ (See [@Stanley-EC Proposition 3.8.5].)
A graded poset $P$ is [*Eulerian*]{} if the Möbius function of every interval $[x,y]$ is given by $\mu([x,y]) = (-1)^{{\rho}(x,y)}$. (Here ${\rho}(x,y)={\rho}([x,y])={\rho}(y) - {\rho}(x)$.)
The first characterization of all linear equalities holding for the flag $f$-vectors of all Eulerian posets was given by Bayer and Billera in [@Bayer-Billera]. The equations of the theorem are called the generalized Dehn-Sommerville equations. Call the subspace of ${\bf R}^{2^n}$ they determine the [*Eulerian subspace*]{}; its dimension is the Fibonacci number $e_n$ ($e_0=e_1=1$, $e_n=e_{n-1}+e_{n-2}$).
\[T\_BB\] Every linear equality holding for the flag $f$-vector of all Eulerian posets of rank $n+1$ is a consequence of the equalities $$\left((-1)^{i-1}+(-1)^{k+1}\right)f_S+\sum _{j=i}^k (-1)^j
f_{S\cup\{j\}}=0$$ for $S\subseteq [1,n]$ and $[i,k]$ a maximal interval of $[1,n]\setminus S$.
Fine discovered that the $ab$-index of a polytope can be written as a polynomial in the noncommuting variables $c{:=}a+b$ and $d{:=}ab + ba$. Bayer and Klapper [@Bayer-Klapper] proved that for a graded poset $P$, the equations of Theorem \[T\_BB\] hold if and only if the $ab$-index is a polynomial with integer coefficients in $c$ and $d$. This polynomial is called the [*$cd$-index*]{} of $P$. Stanley ([@Stanley-flag]) gives an explicit recursion for the $cd$-index in terms of intervals of $P$ for Eulerian posets. (He thus gives another proof of the existence of the $cd$-index for Eulerian posets.)
The flag [$\ell$]{}-vector and the flag -vector {#s_ell}
-----------------------------------------------
The introduction of another vector equivalent to the flag $f$-vector simplifies calculations.
\[D\_lv\] [ *The [*flag ${\ensuremath{\ell}}$-vector*]{} of a graded partially ordered set $P$ of rank $n+1$ is the vector $({\ensuremath{\ell}}_S(P) \:: \: S\subseteq [1,n])$, where $${\ensuremath{\ell}}_S(P){:=}(-1)^{n-|S|} \sum_{T\supseteq [1,n]\setminus S}
(-1)^{|T|} f_T(P).$$* ]{}
As a consequence, $$\label{E_fl}
f_S(P)=\sum _{T\subseteq [1,n]\setminus S} {\ensuremath{\ell}}_T(P).$$ The flag ${\ensuremath{\ell}}$-vector was first considered by Billera and Hetyei ([@Billera-Hetyei]) while describing all linear inequalities holding for the flag $f$-vectors of all graded partially ordered sets. It turned out to give a sparse representation of the cone of flag $f$-vectors described in that paper.
A variant significant for Eulerian posets is the flag ${\mbox{$L$}}$-vector.
\[D\_LV\] [*The [*flag ${\mbox{$L$}}$-vector*]{} of a graded partially ordered set $P$ of rank $n+1$ is the vector $({\mbox{$L$}}_S(P) \::\: S\subseteq [1,n])$, where $${\mbox{$L$}}_S (P){:=}(-1)^{n-|S|}
\sum_{T\supseteq [1,n]\setminus S} \left(-{1\over 2}\right)^{|T|}
f_T(P).$$* ]{}
Inverting the relation of the definition gives $$f_S(P)=2^{|S|}\sum _{T\subseteq [1,n]\setminus S} {\mbox{$L$}}_T(P).$$
When the poset $P$ is Eulerian, the parameters ${\mbox{$L$}}_S (P)$ are actually the coefficients of the [*$ce$-index*]{} of the poset $P$. The $ce$-index was introduced by Stanley ([@Stanley-flag]) as an alternative way of viewing the $cd$-index. The letter $c$ continues to stand for $a+b$; now let $e{:=}a-b$. The $ab$-index of a poset can be written in terms of $c$ and $d$ if and only if it can be written in terms of $c$ and $ee$. It is easy to verify that ${\mbox{$L$}}_S (P)$ is exactly the coefficient in the of $P$ of the word $u_S=u_1u_2\cdots u_n$ where $u_i=c$ if $i\not\in S$, and $u_i=e$ if $i\in S$. Since the existence of the $cd$-index is equivalent to the validity of the generalized Dehn-Sommerville equations, we get the following proposition. (It can be proved directly from the definition of the flag ${\mbox{$L$}}$-vector, yielding an alternative way to prove the existence of the $cd$-index for Eulerian posets.) A subset $S\subseteq [1,n]$ is [*even*]{} if all the maximal intervals contained in $S$ are of even length.
\[P\_BBl\] The generalized Dehn-Sommerville relations hold for a poset $P$ if and only if ${\mbox{$L$}}_S(P)=0$ whenever $S$ is not an even set.
The generalized Dehn-Sommerville relations hold (by chance) for some nonEulerian posets. A poset is Eulerian, however, if these relations hold for all intervals of the poset.
\[P\_EulerLV\] A graded partially ordered set is Eulerian if and only if ${\mbox{$L$}}_S ([x,y])=0$ for every interval $[x,y]\subseteq P$ and every subset $S$ of that is not an even set.
Half-Eulerian posets {#half-Eul}
====================
In this section we find special points in the closed cone of flag vectors of Eulerian posets. First consider the extremes of the closed cone of flag vectors of all graded posets, found by Billera and Hetyei ([@Billera-Hetyei]).
*Given a graded poset $P$ of rank $n+1$, an interval $I\subseteq [1,n]$, and a positive integer $k$, ${\mbox{$D$}}_I^k(P)$ is the graded poset obtained from $P$ by replacing every $x\in P$ with rank in $I$ by $k$ elements $x_1,\ldots,x_k$ and by imposing the following relations.*
1. If for $x,y\in P$, ${\rho}(x)\in I$ and ${\rho}(y)\not\in I$, then $x_i<y$ in ${\mbox{$D$}}_I^k(P)$ if and only if $x<y$ in $P$, and $y<x_i$ in ${\mbox{$D$}}_I^k(P)$ if and only if $y<x$ in $P$.
2. If $\{{\rho}(x),{\rho}(y)\}\subseteq I$, then $x_i<y_j$ in ${\mbox{$D$}}_I^k(P)$ if and only if $i=j$ and $x<y$ in $P$.
Clearly ${\mbox{$D$}}_I^k P$ is a graded poset of the same rank as $P$. Its flag $f$-vector can be computed from that of $P$ in a straightforward manner.
An [*interval system on $[1,n]$*]{} is any set of subintervals of $[1,n]$ that form an antichain (that is, no interval is contained in another). (Much of what follows holds even if the intervals do not form an antichain, but the assumption simplifies the statements of some theorems.) For any interval system ${\ensuremath{{\cal I}}}$ on $[1,n]$, and any positive integer $N$, the poset $P(n,{\ensuremath{{\cal I}}},N)$ is defined to be the poset obtained from a chain of rank $n+1$ by applying ${\mbox{$D$}}_I^N$ for all $I\in{\ensuremath{{\cal I}}}$. It does not matter in which order these operators are applied. (Different values of $N$ can be used for each interval $I$, but we do not need that generality here.) Consider the sequence of posets for a fixed interval system ${\ensuremath{{\cal I}}}$ as $N$ goes to infinity. Billera and Hetyei ([@Billera-Hetyei]) showed that the normalized flag vectors of such a sequence converge to a vector on an extreme ray of the cone of flag vectors of all graded posets. More precisely,
\[Billera-Hetyei theorem\] Suppose ${\ensuremath{{\cal I}}}$ is an interval system of $k$ intervals on $[1,n]$. Then the vector $$\left(\lim_{N\rightarrow \infty} {1\over N^k}
f_S (P(n,{\ensuremath{{\cal I}}}, N)):S\subseteq [1,n]\right)$$ generates an extreme ray of the cone of flag vectors of all graded posets. Moreover, all extreme rays are generated in this way.
Unfortunately, none of the posets $P(n,{\ensuremath{{\cal I}}},N)$ are Eulerian, and none of these extreme rays are contained in the closed cone of flag vectors of Eulerian posets. However some of the posets are “half-Eulerian”, and lead us to extreme rays of the Eulerian cone.
For the interval system ${\cal I}=\{[1,1],[2,2],\ldots,[n,n]\}$, abbreviate ${\mbox{$D$}}_{\cal I}^2(P)$ as ${\mbox{$D$}}P$, and call this the [*horizontal double*]{} of $P$. Thus the horizontal double of $P$ is the poset obtained from $P$ by replacing every $x\in P\setminus \{{\hat{0}},{\hat{1}}\}$ with two elements $x_1,x_2$ such that ${\hat{0}}$ and ${\hat{1}}$ remain the minimum and maximum elements of the partially ordered set, and $x_i<y_j$ if and only if $x<y$ in $P$. (In the Hasse diagram of $P$, every edge is replaced by $\Join$.)
[*A [*half-Eulerian poset*]{} is a graded partially ordered set whose horizontal double is Eulerian.* ]{}
For more information on half-Eulerian posets, see [@Bayer-Hetyei].
The flag $f$-vectors of $P$ and its horizontal double are connected by the formula $f_S ({\mbox{$D$}}P)=2^{|S|} f_S (P)$. Thus, $$\label{elltoL}
{\mbox{$L$}}_S({\mbox{$D$}}P)= {\ensuremath{\ell}}_S(P).$$
Applying the definition of Eulerian to the horizontal double of a poset we get
A graded partially ordered set $P$ is half-Eulerian if and only if for every interval $[x,y]$ of $P$, $$\sum_{i=1}^{{\rho}(x,y)-1} (-1)^{i-1} f_i ([x,y])=
(1+(-1)^{{\rho}(x,y)})/2.$$
Corollary \[P\_EulerLV\] can now be restated for half-Eulerian posets.
\[D\_hE\] A graded partially ordered set is half-Eulerian if and only if ${\ensuremath{\ell}}_S ([x,y])=0$ for every interval $[x,y]\subseteq P$ and every subset $S$ of that is not an even set.
The flag vectors of the horizontal doubles of half-Eulerian posets span the Eulerian subspace, the subspace defined by the generalized Dehn-Sommerville equations. But the cones they determine may be different. Recall ${\cal C}^{n+1}_{\cal E}$ is the closed cone of flag vectors of Eulerian posets. Now write ${\cal C}^{n+1}_{\cal D}$ for the closed cone of flag vectors of horizontal doubles of half-Eulerian posets. We do not know if the inclusion ${\cal C}^{n+1}_{\cal D}\subseteq {\cal C}^{n+1}_{\cal E}$ is actually equality.
For which interval systems ${\ensuremath{{\cal I}}}$ is $P(n,{\ensuremath{{\cal I}}},N)$ half-Eulerian?
[ *An interval system ${\ensuremath{{\cal I}}}$ on $[1,n]$ is [*even*]{} if for every pair of intervals $I,J\in {\ensuremath{{\cal I}}}$ the intersection $I\cap J$ has an even number of elements. (In particular, $|I|$ must be even for every $I\in {\ensuremath{{\cal I}}}$.)* ]{}
Our goal is to show that the posets $P(n,{\ensuremath{{\cal I}}},N)$ are half-Eulerian if and only if ${\ensuremath{{\cal I}}}$ is an even interval system. For this we need to understand the intervals of the posets $P(n,{\ensuremath{{\cal I}}},N)$.
\[P\_limitint\] The interval $[x,y]\subseteq P(n,{\ensuremath{{\cal I}}},N)$ is isomorphic to\
$P({\rho}(x,y)-1,{\cal J},N)$, where ${\cal J}=\{I-{\rho}(x)\::\:
I\in {\ensuremath{{\cal I}}},\, I\subseteq [{\rho}(x)+1, {\rho}(y)-1] \}
$.
Let ${\rho}(x)=r$ and ${\rho}(y)=s$. Construct $P(n,{\ensuremath{{\cal I}}},N)$ by applying the operators ${\mbox{$D$}}_I^N$ for all $I\in{\ensuremath{{\cal I}}}$ to a chain. Since the order of applying these operators is arbitrary, we may choose to apply first those for which $I$ is not a subset of $[r+1,s-1]$. At this point for every $x'$ of rank $r$ and $y'$ of rank $s$ with $y'\ge x'$, the interval $[x',y']$ is isomorphic to a chain of rank ${\rho}(x',y')$. Applying the remaining operators ${\mbox{$D$}}_I^N$ leaves the elements of rank at most $r$ or of rank at least $s$ unchanged, and has the same effect on $[x',y']$ as applying the operators ${\mbox{$D$}}_{I-r}^N$ to a chain of rank ${\rho}(x',y')$.
The effect on the flag $f$-vector of applying the operator ${\mbox{$D$}}_I^N$ to a poset of rank $n+1$ is given by the formula $$\label{E_HDINf}
f_S ({\mbox{$D$}}_I^N(P))
=\left\{
\begin{array}{ll}
N f_S(P) & \mbox{if $I\cap S\neq \emptyset$,}\\
f_S(P) &\mbox{otherwise.}\\
\end{array}
\right.$$ This enables us to write an [$\ell$]{}-vector formula.
\[E\_HDINl\] For $P$ a graded poset of rank $n+1$, $S\subseteq [1,n]$, and $N$ a positive integer, $${\ensuremath{\ell}}_S ({\mbox{$D$}}_I^N(P))
=N {\ensuremath{\ell}}_S(P)-(N-1)\sum_{T\cup I=S} {\ensuremath{\ell}}_T(P).$$
From the definition of $\ell_S$ and equation (\[E\_HDINf\]), $$\begin{aligned}
{\ensuremath{\ell}}_S({\mbox{$D$}}_I^N(P))
&=&(-1)^{n-|S|} \sum_{R\supseteq [1,n]\setminus S}
(-1)^{|R|} f_R ({\mbox{$D$}}_I^N(P))\\
&=&(-1)^{n-|S|} \sum_{R\supseteq [1,n]\setminus S}
\!\!\! (-1)^{|R|} N f_R (P) \\
& &
\mbox{} -(-1)^{n-|S|} \sum_{R\supseteq [1,n]\setminus S \atop
R\subseteq [1,n]\setminus I}
\!\!\!(-1)^{|R|} (N-1) f_R (P)\\
&=& N {\ensuremath{\ell}}_S(P)
-(-1)^{n-|S|} \sum_{R\supseteq [1,n]\setminus S \atop
R\subseteq [1,n]\setminus I}
(-1)^{|R|} (N-1) f_R (P)\\\end{aligned}$$ By (\[E\_fl\]), the coefficient in $-(-1)^{n-|S|} \sum_{R\supseteq [1,n]\setminus S \atop
R\subseteq [1,n]\setminus I} (-1)^{|R|} (N-1) f_R (P)$ of ${\ensuremath{\ell}}_T(P)$ is $$-(N-1)(-1)^{n-|S|} \sum_{R\supseteq [1,n]\setminus S \atop
R\subseteq [1,n]\setminus (T\cup I)}
(-1)^{|R|}\, ,$$ which is an empty sum if $(T\cup I)$ is not contained in $S$, zero if $(T\cup I)$ is properly contained in $S$, and $-(N-1)(-1)^{n-|S|}(-1)^{|[1,n]\setminus S|}=-(N-1)$ if $(T\cup I)=S$. This gives the recursion of the lemma.
From this we can determine which of the posets $P(n,{\ensuremath{{\cal I}}}, N)$ are half-Eulerian.
\[P\_limitHE\] Let ${\ensuremath{{\cal I}}}$ be an interval system on $[1,n]$.
1. If ${\ensuremath{{\cal I}}}$ is an even system of intervals, then for all $N$ the partially ordered set $P(n,{\ensuremath{{\cal I}}},N)$ is half-Eulerian.
2. If for some $N>1$, $P(n,{\ensuremath{{\cal I}}},N)$ is half-Eulerian, then ${\ensuremath{{\cal I}}}$ is an even system of intervals.
Using Lemma \[E\_HDINl\] we can show by induction on $|{\ensuremath{{\cal I}}}|$ that for every $N$, ${\ensuremath{\ell}}^{n+1}_S\left(P(n,{\ensuremath{{\cal I}}}, N)\right)$ is zero unless $S$ is the union of some intervals of ${\ensuremath{{\cal I}}}$. In particular, if ${\ensuremath{{\cal I}}}$ is an even system of intervals, then ${\ensuremath{\ell}}_S\left(P(n,{\ensuremath{{\cal I}}}, N)\right)=0$ whenever $S$ is not an even set. The same observation holds for every interval $[x,y]\subseteq P(n,{\ensuremath{{\cal I}}}, N)$ as well, since by Proposition \[P\_limitint\] $[x,y]$ is isomorphic to $P(m,{\cal J}, N)$ for some $m\leq n$ and some even system of intervals $\cal J$. Therefore the conditions of Proposition \[D\_hE\] are satisfied by $P(n,{\ensuremath{{\cal I}}}, N)$ for every $N$, if ${\ensuremath{{\cal I}}}$ is an even system of intervals.
Now assume ${\ensuremath{{\cal I}}}$ is a system of intervals that is not even. First consider the case where ${\ensuremath{{\cal I}}}$ contains an interval $I_m=[a,b]$ with $b-a$ even (hence $I_m$ is odd). Let ${\cal J}=\{I_m-a+1\}=\{[1,b-a+1]\}$. For $S$ nonempty, $f_S(P(b-a+1,{\cal J},N))=N$, so $$\begin{aligned}
\lefteqn{{\ensuremath{\ell}}_{[1,b-a+1]}(P(b-a+1,{\cal J},N))}\hspace*{.5in}\\
&=& \sum_{T\subseteq[1,b-a+1]} (-1)^{|T|}f_T(P(b-a+1,{\cal J},N))\\
&=& 1+\sum_{{T\subseteq[1,b-a+1]}\atop{T\ne\emptyset}} (-1)^{|T|}N=1-N.\end{aligned}$$ So ${\ensuremath{\ell}}_{[1,b-a+1]}(P(b-a+1,{\cal J},N))\ne 0$ for $N>1$. Fix $N>1$, and choose $x$ and $y$ in $P(n,{\ensuremath{{\cal I}}},N)$ with ${\rho}(x)=a-1$, ${\rho}(y)=b+1$, and $x\le y$. Then by Proposition \[P\_limitint\], ${\ensuremath{\ell}}_{[1,{\rho}(x,y)-1]}([x,y])
={\ensuremath{\ell}}_{[1,b-a+1]}(P(b-a+1,{\cal J},N))\ne 0$, with $|[1,b-a+1]|$ odd. So $P(n,{\ensuremath{{\cal I}}},N)$ is not half-Eulerian.
Now suppose ${\ensuremath{{\cal I}}}$ contains only even intervals, but some two intervals have an odd overlap. Let $I_p=[a,d]$ and $I_q=[c,b]$, where $a<c\le d<b$ and $d-a$ and $b-c$ are odd, but $d-c$ is even. Then $b-a$ is also even. We show that we may assume no other interval of ${\ensuremath{{\cal I}}}$ is in the union $I_p\cup I_q$. Suppose $I_r=[e,f]$ is another interval of ${\ensuremath{{\cal I}}}$ with $[e,f]\subset [a,b]$ (and $f-e$ is odd). Since ${\ensuremath{{\cal I}}}$ is an antichain, $a<e<c\le d<f<b$. If $e-a$ is even, then $|I_q\cap I_r|=|[c,f]|=f-c+1 = (f-e)+(e-a)-(d-a)+(d-c)+1$, which is odd, because it is the sum of three odds and two evens. If $e-a$ is odd, then $|I_p\cap I_r|=|[e,d]|=d-e+1=(d-a)-(e-a)+1$, which is odd because it is the sum of three odds. Thus, if two intervals of ${\ensuremath{{\cal I}}}$ have odd intersection and their union contains a third interval of ${\ensuremath{{\cal I}}}$, then two intervals of ${\ensuremath{{\cal I}}}$ with smaller union have odd intersection.
So we may assume $I_p=[a,d]$ and $I_q=[c,b]$ have odd intersection, and their union $[a,b]$ contains no other interval of ${\ensuremath{{\cal I}}}$. Let ${\cal J}
=\{I_p-a+1, \linebreak I_q-a+1\}
=\{[1,d-a+1],[c-a+1,b-a+1]\}
$. Then $$\begin{aligned}
\lefteqn{
f_S(P(b-a+1,{\cal J},N))}\hspace*{.5in}\\
&=&\left\{\begin{array}{ll}
1 & \mbox{if $S=\emptyset$}\\
N^2 & \mbox{if $S\cap(I_p-a+1)\ne\emptyset$ and
$S\cap(I_q-a+1)\ne\emptyset$}\\
N & \mbox{otherwise.}
\end{array} \right. \end{aligned}$$ So $$\begin{aligned}
\lefteqn{
{\ensuremath{\ell}}_{[1,b-a+1]}(P(b-a+1,{\cal J},N))}\hspace*{.25in}\\
&=&
\sum_{T\subseteq[1,b-a+1]} (-1)^{|T|}f_T(P(b-a+1,{\cal J},N))\\
&=& \sum_{T\subseteq[1,b-a+1]} (-1)^{|T|}N^2
+ \sum_{T\subseteq[1,c-a]} (-1)^{|T|}(N-N^2)\\
&+& \sum_{T\subseteq[d-a+2,b-a+1]} (-1)^{|T|}(N-N^2)
+ (1-2N+N^2) = (1-N)^2. \end{aligned}$$ So ${\ensuremath{\ell}}_{[1,b-a+1]}(P(b-a+1,{\cal J},N))\ne 0$ for $N>1$. Fix $N>1$, and choose $x$ and $y$ in $P(n,{\ensuremath{{\cal I}}},N)$ with ${\rho}(x)=a-1$, ${\rho}(y)=b+1$, and $x\le y$. Then by Proposition \[P\_limitint\], ${\ensuremath{\ell}}_{[1,{\rho}(x,y)-1]}([x,y])
={\ensuremath{\ell}}_{[1,b-a+1]}(P(b-a+1,{\cal J},N))\ne 0$, with $|[1,b-a+1]|$ odd. So $P(n,{\ensuremath{{\cal I}}},N)$ is not half-Eulerian.
As will be seen later, even interval systems give rise to extreme rays of the cone of flag vectors of Eulerian posets. It is of interest, therefore, to count them.
The number of even interval systems on $[1,n]$ is ${n \choose \lfloor n/2\rfloor}$.
We define a one-to-one correspondence between even interval systems on $[1,n]$ and sequences $\lambda=(\lambda_1,\lambda_2,\ldots, \lambda_n)\in \{-1,1\}^n$ satisfying $\sum_i \lambda_i = 0$ if $n$ is even and $\sum_i \lambda_i = 1$ if $n$ is odd. Clearly there are ${n \choose \lfloor n/2\rfloor}$ such sequences.
For ${\ensuremath{{\cal I}}}$ an even interval system, define $\lambda({\ensuremath{{\cal I}}})=(\lambda_1,\lambda_2,\ldots, \lambda_n)\in \{-1,1\}^n$, where $\lambda_i = (-1)^i$ if $i$ is an endpoint of an interval of ${\ensuremath{{\cal I}}}$, and $\lambda_i = (-1)^{i-1}$ otherwise. (Note that for an even interval system, no number can be an endpoint of more than one interval.) For [${\cal I}$]{} an even interval system, summing $(-1)^i$ over the endpoints of intervals gives 0. So $$\begin{aligned}
\sum_{i=1}^n \lambda_i &=& \sum_{i=1}^n (-1)^{i-1} +
\sum_{{\mbox{{\scriptsize $i$ endpoint}}}\atop
{\mbox{{\scriptsize of interval}}}}
2(-1)^i\\
&=&
\sum_{i=1}^n(-1)^{i-1}
= \left\{\begin{array}{rl}
0 & \mbox{if $n$ is even}\\
1 & \mbox{if $n$ is odd} \end{array}\right. .\end{aligned}$$
On the other hand, given a sequence $\lambda=(\lambda_1,\lambda_2,\ldots, \lambda_n)\in \{-1,1\}^n$ satisfying $\sum_i \lambda_i = 0$ if $n$ is even and $\sum_i \lambda_i = 1$ if $n$ is odd, construct an even interval system as follows. Let $s_1 < s_2 < \cdots < s_k$ be the sequence of indices $s$ for which $\lambda_s = (-1)^s$. Then $ \sum_{i=1}^n (-1)^{i-1} =\sum_{i=1}^n \lambda_i
=\sum_{i=1}^n (-1)^{i-1} + \sum_{j=1}^k 2(-1)^{s_j}$, so $ \sum_{j=1}^k (-1)^{s_j}=0$. Thus the sequence of $s_j$’s contains the same number of even numbers as odd. Construct an interval system ${\ensuremath{{\cal I}}}= \{[a_1,b_1], [a_2,b_2], \ldots, [a_m,b_m]\}$ ($2m=k$) recursively as follows. Let $a_1=s_1$ and let $b_1=s_j$ where $j$ is the least index such that $s_1$ and $s_j$ are of opposite parity. Then ${\ensuremath{{\cal I}}}= {[a_1,b_1]}\cup {\ensuremath{{\cal I}}}'$, where ${\ensuremath{{\cal I}}}'$ is the interval system associated with $s_2 < s_3 < s_4 < \cdots < s_k$ with $b_1=s_j$ removed. Clearly $[a_1,b_1]$ is of even length. If $[a_1,b_1]\cap [a_i,b_i]\ne \emptyset$ for some interval $[a_i,b_i]$ of ${\ensuremath{{\cal I}}}'$, then $a_i<b_1$, so by the choice of $b_1$, $a_i$ has the same parity as $a_1$. Thus $[a_1,b_1]\cap [a_i,b_i]=[a_i,b_1]$ is of even length. Furthermore, $b_i$ and $b_1$ are of the same parity, since $a_i$ and $a_1$ are, so again by the choice of $b_1$, $b_i>b_1$. So the interval $[a_i,b_i]$ is not contained in the interval $[a_1,b_1]$. The interval system $\{[a_m,b_m]\}$, is even, so by induction ${\ensuremath{{\cal I}}}$ is an even interval system.
These constructions are inverses, giving the desired bijection.
Recall that Billera and Hetyei ([@Billera-Hetyei]) found extremes of the cone of flag vectors of graded posets as limits of the normalized flag vectors of the posets $P(n,{\ensuremath{{\cal I}}},N)$. The next proposition follows easily by induction from Lemma \[E\_HDINl\].
\[P\_limitl\] Let ${\ensuremath{{\cal I}}}=\{I_1,I_2,\ldots,I_k\}$ be a system of $k\geq 0$ intervals on $[1,n]$. Then $$\begin{aligned}
\lefteqn{
\lim_{N\longrightarrow \infty} {1\over N^k} {\ensuremath{\ell}}_S
\left(P(n,{\ensuremath{{\cal I}}}, N)\right)} \hspace*{.5in}\\
&=&\sum _{j=0}^k (-1)^j
\left|\left\{1\leq i_1<\cdots <i_j\leq k \::\:
I_{i_1}\cup \cdots \cup I_{i_j}=S
\right\}\right|. \end{aligned}$$
Write $f_S(P(n,{\ensuremath{{\cal I}}}))=
\lim_{N\rightarrow\infty} f_S(P(n,{\ensuremath{{\cal I}}},N))/N^{|{\ensuremath{{\cal I}}}|}$. The vector these form (as $S$ ranges over all subsets of $[1,n]$) is not the flag $f$-vector of an actual poset, but it is in the closed cone of flag $f$-vectors of all graded posets. We call the symbol $P(n,{\ensuremath{{\cal I}}})$ a “limit poset” and refer to the flag vector of the limit poset. If ${\ensuremath{{\cal I}}}$ is an even interval system, then $(f_S(P(n,{\ensuremath{{\cal I}}})): S\subseteq [1,n])$ is in the closed cone of flag vectors of half-Eulerian posets. To get Eulerian posets the horizontal double operator is applied to $P(n,{\ensuremath{{\cal I}}},N)$. The vector $(f_S({\mbox{$D$}}P(n,{\ensuremath{{\cal I}}})): S\subseteq [1,n])$ is defined as a limit of the resulting normalized flag $f$-vectors, and satisfies $f_S({\mbox{$D$}}P(n,{\ensuremath{{\cal I}}}))=2^{|S|}f_S(P(n,{\ensuremath{{\cal I}}}))$. It lies in the cone ${\cal C}^{n+1}_{{\cal D}}$ of flag vectors of doubles of half-Eulerian posets, a subcone of the Eulerian cone.
Recall (equation (\[elltoL\])) that the [$\ell$]{}-vector of a poset $P$ equals the -vector of its horizontal double ${\mbox{$D$}}P$. The same holds after passing to the limit posets. Thus, Proposition \[P\_limitl\] gives $$L_S({\mbox{$D$}}P(n,{\cal I}))
=\sum _{j=0}^k (-1)^j
\left|\left\{1\leq i_1<\cdots <i_j\leq k :\:
I_{i_1}\cup \cdots \cup I_{i_j}=S
\right\}\right|,$$ where ${\ensuremath{{\cal I}}}=\{I_1,I_2,\ldots,I_k\}$.
We look at the associated $cd$-indices of the “doubled limit posets.” Think of a word in $c$ and $d$ as a string with each $c$ occupying one position and each $d$ occupying two positions. The [*weight*]{} of a $cd$-word $w$ is then the number of positions of the string. Associated to each $cd$-word $w$ is the even set $S(w)$ consisting of the positions occupied by the $d$’s.
\[cd\] For each $cd$-word $w$ with $k$ $d$’s and weight $n$, there exists an even interval system ${\ensuremath{{\cal I}}}_w$ for which the $cd$-index of ${\mbox{$D$}}P(n,{\ensuremath{{\cal I}}}_w)$ is $2^k w$.
Fix a $cd$-word $w$ with $k$ $d$’s and weight $n$. Write the elements of $S(w)$ in increasing order as $i_1$, $i_1+1$, $i_2$, $i_2+1$, …, $i_k$, $i_k+1$, and let ${\ensuremath{{\cal I}}}_w$ be the interval system $\{[i_1, i_1+1], [i_2, i_2+1], \ldots, [i_k, i_k+1]\}$. Let $\Phi=2^k w$. Rewrite the $cd$-polynomial $\Phi$ as a $ce$-polynomial. Recall from Sections \[ss\_gpo\] and \[s\_ell\] that $c=a+b$, $d=ab+ba$, and $e=a-b$, so $d = (cc-ee)/2$. Thus, $\Phi$ is rewritten as a sum of $2^k$ terms. Each is the result of replacing some subset of the $d$’s by $cc$, and the rest by $ee$; the coefficient is $\pm 1$, depending on whether the number of $d$’s replaced by $ee$ is even or odd. Thus $$2^k w = \sum_{J\subseteq [1,k]} (-1)^J w_J,$$ where $w_J=w_1w_2\cdots w_n$, with $w_{i_j}=w_{i_j+1}=e$ if $j\in J$ and the remaining $w_i$’s are $c$. By the -vector version of Proposition \[P\_limitl\], this is precisely the $ce$-index of ${\mbox{$D$}}P(n,{\ensuremath{{\cal I}}}_w)$.
In [@Stanley-flag] Stanley first found for each $cd$-word $w$ a sequence of Eulerian posets whose normalized $cd$-indices converge to $w$. Our limit posets are closely related to Stanley’s, but this particular construction highlights the important link between the half-Eulerian and Eulerian cones.
Before turning to inequalities satisfied by the flag vectors of Eulerian posets, we consider the question of whether the two cones ${\cal C}^{n+1}_{\cal D}$ and ${\cal C}^{n+1}_{\cal E}$ are equal. For low ranks the two cones are the same, as seen below. We know of no example in any rank of an Eulerian poset whose flag vector is not contained in the cone ${\cal C}^{n+1}_{\cal D}$ of doubled half-Eulerian posets. To look for such an example we turn to the best known examples of Eulerian posets, the face lattices of polytopes. In [@Stanley-flag] Stanley proved the nonnegativity of the $cd$-index for “$S$-shellable regular CW-spheres”, a class of Eulerian posets that includes all polytopes. By a result of Billera, Ehrenborg, and Readdy ([@Billera-Ehrenborg-Readdy]), the lattice of regions of any oriented matroid also has a nonnegative $cd$-index. Proposition \[cd\] implies that nonnegative $cd$-indices (and the associated flag vectors) are in the cone generated by the $cd$-indices (flag vectors) of the doubles of limit posets associated with even interval systems.
${\cal C}^{n+1}_{\cal D}$ contains the flag vectors of all Eulerian posets with nonnegative $cd$-indices. This includes the face lattices of polytopes and the lattices of regions of oriented matroids.
The closed cone ${\cal C}^{n+1}_{\cal E}$ of flag vectors of Eulerian posets is the same as the closed cone ${\cal C}^{n+1}_{\cal D}$ of flag vectors of horizontal doubles of half-Eulerian posets.
Inequalities {#sec-inequalities}
============
Throughout this section we use the following notation.
[*The [*interval system ${\ensuremath{{\cal I}}}[S]$ of a set $S\subseteq [1,n]$*]{} is the family of intervals ${\ensuremath{{\cal I}}}[S]=\{[a_1,b_1],\ldots,[a_k,b_k]\}$, where $S=[a_1,b_1]\cup \cdots\cup [a_k,b_k]$ and $b_{i-1}<a_i-1$ for $i\geq 2$. In other words, ${\ensuremath{{\cal I}}}[S]$ is the collection of the maximal intervals contained in $S$.* ]{}
Note that $S$ is an even set if and only if ${\ensuremath{{\cal I}}}[S]$ is an even interval system.
The following flag vector forms can be proved nonnegative by writing them as convolutions of basic nonnegative forms [@Billera-Liu; @Kalai]. (See \[S\_ring\].) The issue of whether they give all linear inequalities on flag vectors of Eulerian posets was raised by Billera and Liu (see the discussion after Proposition 1.3 in [@Billera-Liu]). We give here a simple direct argument for their nonnegativity that avoids convolutions.
\[ineqlemma\] Let $T$ and $V$ be subsets of $[1,n]$ such that for every $I\in {\ensuremath{{\cal I}}}[V]$, $|I\cap T|\le 1$. Write $S=[1,n]\setminus V$. For $P$ any rank $n+1$ Eulerian poset, $$\sum_{R\subseteq T}(-2)^{|T\setminus R| } f_{S\cup R}(P)\ge 0.$$ Equivalently, $$(-1)^{|T|}\sum_{T\subseteq Q\subseteq V} L_Q(P)\ge 0.$$
The idea is that since no two elements of $T$ are in the same gap of $S$, elements with ranks in $T$ can be inserted independently in chains with rank set $S$. For $C$ an $S$-chain (i.e., a chain with rank set $S$) and $t\in T$, let $n_t(C)$ be the number of rank $t$ elements $x\in P$ such that $C\cup \{x\}$ is a chain of $P$. Since every interval of an Eulerian poset is Eulerian, $n_t(C)\ge 2$ for all $C$ and $t$. So $$\begin{aligned}
\sum_{R\subseteq T} (-2)^{|T\setminus R|} f_{S\cup R} (P)
&=& \sum_{R\subseteq T} (-2)^{|T\setminus R|}
\sum_{\mbox{\scriptsize $C$ an $S$-chain}} \prod_{t\in R} n_t(C)\\
&=& \sum_{\mbox{\scriptsize $C$ an $S$-chain}}
\sum_{R\subseteq T} (-2)^{|T\setminus R|} \prod_{t\in R} n_t(C)\\
&=& \sum_{\mbox{\scriptsize $C$ an $S$-chain}}
\prod_{t\in T} (n_t(C)-2)\ge 0.\\\end{aligned}$$ So the flag vector inequality is proved. The second inequality is simply the translation into -vector form.
Here are some new inequalities.
\[ijk-ineqs\] Let $1 \le i < j < k \le n$. For $P$ any rank $n+1$ Eulerian poset, $$f_{ik}(P) - 2f_i(P) - 2f_k(P) + 2f_j(P) \ge 0.$$
First order the rank $j$ elements of $P$ in the following way. Choose any order, $G_1$, $G_2$, …, $G_m$ for the components of the Hasse diagram of the rank-selected poset $P_{\{i,j,k\}}$. For each rank $j$ element $y$ of $P$, identify the component containing $y$ by $y\in G_{g(y)}$. Order the rank $j$ elements of $P$ in any way consistent with the ordering of components. That is, choose an order $y_1$, $y_2$, …, $y_r$ such that $y_s<y_t$ implies $g(y_s)\le g(y_t)$.
A rank $i$ element $x$ [*belongs*]{} to $y_q$ if $q$ is the least index such that $x < y_q$ in $P$. Write $I_q$ for the number of rank $i$ elements belonging to $y_q$, and $I'_q$ for the number of rank $i$ elements $x$ such that $x < y_q$, but $x$ does not belong to $y_q$. Similarly, a rank $k$ element $z$ [*belongs*]{} to $y_q$ if $q$ is the least index such that $y_q < z$ in $P$. Write $K_q$ for the number of rank $k$ elements belonging to $y_q$, and $K'_q$ for the number of rank $k$ elements $z$ such that $y_q < z$, but $z$ does not belong to $y_q$. Note that $I_q + I'_q \ge 2$ and $K_q + K'_q \ge 2$, since $P$ is Eulerian. A flag $x<z$ [*belongs*]{} to $y_q$ if $x<y_q<z$ and $q$ is the least index such that either $x<y_q$ or $y_q<z$.
Let $F= f_{ik}(P) - 2f_i(P) - 2f_k(P) + 2f_j(P)$. Let $F_q$ be the contribution to $F$ by elements and flags belonging to $y_q$. Thus, $$F_q = I_q K_q + I'_q K_q + I_q K'_q - 2I_q - 2K_q + 2.$$
If $I'_q\ge 2$, then $F_q=I_q(K_q+K'_q-2)+(I'_q-2)K_q+2 \ge 2$.
If $I'_q=K'_q=0$, then $F_q=(I_q-2)(K_q-2)-2\ge -2$.
In all other cases it is easy to check that $F_q\ge 0$.
Suppose that the rank $j$ elements in component $G_\ell$ are $y_s$, $y_{s+1}$, …, $y_t$. Then $I'_s = K'_s = 0$, so $F_s\ge -2$. Furthermore, $I_t = K_t = 0$, because any rank $i$ element $x$ related to $y_t$ must also be related to at least one other rank $j$ element, and it is in the same component. That rank $j$ element has index less than $t$, so $x$ does not belong to $y_t$. This in turn implies $I'_t\ge 2$, so $F_t\ge 2$. For all $q$, $s < q < t$, either $I'_q>0$ or $K'_q>0$, by the connectivity of the component, so $F_q\ge 0$. Thus $\sum_{q=s}^t F_q \ge 0$. This is true for each component $G_\ell$, so $F=\sum_{q=1}^r F_q \ge 0$.
These inequalities can be used to generate others by convolution (see \[S\_ring\].)
Evaluating the flag vector inequalities of Proposition \[ineqlemma\] for the horizontal double ${\mbox{$D$}}P$ of a half-Eulerian poset $P$ gives the inequalities, for $S$ and $T$ satisfying the hypotheses of Proposition \[ineqlemma\], $$\label{halfEulineq1}
\sum_{R\subseteq T}(-1)^{|T\setminus R| } f_{S\cup R}(P)\ge 0.$$ These inequalities are valid not just for half-Eulerian posets but for [*all*]{} graded posets. The proof of Proposition \[ineqlemma\] uses only the fact that in every open interval of an Eulerian poset there are at least two elements of each rank. If the proof is rewritten using the assumption that in every open interval there is at least one element of each rank, the inequalities (\[halfEulineq1\]) are proved for all graded posets.
Similarly, the flag vector inequalities of Theorem \[ijk-ineqs\] give inequalities for half-Eulerian posets, $$f_{ik}(P) - f_i(P) - f_k(P) + f_j(P) \ge 0.$$ The proof of Theorem \[ijk-ineqs\] can be modified in the same way to show these inequalities are valid for all graded posets. The first instance of this class of inequalities was found by Billera and Liu ([@Billera-Liu]).
We conjecture that all inequalities valid for half-Eulerian posets come from inequalities valid for all graded posets. Inequalities for half-Eulerian posets are to be interpreted as conditions in the subspace of ${\bf R}^{2^n}$ spanned by flag vectors of half-Eulerian posets, but we are describing them in ${\bf R}^{2^n}$. Giving inequalities using linear forms in the flag numbers $f_S$ over ${\bf R}^{2^n}$, the statement is as follows.
Every linear form that is nonnegative for the flag vectors of all half-Eulerian posets is the sum of a linear form that is nonnegative for all graded posets and a linear form that is zero for all half-Eulerian posets.
Extreme Rays and Facets of the Cone {#sec-cone}
===================================
We have described some points in the Eulerian cone ${\cal C}^{n+1}_{{\cal E}}$ and some inequalities satisfied by all points in the cone. We turn now to identifying which of these give extreme rays and facets.
If ${\ensuremath{{\cal I}}}$ is an even interval system, then $(f_S(P(n,{\ensuremath{{\cal I}}})): S\subseteq [1,n])$ is on an extreme ray in the closed cone of flag vectors of all graded posets, and is in the subcone of flag $f$-vectors of half-Eulerian posets. Therefore it is on an extreme ray of the subcone.
For every even interval system ${\ensuremath{{\cal I}}}$, the flag vector of the limit poset $P(n,{\ensuremath{{\cal I}}})$ generates an extreme ray of the cone of flag vectors of half-Eulerian posets.
What does this say about the extreme rays of the cone of flag vectors of Eulerian posets? For every even interval system [${\cal I}$]{}, the flag vector of ${\mbox{$D$}}P(n,{\ensuremath{{\cal I}}})$ lies on an extreme ray of the subcone ${\cal C}^{n+1}_{{\cal D}}$, but we cannot conclude directly that it lies on an extreme ray of the cone ${\cal C}^{n+1}_{{\cal E}}$. A separate proof is needed.
For the following proofs, we use the computation of ${\ensuremath{\ell}}_Q(P(n,{\ensuremath{{\cal I}}}))$ (and ${\mbox{$L$}}_Q({\mbox{$D$}}P(n,{\ensuremath{{\cal I}}}))$) from the decompositions of $Q$ as the union of intervals of $\cal I$ (Proposition \[P\_limitl\]).
For every even interval system ${\ensuremath{{\cal I}}}$, the flag vector of the doubled limit poset ${\mbox{$D$}}P(n,{\ensuremath{{\cal I}}})$ generates an extreme ray of the cone of flag vectors of Eulerian posets.
We work in the closed cone of -vectors of Eulerian posets. The cone of -vectors of Eulerian posets is contained in the subspace of ${\bf R}^{2^n}$ determined by the equations $L_S=0$ for $S$ not an even set. To prove that the -vector of ${\mbox{$D$}}P(n,{\ensuremath{{\cal I}}})$ generates an extreme ray, we show that it lies on linearly independent supporting hyperplanes, one for each nonempty even set $V$ in $[1,n]$. Fix an even interval system ${\ensuremath{{\cal I}}}$. For each nonempty even set $V\subseteq[1,n]$, we find a set $T$ such that $T$ and $V$ satisfy the hypothesis of Proposition \[ineqlemma\] and $\sum_{T\subseteq Q\subseteq V} L_Q({\mbox{$D$}}P(n,{\ensuremath{{\cal I}}}))=0$.
[*Case 1.*]{} Suppose $V$ is the union of some intervals in [${\cal I}$]{}. Let $I_1$, $I_2$, …, $I_k$ be all the intervals of [${\cal I}$]{} contained in $V$. Set $T=\emptyset$. Then for each subset $J\subseteq [1,k]$, the corresponding union of intervals contributes $(-1)^{|J|}$ to $L_Q({\mbox{$D$}}P(n,{\ensuremath{{\cal I}}}))$, for $Q=\cup_{j\in J} I_j$. Thus $\sum_{T\subseteq Q\subseteq V} L_Q({\mbox{$D$}}P(n,{\ensuremath{{\cal I}}}))=
\sum_{J\subseteq [1,k]}(-1)^{|J|}=0$.
[*Case 2.*]{} If $V$ is not the union of some intervals in [${\cal I}$]{}, let $W$ be the union of all those intervals of [${\cal I}$]{} contained in $V$. Choose $t\in V\setminus W$, and set $T=\{t\}$. For $Q\subseteq V$, $L_Q({\mbox{$D$}}P(n,{\ensuremath{{\cal I}}}))=0$ unless $Q\subseteq W$. But if $Q\subseteq W$ then $t$ cannot be in $Q$. So $\sum_{\{t\}\subseteq Q\subseteq V} L_Q({\mbox{$D$}}P(n,{\ensuremath{{\cal I}}}))=0$.
Now $\sum_{T\subseteq Q\subseteq V} L_Q(P)=0$ determines a supporting hyperplane of the closed cone of -vectors of Eulerian posets, because the inequality of Proposition \[ineqlemma\] is valid, and the poset ${\mbox{$D$}}P(n,{\ensuremath{{\cal I}}})$ lies on the hyperplane. The hyperplane equations each involve a distinct maximal set $V$, which is even, so they are linearly independent on the subspace determined by the equations $L_S=0$ for $S$ not an even set. So the doubled limit poset ${\mbox{$D$}}P(n,{\ensuremath{{\cal I}}})$ is on an extreme ray of the cone.
Note how far we are, however, from a complete description of the extreme rays.
For every positive integer $n$, the closed cone of flag $f$-vectors of Eulerian posets of rank $n+1$ is finitely generated.
\[facetlemma\] Assume $\sum_{Q\subseteq[1,n]} a_QL_Q(P)\ge 0$ for all Eulerian posets $P$ of rank $n+1$. Let $M\subseteq [1,n]$ be a fixed even set. Suppose for all even sets $R\subseteq [1,n]$, $R\ne M$, there exists an interval system ${\ensuremath{{\cal I}}}(R)$ consisting of disjoint even intervals whose union is $R$ and such that $\sum_{Q\subseteq[1,n]} a_Q{\ensuremath{\ell}}_Q(P(n,{\ensuremath{{\cal I}}}(R)))= 0$. Then $\sum_{Q\subseteq[1,n]} a_QL_Q(P)= 0$ determines a facet of the closed cone of -vectors of Eulerian posets.
(Note that ${\ensuremath{{\cal I}}}(R)$ need not be ${\ensuremath{{\cal I}}}[R]$.)
The dimension of the cone ${\cal C}^{n+1}_{\cal E}$ equals the number of even subsets (a Fibonacci number). So it suffices to show that the vectors $({\ensuremath{\ell}}_Q(P(n,{\ensuremath{{\cal I}}}(R))))$ $\mbox{}=
(L_Q({\mbox{$D$}}P(n,{\ensuremath{{\cal I}}}(R) )))$ are linearly independent. To see this, note that for every set $Q$ not contained in $R$, ${\ensuremath{\ell}}_Q(P(n,{\ensuremath{{\cal I}}}(R)))=0$. By the disjointness of the intervals in ${\ensuremath{{\cal I}}}(R)$, there is a unique way to write $R$ as the union of intervals in ${\ensuremath{{\cal I}}}(R)$. So by Proposition \[P\_limitl\], $({\ensuremath{\ell}}_R(P(n,{\ensuremath{{\cal I}}}(R))))=(-1)^{|{\ensuremath{{\cal I}}}(R)|}$. Thus, $R$ is the unique maximal set $Q$ for which $({\ensuremath{\ell}}_Q(P(n,{\ensuremath{{\cal I}}}(R))))\ne 0$. So the -vectors of the posets ${\mbox{$D$}}P(n,{\ensuremath{{\cal I}}}(R))$, as $R$ ranges over sets different from $M$, are linearly independent.
\[ineq0\] The inequality $\sum_{Q\subseteq[1,n]} L_Q(P)\ge 0$ (or, equivalently, $f_\emptyset(P)\ge 0$) determines a facet of the closed cone of -vectors of Eulerian posets of rank $n+1$.
Apply the Facet Lemma \[facetlemma\] with $M=\emptyset$. For a nonempty even set $R$, the interval system ${\ensuremath{{\cal I}}}[R]$ of $R$ is nonempty, so $\sum_{Q\subseteq[1,n]} {\ensuremath{\ell}}_Q(P(n,{\ensuremath{{\cal I}}}[R]))
=\sum_{{\cal J}\subseteq {\ensuremath{{\cal I}}}[R]} (-1)^{|{\cal J}|} =0$.
\[facets\] Let $V$ be a subset of $[1,n]$ such that every $I\in {\ensuremath{{\cal I}}}[V]$ has cardinality at least $2$, and every $I\in {\ensuremath{{\cal I}}}[[0,n+1]\setminus V]$ has cardinality at most $3$. Assume that $M$ is a subset of $V$ such that every $[a,b]\in{\ensuremath{{\cal I}}}[V]$ satisfies the following:
1. $M\cap [a,b]= \emptyset,$ $[a,a+1]$, or $[b-1,b]$.
2. If $a\not \in M$ then $a-2\in \{-1\}\cup M$.
3. If $b\not \in M$ then $b+2\in \{n+2\}\cup M$.
Then $$\begin{aligned}
\label{facetineq}
(-1)^{|M|/2}\sum_{M\subseteq Q\subseteq V}L_Q(P)\ge 0\end{aligned}$$ determines a facet of ${\cal C}^{n+1}_{\cal E}$. Furthermore, if we strengthen [*($i$)*]{} by also requiring $M\cap [a,a+2]=\emptyset$ for every $[a,a+2]\in{\ensuremath{{\cal I}}}[V]$, then distinct pairs $(M,V)$ give distinct facets.
If $M=\emptyset$, then conditions ($ii$) and ($iii$) force $V=[1,n]$ (or $V=\emptyset$ if $n\le 1$). The resulting inequality, $\sum_{Q\subseteq[1,n]} L_Q(P)\ge 0$, gives a facet, as shown in Proposition \[ineq0\]. Now assume that $M\neq\emptyset$.
Step 1 is to prove that inequality (\[facetineq\]) holds for all Eulerian posets. Note that ${\ensuremath{{\cal I}}}[M]$ is a nonempty collection of intervals of length two. From each such interval choose one endpoint adjacent to an element of . Let $T$ be the set of these chosen elements. The Inequality Lemma \[ineqlemma\] applies to these $T$ and $V$ because each interval of $V$ contains at most one interval of ${\ensuremath{{\cal I}}}[M]$, and hence at most one element of $T$. The resulting inequality is $(-1)^{|T|}\sum_{T\subseteq Q\subseteq V} L_Q \ge 0$. Now $L_Q(P)=0$ for all $P$ if ${\ensuremath{{\cal I}}}[Q]$ contains an odd interval. So we can restrict the sum to even sets $Q$. Since $Q$ must be contained in $V$, such a $Q$ must contain the intervals of $M$. Thus, $(-1)^{|M|/2}\sum_{M\subseteq Q\subseteq V}L_Q(P)\ge 0$.
Step 2 is to prove that if $I\subseteq [1,n]$ is an interval of cardinality at least 2 and $I$ contains an element $i$ not in $V$, then $I$ contains an element adjacent to an interval of $M$. If an interval from ${\ensuremath{{\cal I}}}[V]$ ends at $i-1$, then either $i-1\in M$ or $i+1\in M$ by ($iii$) (since $i+1<n+2$). Similarly, if an interval from ${\ensuremath{{\cal I}}}[V]$ begins at $i+1$, then either $i-1\in M$ or $i+1\in M$. So assume no interval from ${\ensuremath{{\cal I}}}[V]$ begins at $i-1$ or ends at $i+1$. The hypothesis of the theorem states that every interval from ${\ensuremath{{\cal I}}}[[0,n+1]\setminus V]$ has cardinality at most three. Thus the interval $[i-1,i+1]$ belongs to ${\ensuremath{{\cal I}}}[[0,n+1]\setminus V]$. Hence $i-2\in \{-1\}\cup V$ and $i+2\in \{n+2\}\cup V$. If $i-2=-1$ then $I\supseteq
[i,i+1]=[1,2]$, condition ($ii$) applied to $a=3$ yields $3\in M$, and $2\in I$ is adjacent to $3$. The case when $i+2=n+2$ is dealt with similarly. Finally, if $i-2$ and $i+2$ are both endpoints of intervals from ${\ensuremath{{\cal I}}}[V]$, then, since $i\not\in M\cup\{-1,n+2\}$, condition ($ii$) applied to $a=i+2$ and condition ($iii$) applied to $b=i-2$ yield $i+2\in M$ and $i-2\in M$. Either $i-1$ or $i+1$ belongs to $I$ and each of them is adjacent to an element of $M$.
Recall that for ${\ensuremath{{\cal I}}}$ an even interval system, the vector $({\ensuremath{\ell}}_Q(P(n,{\ensuremath{{\cal I}}})): \mbox{$Q\subseteq [1,n]$})$ is in the closed cone of ${\ensuremath{\ell}}$-vectors of half-Eulerian posets. Step 3 is to show that for each even set $R\ne M$, there exists an even interval system ${\ensuremath{{\cal I}}}$ with $\cup_{i\in{\ensuremath{{\cal I}}}}I=R$ such that $(-1)^{|M|/2}\sum_{M\subseteq Q\subseteq V}
{\ensuremath{\ell}}_Q(P(n,{\ensuremath{{\cal I}}}))=0$.
Let $R$ be an even set not equal to $M$. If $M\not\subseteq R$, then for every $Q$ containing $M$, ${\ensuremath{\ell}}_Q(P(n,{\ensuremath{{\cal I}}}[R]))=0$. Now suppose $M\subseteq R$, but $R\not\subseteq V$. Let $I$ be an interval of ${\ensuremath{{\cal I}}}[R]$ such that $I\not\subseteq V$. Then $I$ contains an element adjacent to an interval of $M$. Since $M\subseteq R$ and $I$ is a maximal interval in $R$, $I\cap M\ne\emptyset$. Thus every union of intervals of ${\ensuremath{{\cal I}}}[R]$ containing $M$ must contain $I$ and thus an element not in $V$. So $\sum_{M\subseteq Q\subseteq V}{\ensuremath{\ell}}_Q(P(n,{\ensuremath{{\cal I}}}[R]))=0$, because all terms are zero.
Finally, suppose $M\subseteq R\subseteq V$ and $R\ne M$. Let ${\ensuremath{{\cal I}}}$ be the interval system of $R$ consisting only of intervals of length 2. Then every interval of $M$ is in ${\ensuremath{{\cal I}}}$. This is because every interval of $M$ is of length 2, with at least one of its endpoints adjacent to an element not in $V$. So $\sum_{M\subseteq Q\subseteq V} {\ensuremath{\ell}}_Q(P(n,{\ensuremath{{\cal I}}}))
=\sum_{{\ensuremath{{\cal I}}}[M]\subseteq {\cal J}\subseteq {\ensuremath{{\cal I}}}} (-1)^{|{\cal J}|}=0$, since $R\ne M$ implies ${\ensuremath{{\cal I}}}\ne{\ensuremath{{\cal I}}}[M]$.
By the Facet Lemma \[facetlemma\], the inequality $(-1)^{|M|/2}\sum_{M\subseteq Q\subseteq V}L_Q(P)\ge 0$ gives a facet of ${\cal C}^{n+1}_{\cal E}$.
Now we show that under the added condition $M\cap [a,a+2]=\emptyset$ for every $[a,a+2]\in {\ensuremath{{\cal I}}}[V]$, the facets obtained are distinct.
Note that two $(M,V)$ pairs can give the same inequality only if they have the same $M$, because ${\mbox{$L$}}_M$ is included in the linear form for $(M,V)$, and $M$ is the minimal (by set inclusion) set for which $L_M$ is in the form. Now for fixed $M$, we show that $(M,V_1)$ and $(M,V_2)$ give distinct linear inequalities when $V_1\neq V_2$. Since the sets $V_1$ and $V_2$ are different, there is an interval $[a,b]$ such that $[a, b]$ occurs in exactly one of ${\ensuremath{{\cal I}}}[V_1]$ or ${\ensuremath{{\cal I}}}[V_2]$. Let $[a,b]$ be a maximal interval with this property. Without loss of generality assume $[a,b]\in{\ensuremath{{\cal I}}}[V_1]$. Then $[a,b]$ is contained in no interval of ${\ensuremath{{\cal I}}}[V_2]$.
[*Case 1.*]{} $M\cap [a,b]=\emptyset$. Then for every $i$, $a\le i\le b-1$, the term $L_{[i,i+1]\cup M}$ occurs in the inequality for $(M,V_1)$. At least one of these terms does not occur in the inequality for $(M,V_2)$, because $[a,b]\not\subseteq V_2$.
[*Case 2.*]{} $M\cap [a,b]=[a,a+1]$. Since $M\subseteq V_2$ and $[a,b]\not\subseteq V_2$, $b>a+1$. By the strengthened hypothesis on $M$, $b\geq a+3$. Then for every $i$, $a+2\le i\le b-1$, the term $L_{[i,i+1]\cup M}$ occurs in the inequality for $(M,V_1)$. At least one of these terms does not occur in the inequality for $(M,V_2)$, because $[a,b]\not\subseteq V_2$.
[*Case 3.*]{} $M\cap [a,b]=[b-1,b]$. The proof is similar to Case 2.
Thus, with the condition $M\cap [a,a+2]=\emptyset$ for every $[a,a+2]\in {\ensuremath{{\cal I}}}[V]$, the facets given by the theorem are all distinct.
Theorem \[facets\] may be restated and interpreted in terms of the convolution of chain operators. We refer the interested reader to \[S\_ring\] for that approach.
With the aid of PORTA ([@porta]), we verified that the theorems above give all the extremes and facets of the Eulerian cone for rank at most 6.
\[rank6\] For rank $n+1\le 6$, the closed cone ${\cal C}^{n+1}_{{\cal E}}$ of flag vectors of Eulerian posets is finitely generated. It has ${n \choose \lfloor n/2\rfloor}$ extreme rays, all generated by the flag vectors of the limit posets ${\mbox{$D$}}P(n,{\ensuremath{{\cal I}}})$ for ${\ensuremath{{\cal I}}}$ even interval systems on $[1,n]$. It has ${n \choose \lfloor n/2\rfloor}$ facets, all given by Proposition \[ineq0\] and Theorem \[facets\].
At rank 7 the situation changes for both extreme rays and facets.
\[rank7\] [*($i$)*]{} The cone ${\cal C}^7_{\cal E}$ is finitely generated, with 24 extreme rays. Twenty of the extreme rays are generated by the flag vectors of the limit posets ${\mbox{$D$}}P(n,{\ensuremath{{\cal I}}})$ for ${\ensuremath{{\cal I}}}$ even interval systems on $[1,6]$.
[*($ii$)*]{} The cone ${\cal C}^7_{\cal E}$ has 23 facets. Fifteen of the facets are given by the inequalities of Theorem \[facets\]. Four additional facets come from the Inequality Lemma \[ineqlemma\]. The remaining four come from Theorem \[ijk-ineqs\].
The four special extreme rays of the rank 7 Eulerian cone have corresponding rays in the half-Eulerian cone. The generators for the half-Eulerian cone are all obtained by adding the flag vectors of limit posets associated with noneven interval systems. The summands do not satisfy the conditions of Proposition \[D\_hE\] for half-Eulerian posets, but the sum does. The calculations are easily done in terms of the ${\ensuremath{\ell}}$-vector, using Proposition \[P\_limitl\]. Specific sequences of half-Eulerian posets have been constructed whose flag vectors converge to these four extremes. The half-Eulerian posets are obtained by “gluing together” posets for each summand. These are then converted to Eulerian posets by the horizontal doubling operation. Below are the sums of limit posets used. Descriptions of the half-Eulerian posets are found in \[limit\_posets\].
Extreme 1: $P\left(6,\{[1,2],[2,6]\}+\{[2,5],[5,6]\}\right)$
Extreme 2: $P\left(6,\{[1,3],[3,4],[4,6]\}+\{[1,2],[2,3]\}
+\{[4,5],[5,6]\}\right)$
Extreme 3: $P\left(6,\{[1,2],[3,4],[4,5]\}+\{[3,5],[5,6]\}
+\{[1,2],[2,5]\}\right)$
Extreme 4: $P\left(6,\{[1,2],[2,4]\}+\{[2,5],[5,6]\}
+\{[2,3],[3,4],[5,6]\}\right)$
Note that for rank at most 7, the two cones ${\cal C}^{n+1}_{\cal D}$ and ${\cal C}^{n+1}_{\cal E}$ are equal, because the generators of extreme rays specified in Theorems \[rank6\] and \[rank7\] are horizontal doubles of half-Eulerian limit posets.
Perhaps all the extreme rays of the half-Eulerian cone (if not the Eulerian cone) can be obtained by gluing together Billera-Hetyei limit posets.
A complete description of the closed cone of flag vectors of Eulerian posets remains open, and, as mentioned before, the cone is not even known to be finitely generated. We do not know if convolutions of the inequalities of Proposition \[ineqlemma\] and Theorem \[ijk-ineqs\] completely determine the cone. A better understanding of the construction of extreme rays as sums of Billera-Hetyei limit posets would be valuable.
The study of Eulerian posets is motivated in part by questions about convex polytopes. Is the cone of flag vectors of all Eulerian posets the same as or close to the cone of flag vectors of polytopes? The answer is no. The inequalities of Proposition \[ineqlemma\] can be strengthened considerably for polytopes. The proof of Proposition \[ineqlemma\] uses only the fact that in an Eulerian poset each interval has at least two elements of each rank. For convex polytopes, each interval is at least the size of a Boolean algebra of the same rank. Thus, for example, where Proposition \[ineqlemma\] gives that $f_{1479}(P)-2f_{179}(P)\ge 0$ for Eulerian posets, for convex polytopes the inequality $f_{1479}(P)-20f_{179}(P)\ge 0$ holds, because the rank 6 Boolean algebra has ${6\choose 3}=20$ elements of rank 3. For ranks 4 through 7, we have verified that none of the extreme rays of the Eulerian cone is in the closed cone of flag vectors of convex polytopes.
Some half-Eulerian limit posets of rank $7$ {#limit_posets}
===========================================
Here are the constructions of half-Eulerian posets whose doubles give Extremes 1, 2 and 3 of ${\cal C}^7_{\cal E}$. Extreme 4 is the dual of Extreme 3.
In the following, $C^7$ denotes a chain of rank $7$.
$P\left(6,\{[1,2],[2,6] \}+\{[2,5],[5,6] \}\right)$
---------------------------------------------------
Take ${\mbox{$D$}}^{N}_{[1,2]}{\mbox{$D$}}^{N}_{[2,6]}\left(C^7\right)$ and ${\mbox{$D$}}^{N}_{[1,5]}{\mbox{$D$}}^{N}_{[5,6]}\left(C^7\right)$. Identify the elements of both posets at rank $1$ and at rank $6$. Figure \[F\_lpos1\] represents the resulting poset for $N=2$.
(3766,4366)(1118,-4644) (3001,-4561) (1801,-3361) (1201,-3361) (2401,-3361) (3001,-3361) (1801,-2161) (2401,-2161) (2401,-961) (3601,-961) (3601,-3961) (2401,-3961) (1801,-2761) (2401,-2761) (1801,-1561) (2401,-1561) (3001,-361) (3001,-1561) (3601,-1561) (4201,-1561) (4801,-1561) (3601,-3361) (4201,-3361) (3601,-2161) (4201,-2161) (3601,-2761) (4201,-2761)
(1276,-3361)[(1.6667,11.6667)[.]{}]{}
(2401,-3961)[( 1,-1)[600]{}]{} (3001,-4561)[( 1, 1)[600]{}]{}
(1201,-3361)[( 2,-1)[1200]{}]{}
(2401,-3961)[(-1, 1)[600]{}]{}
(2401,-3361)[( 2,-1)[1200]{}]{}
(3601,-3961)[(-1, 1)[600]{}]{}
(1201,-3361)[( 1, 1)[600]{}]{}
(1801,-3361)[( 1, 1)[600]{}]{}
(1801,-2761)[( 1,-1)[600]{}]{}
(2401,-2761)[( 1,-1)[600]{}]{}
(1801,-2761)[( 0, 1)[1200]{}]{}
(2401,-2761)[( 0, 1)[1200]{}]{}
(1801,-1561)[( 1, 1)[600]{}]{}
(2401,-1561)[( 2, 1)[1200]{}]{}
(2401,-961)[( 1, 1)[600]{}]{} (3001,-361)[( 1,-1)[600]{}]{}
(2401,-961)[( 1,-1)[600]{}]{}
(2401,-961)[( 2,-1)[1200]{}]{}
(3601,-961)[( 1,-1)[600]{}]{}
(3601,-961)[( 2,-1)[1200]{}]{}
(3001,-1561)[( 1,-1)[600]{}]{}
(3601,-1561)[( 1,-1)[600]{}]{}
(4201,-1561)[(-1,-1)[600]{}]{}
(4801,-1561)[(-1,-1)[600]{}]{}
(3601,-2161)[( 0,-1)[1200]{}]{}
(4201,-2161)[( 0,-1)[1200]{}]{}
(2401,-3961)[( 2, 1)[1200]{}]{}
(3601,-3961)[( 1, 1)[600]{}]{}
$P\left(6,\{[1,3],[3,4],[4,6]\}
+\{[1,2],[2,3]\}
+\{[4,5],[5,6] \}\right)$
-------------------------------
Take $$\begin{aligned}
P^I(N)&=&
{\mbox{$D$}}^{N}_{[1,3]}{\mbox{$D$}}^{N}_{[3,4]}{\mbox{$D$}}^{N}_{[4,6]}{\mbox{$D$}}^{N+1}_{[4,5]}
(C^7)\\ \\
P^{II}(N)&=&
{\mbox{$D$}}^{N+1}_{[1,2]}{\mbox{$D$}}^{N}_{[1,6]}{\mbox{$D$}}^{N}_{[2,4]}(C^7),
\quad\mbox{and}\\ \\
P^{III}(N)&=&
{\mbox{$D$}}^{N}_{[1,5]}{\mbox{$D$}}^{N}_{[3,5]}{\mbox{$D$}}^{N}_{[5,6]}(C^7).\\\end{aligned}$$ Identify the elements of $P^{I}(N)$ with the elements of $P^{II}(N)$ at ranks $1,4,5$, and $6$. Identify the elements of $P^{I}(N)$ with the elements of $P^{III}(N)$ at ranks $1,2,3$, and $6$. Figure \[F\_lpos2\] represents the resulting poset for $N=2$.
(6166,4366)(2918,-4644) (6001,-361) (6001,-4561) (7201,-3361) (7801,-3361) (8401,-3361) (4801,-3361) (4201,-3361) (3601,-3361) (5401,-3961) (6601,-3961) (3001,-3361) (9001,-3361) (6601,-3361) (5401,-3361) (3601,-2761) (4201,-2761) (4801,-2761) (7201,-2761) (7801,-2761) (8401,-2761) (3601,-2161) (4201,-2161) (4801,-2161) (5401,-2761) (6601,-2761) (5401,-2161) (7201,-2161) (7801,-2161) (8401,-2161) (6601,-2161) (5401,-1561) (6601,-1561) (6601,-961) (5401,-961) (4801,-1561) (4201,-1561) (3601,-1561) (3001,-1561) (7201,-1561) (7801,-1561) (8401,-1561) (9001,-1561)
(6601,-1561)[( 1,-1)[600]{}]{}
(6001,-361)[(-1,-1)[600]{}]{} (5401,-961)[( 0,-1)[600]{}]{}
(6001,-361)[( 1,-1)[600]{}]{} (6601,-961)[( 0,-1)[600]{}]{}
(5401,-1561)[( 2,-1)[1200]{}]{}
(6601,-1561)[(-2,-1)[1200]{}]{}
(5401,-1561)[(-1,-1)[600]{}]{}
(6601,-2161)[( 0,-1)[600]{}]{} (6601,-2761)[( 1, 1)[600]{}]{} (7201,-2161)[( 0,-1)[600]{}]{} (7201,-2761)[(-1, 1)[600]{}]{}
(4801,-2761)[( 1,-1)[600]{}]{}
(5401,-2761)[( 2,-1)[1200]{}]{}
(5401,-3361)[( 2, 1)[1200]{}]{}
(6601,-3361)[( 1, 1)[600]{}]{}
(5401,-3361)[( 0,-1)[600]{}]{}
(6601,-3361)[( 0,-1)[600]{}]{}
(5401,-3961)[( 1,-1)[600]{}]{} (6001,-4561)[( 1, 1)[600]{}]{}
(3601,-2761)[( 2, 1)[1200]{}]{}
(6601,-2161)[( 2,-1)[1200]{}]{}
(7201,-2161)[( 2,-1)[1200]{}]{}
(4201,-2761)[( 2, 1)[1200]{}]{}
(4801,-2161)[( 0,-1)[600]{}]{} (4801,-2761)[( 1, 1)[600]{}]{} (5401,-2161)[( 0,-1)[600]{}]{} (5401,-2761)[(-1, 1)[600]{}]{}
(3601,-2761)[(-1,-1)[600]{}]{}
(3601,-2761)[( 0,-1)[600]{}]{}
(3601,-3361)[( 5,-1)[3000]{}]{}
(3001,-3361)[( 4,-1)[2400]{}]{}
(4201,-2761)[( 0,-1)[600]{}]{}
(4201,-2761)[( 1,-1)[600]{}]{}
(4201,-3361)[( 2,-1)[1200]{}]{}
(4801,-3361)[( 3,-1)[1800]{}]{}
(7801,-2761)[(-1,-1)[600]{}]{}
(7801,-2761)[( 0,-1)[600]{}]{}
(8401,-2761)[( 0,-1)[600]{}]{}
(8401,-2761)[( 1,-1)[600]{}]{}
(5401,-3961)[( 3, 1)[1800]{}]{}
(6601,-3961)[( 2, 1)[1200]{}]{}
(5401,-3961)[( 5, 1)[3000]{}]{}
(6601,-3961)[( 4, 1)[2400]{}]{}
(5401,-2761)[(-2, 1)[1200]{}]{}
(4801,-2761)[(-2, 1)[1200]{}]{}
(6601,-2761)[( 2, 1)[1200]{}]{}
(7201,-2761)[( 2, 1)[1200]{}]{}
(3601,-2161)[(-1, 1)[600]{}]{} (3001,-1561)[( 4, 1)[2400]{}]{}
(3601,-2161)[( 0, 1)[600]{}]{} (3601,-1561)[( 5, 1)[3000]{}]{}
(4201,-2161)[( 0, 1)[600]{}]{} (4201,-1561)[( 2, 1)[1200]{}]{}
(4201,-2161)[( 1, 1)[600]{}]{} (4801,-1561)[( 3, 1)[1800]{}]{}
(7801,-2161)[(-1, 1)[600]{}]{} (7201,-1561)[(-3, 1)[1800]{}]{}
(7801,-2161)[( 0, 1)[600]{}]{} (7801,-1561)[(-2, 1)[1200]{}]{}
(8401,-2161)[( 0, 1)[600]{}]{} (8401,-1561)[(-5, 1)[3000]{}]{}
(8401,-2161)[( 1, 1)[600]{}]{} (9001,-1561)[(-4, 1)[2400]{}]{}
$P\left(6,\{[1,2],[3,4],[4,5]\}
+\{[3,5],[5,6]\}
+\{[1,2],[2,5] \}\right)$
-------------------------------
Take
$$\begin{array}{rcll}
\vspace*{6pt}
P^I(N)&=&
{\mbox{$D$}}^{N+1}_{[1,2]}{\mbox{$D$}}^{N+1}_{[3,4]}{\mbox{$D$}}^{N}_{[3,6]}{\mbox{$D$}}^{N+1}_{[4,5]}
(C^7)&\mbox{(Figure \ref{F_lpos3a})}\\
\vspace*{6pt}
P^{II}(N)&=&
{\mbox{$D$}}^{N+1}_{[1,5]}{\mbox{$D$}}^{N^2}_{[3,5]}{\mbox{$D$}}^{N}_{[5,6]}(C^7)
&\mbox{(Figure \ref{F_lpos3b}), and}\\
P^{III}(N)&=&
{\mbox{$D$}}^{N+2}_{[1,2]}{\mbox{$D$}}^{N^2-N+2}_{[2,5]}{\mbox{$D$}}^{N}_{[1,6]}(C^7)&
\mbox{(Figure \ref{F_lpos3c}).}
\end{array}$$ Identify the elements of $P^{I}(N)$ with the elements of $P^{II}(N)$ at ranks $1,2$, and $6$. Identify the elements of $P^{I}(N)$ with the elements of $P^{III}(N)$ at rank $6$. Figure \[F\_lpos3\] represents the resulting poset for $N=2$.
(9766,9166)(1718,-9444) (2701,-2761) (2401,-2761) (2101,-2761) (3601,-961) (11101,-2761) (10801,-2761) (10501,-2761) (9601,-961) (6601,-361) (9001,-3961) (9301,-3961) (9601,-3961) (9901,-3961) (10201,-3961) (10501,-3961) (10801,-3961) (11101,-3961) (11401,-3961) (4201,-3961) (3901,-3961) (3601,-3961) (3301,-3961) (3001,-3961) (2701,-3961) (2401,-3961) (2101,-3961) (1801,-3961) (4201,-5161) (3901,-5161) (4501,-5161) (6601,-6361) (8101,-6361) (5101,-6361) (6001,-7561) (6601,-7561) (7201,-7561) (6601,-9361) (8701,-5161) (8701,-5161) (9001,-5161) (9301,-5161)
(2101,-2761)[( 5, 6)[1500]{}]{}
(2401,-2761)[( 2, 3)[1200]{}]{}
(2701,-2761)[( 1, 2)[900]{}]{}
(10501,-2761)[(-1, 2)[900]{}]{}
(10801,-2761)[(-2, 3)[1200]{}]{}
(11101,-2761)[(-5, 6)[1500]{}]{}
(3601,-961)[( 5, 1)[3000]{}]{} (6601,-361)[( 5,-1)[3000]{}]{}
(4201,-3961)[(-5, 4)[1500]{}]{}
(3901,-3961)[(-5, 4)[1500]{}]{}
(3601,-3961)[(-5, 4)[1500]{}]{}
(3301,-3961)[(-1, 2)[600]{}]{}
(3001,-3961)[(-1, 2)[600]{}]{}
(2701,-3961)[(-1, 2)[600]{}]{}
(2401,-3961)[( 1, 4)[300]{}]{}
(2101,-3961)[( 1, 4)[300]{}]{}
(1801,-3961)[( 1, 4)[300]{}]{}
(9001,-3961)[( 5, 4)[1500]{}]{}
(9301,-3961)[( 5, 4)[1500]{}]{}
(9601,-3961)[( 5, 4)[1500]{}]{}
(9901,-3961)[( 1, 2)[600]{}]{}
(10201,-3961)[( 1, 2)[600]{}]{}
(10501,-3961)[( 1, 2)[600]{}]{}
(10801,-3961)[(-1, 4)[300]{}]{}
(11101,-3961)[(-1, 4)[300]{}]{}
(11401,-3961)[(-1, 4)[300]{}]{}
(4201,-3961)[( 1,-4)[300]{}]{}
(3901,-3961)[( 1,-2)[600]{}]{}
(3601,-3961)[( 3,-4)[900]{}]{}
(3301,-3961)[( 3,-4)[900]{}]{}
(3001,-3961)[( 1,-1)[1200]{}]{}
(2701,-3961)[( 5,-4)[1500]{}]{}
(1801,-3961)[( 5,-3)[1500]{}]{} (3301,-4861)[( 2,-1)[600]{}]{}
(2101,-3961)[( 3,-2)[1800]{}]{}
(2401,-3961)[( 5,-4)[1500]{}]{}
(4501,-5161)[( 3,-1)[3600]{}]{}
(4201,-5161)[( 3,-1)[2700]{}]{} (6901,-6061)[( 4,-1)[1200]{}]{}
(3901,-5161)[( 3,-1)[2700]{}]{} (6601,-6061)[( 5,-1)[1500]{}]{}
(3901,-5161)[( 1,-1)[1200]{}]{}
(4201,-5161)[( 3,-4)[900]{}]{}
(4501,-5161)[( 1,-2)[600]{}]{}
(6001,-7561)[( 1,-3)[600]{}]{}
(6601,-7561)[( 0,-1)[1800]{}]{}
(7201,-7561)[(-1,-3)[600]{}]{}
(5101,-6361)[( 3,-4)[900]{}]{}
(6601,-6361)[( 0,-1)[1200]{}]{}
(8101,-6361)[(-3,-4)[900]{}]{}
(9001,-3961)[(-1,-4)[300]{}]{}
(9301,-3961)[(-1,-2)[600]{}]{}
(9601,-3961)[(-3,-4)[900]{}]{}
(9901,-3961)[(-3,-4)[900]{}]{}
(10201,-3961)[(-1,-1)[1200]{}]{}
(10501,-3961)[(-5,-4)[1500]{}]{}
(10801,-3961)[(-5,-4)[1500]{}]{}
(11101,-3961)[(-3,-2)[1800]{}]{}
(11401,-3961)[(-5,-3)[1500]{}]{} (9901,-4861)[(-2,-1)[600]{}]{}
(8101,-6361)[( 1, 1)[1200]{}]{}
(9001,-5161)[(-3,-4)[900]{}]{}
(8701,-5161)[(-1,-2)[600]{}]{}
(8701,-5161)[(-5,-3)[1500]{}]{} (7201,-6061)[(-2,-1)[600]{}]{}
(9001,-5161)[(-5,-3)[1500]{}]{} (7501,-6061)[(-3,-1)[900]{}]{}
(9301,-5161)[(-5,-3)[1500]{}]{} (7801,-6061)[(-4,-1)[1200]{}]{}
(3901,-5161)[( 5,-3)[1500]{}]{} (5401,-6061)[( 4,-1)[1200]{}]{}
(4201,-5161)[( 5,-3)[1500]{}]{} (5701,-6061)[( 3,-1)[900]{}]{}
(4501,-5161)[( 5,-3)[1500]{}]{} (6001,-6061)[( 2,-1)[600]{}]{}
(7366,9166)(2918,-9444) (6001,-2761) (5701,-2761) (5401,-2761) (5101,-2761) (4801,-2761) (4501,-2761) (4201,-2761) (3901,-2761) (3601,-2761) (3301,-2761) (3001,-2761) (6301,-2761) (3601,-961) (10201,-2761) (9901,-2761) (9601,-2761) (9301,-2761) (9001,-2761) (8701,-2761) (8401,-2761) (8101,-2761) (7801,-2761) (7501,-2761) (7201,-2761) (6901,-2761) (9601,-961) (6601,-361) (7201,-3961) (7501,-3961) (7801,-3961) (8101,-3961) (8401,-3961) (8701,-3961) (6001,-3961) (5701,-3961) (5401,-3961) (5101,-3961) (4801,-3961) (4501,-3961) (6001,-5161) (5701,-5161) (5401,-5161) (5101,-5161) (4801,-5161) (6301,-5161) (6601,-6361) (8101,-6361) (5101,-6361) (6001,-7561) (6601,-7561) (7201,-7561) (6601,-9361) (6901,-5161) (7201,-5161) (7501,-5161) (7801,-5161) (8101,-5161) (8401,-5161)
(3001,-2761)[( 1, 3)[600]{}]{}
(3301,-2761)[( 1, 6)[300]{}]{}
(3601,-2761)[( 0, 1)[1800]{}]{}
(3601,-961)[( 1,-6)[300]{}]{}
(4201,-2761)[(-1, 3)[600]{}]{}
(4501,-2761)[(-1, 2)[900]{}]{}
(4801,-2761)[(-2, 3)[1200]{}]{}
(5101,-2761)[(-5, 6)[1500]{}]{}
(5401,-2761)[(-1, 1)[1800]{}]{}
(5701,-2761)[(-6, 5)[1800]{}]{} (3901,-1261)[(-1, 1)[300]{}]{}
(6001,-2761)[(-4, 3)[2400]{}]{}
(6301,-2761)[(-3, 2)[2700]{}]{}
(6901,-2761)[( 3, 2)[2700]{}]{}
(7201,-2761)[( 4, 3)[2400]{}]{}
(7501,-2761)[( 6, 5)[1800]{}]{} (9301,-1261)[( 1, 1)[300]{}]{}
(7801,-2761)[( 1, 1)[1800]{}]{}
(8101,-2761)[( 5, 6)[1500]{}]{}
(8401,-2761)[( 2, 3)[1200]{}]{}
(8701,-2761)[( 1, 2)[900]{}]{}
(9001,-2761)[( 1, 3)[600]{}]{}
(9301,-2761)[( 1, 6)[300]{}]{}
(9601,-2761)[( 0, 1)[1800]{}]{}
(9601,-961)[( 1,-6)[300]{}]{}
(10201,-2761)[(-1, 3)[600]{}]{}
(3601,-961)[( 5, 1)[3000]{}]{} (6601,-361)[( 5,-1)[3000]{}]{}
(6301,-2761)[( 2,-1)[2400]{}]{}
(6001,-2761)[( 2,-1)[2400]{}]{}
(5701,-2761)[( 2,-1)[2400]{}]{}
(5401,-2761)[( 2,-1)[2400]{}]{}
(5101,-2761)[( 2,-1)[2400]{}]{}
(4801,-2761)[( 2,-1)[2400]{}]{}
(6901,-2761)[(-2,-1)[2400]{}]{}
(7201,-2761)[(-2,-1)[2400]{}]{}
(7501,-2761)[(-2,-1)[2400]{}]{}
(7801,-2761)[(-2,-1)[2400]{}]{}
(8101,-2761)[(-2,-1)[2400]{}]{}
(8401,-2761)[(-2,-1)[2400]{}]{}
(6001,-3961)[(-5, 4)[1500]{}]{}
(5701,-3961)[(-5, 4)[1500]{}]{}
(5401,-3961)[(-5, 4)[1500]{}]{}
(5101,-3961)[(-5, 4)[1500]{}]{}
(4801,-3961)[(-5, 4)[1500]{}]{}
(4501,-3961)[(-5, 4)[1500]{}]{}
(7201,-3961)[( 5, 4)[1500]{}]{}
(7501,-3961)[( 5, 4)[1500]{}]{}
(7801,-3961)[( 5, 4)[1500]{}]{}
(8101,-3961)[( 5, 4)[1500]{}]{}
(8401,-3961)[( 5, 4)[1500]{}]{}
(8701,-3961)[( 5, 4)[1500]{}]{}
(6001,-3961)[( 1,-4)[300]{}]{}
(5701,-3961)[( 1,-4)[300]{}]{}
(5401,-3961)[( 1,-4)[300]{}]{}
(5101,-3961)[( 1,-4)[300]{}]{}
(4801,-3961)[( 1,-4)[300]{}]{}
(4501,-3961)[( 1,-4)[300]{}]{}
(4801,-5161)[( 1,-4)[300]{}]{}
(5701,-5161)[(-1,-2)[600]{}]{}
(5101,-5161)[( 5,-4)[1500]{}]{}
(6001,-5161)[( 1,-2)[600]{}]{}
(5401,-5161)[( 5,-2)[1500]{}]{} (6901,-5761)[( 2,-1)[1200]{}]{}
(6301,-5161)[( 3,-2)[1800]{}]{}
(6001,-7561)[( 1,-3)[600]{}]{}
(6601,-7561)[( 0,-1)[1800]{}]{}
(7201,-7561)[(-1,-3)[600]{}]{}
(5101,-6361)[( 3,-4)[900]{}]{}
(6601,-6361)[( 0,-1)[1200]{}]{}
(8101,-6361)[(-3,-4)[900]{}]{}
(7201,-3961)[(-1,-4)[300]{}]{}
(7501,-3961)[(-1,-4)[300]{}]{}
(7801,-3961)[(-1,-4)[300]{}]{}
(8101,-3961)[(-1,-4)[300]{}]{}
(8401,-3961)[(-1,-4)[300]{}]{}
(8701,-3961)[(-1,-4)[300]{}]{}
(8401,-5161)[(-1,-4)[300]{}]{}
(7501,-5161)[( 1,-2)[600]{}]{}
(8101,-5161)[(-5,-4)[1500]{}]{}
(7201,-5161)[(-1,-2)[600]{}]{}
(7801,-5161)[(-5,-2)[1500]{}]{} (6301,-5761)[(-2,-1)[1200]{}]{}
(6901,-5161)[(-3,-2)[1800]{}]{}
(12766,9166)(218,-9444) (1801,-2761) (1501,-2761) (1201,-2761) (901,-2761) (3601,-961) (12001,-2761) (11701,-2761) (11401,-2761) (12301,-2761) (9601,-961) (6601,-361) (11701,-3961) (12001,-3961) (12301,-3961) (1501,-3961) (1201,-3961) (901,-3961) (12601,-3961) (601,-3961) (1201,-6361) (1501,-6361) (1801,-6361) (2101,-6361) (2401,-6361) (2701,-6361) (3001,-6361) (3301,-6361) (3601,-6361) (3901,-6361) (4201,-6361) (4501,-6361) (901,-6361) (601,-6361) (301,-6361) (4801,-6361) (2701,-5161) (2401,-5161) (2101,-5161) (1801,-5161) (1201,-7561) (2401,-7561) (3601,-7561) (4501,-7561) (6601,-9361) (10501,-5161) (10801,-5161) (11101,-5161) (11401,-5161) (8701,-7561) (9601,-7561) (10801,-7561) (12001,-7561) (8401,-6361) (8701,-6361) (9001,-6361) (9301,-6361) (9601,-6361) (9901,-6361) (10201,-6361) (10501,-6361) (10801,-6361) (11101,-6361) (11401,-6361) (11701,-6361) (12001,-6361) (12301,-6361) (12601,-6361) (12901,-6361)
(901,-2761)[( 3, 2)[2700]{}]{}
(1201,-2761)[( 4, 3)[2400]{}]{}
(1501,-2761)[( 6, 5)[1800]{}]{} (3301,-1261)[( 1, 1)[300]{}]{}
(1801,-2761)[( 1, 1)[1800]{}]{}
(11401,-2761)[(-1, 1)[1800]{}]{}
(11701,-2761)[(-6, 5)[1800]{}]{} (9901,-1261)[(-1, 1)[300]{}]{}
(12001,-2761)[(-4, 3)[2400]{}]{}
(12301,-2761)[(-3, 2)[2700]{}]{}
(3601,-961)[( 5, 1)[3000]{}]{} (6601,-361)[( 5,-1)[3000]{}]{}
(1801,-2761)[(-1,-4)[300]{}]{}
(1501,-2761)[(-1,-4)[300]{}]{}
(1201,-2761)[(-1,-4)[300]{}]{}
(901,-2761)[(-1,-4)[300]{}]{}
(11701,-3961)[(-1, 4)[300]{}]{}
(12001,-3961)[(-1, 4)[300]{}]{}
(12301,-3961)[(-1, 4)[300]{}]{}
(12601,-3961)[(-1, 4)[300]{}]{}
(601,-3961)[( 1,-1)[1200]{}]{} (1801,-5161)[(-5,-4)[1500]{}]{}
(901,-3961)[( 1,-1)[1200]{}]{} (2101,-5161)[(-5,-4)[1500]{}]{}
(1201,-3961)[( 1,-1)[1200]{}]{} (2401,-5161)[(-5,-4)[1500]{}]{}
(1501,-3961)[( 1,-1)[1200]{}]{} (2701,-5161)[(-5,-4)[1500]{}]{}
(1801,-5161)[(-1,-4)[300]{}]{}
(2101,-5161)[(-1,-4)[300]{}]{}
(2401,-5161)[(-1,-4)[300]{}]{}
(2701,-5161)[(-1,-4)[300]{}]{}
(1801,-5161)[( 3,-4)[900]{}]{}
(2101,-5161)[( 3,-4)[900]{}]{}
(2401,-5161)[( 3,-4)[900]{}]{}
(2701,-5161)[( 3,-4)[900]{}]{}
(1801,-5161)[( 5,-3)[1500]{}]{} (3301,-6061)[( 2,-1)[600]{}]{}
(2101,-5161)[( 5,-3)[1500]{}]{} (3601,-6061)[( 2,-1)[600]{}]{}
(2401,-5161)[( 5,-3)[1500]{}]{} (3901,-6061)[( 2,-1)[600]{}]{}
(2701,-5161)[( 5,-3)[1500]{}]{} (4201,-6061)[( 2,-1)[600]{}]{}
(301,-6361)[( 3,-4)[900]{}]{}
(601,-6361)[( 1,-2)[600]{}]{}
(901,-6361)[( 1,-4)[300]{}]{}
(1201,-6361)[( 0,-1)[1200]{}]{}
(1501,-6361)[( 3,-4)[900]{}]{}
(1801,-6361)[( 1,-2)[600]{}]{}
(2101,-6361)[( 1,-4)[300]{}]{}
(2401,-6361)[( 0,-1)[1200]{}]{}
(1201,-7561)[( 3,-1)[5400]{}]{}
(2401,-7561)[( 5,-2)[3000]{}]{} (5401,-8761)[( 2,-1)[1200]{}]{}
(3601,-7561)[( 5,-3)[3000]{}]{}
(4501,-7561)[( 6,-5)[1800]{}]{} (6301,-9061)[( 1,-1)[300]{}]{}
(2701,-6361)[( 3,-4)[900]{}]{}
(3001,-6361)[( 1,-2)[600]{}]{}
(3301,-6361)[( 1,-4)[300]{}]{}
(3601,-6361)[( 0,-1)[1200]{}]{}
(3901,-6361)[( 1,-2)[600]{}]{}
(4201,-6361)[( 1,-4)[300]{}]{}
(4501,-6361)[( 0,-1)[1200]{}]{}
(4801,-6361)[(-1,-4)[300]{}]{}
(11701,-3961)[(-1,-1)[1200]{}]{}
(12001,-3961)[(-1,-1)[1200]{}]{}
(12301,-3961)[(-1,-1)[1200]{}]{}
(12601,-3961)[(-1,-1)[1200]{}]{}
(11401,-5161)[( 5,-4)[1500]{}]{}
(11101,-5161)[( 5,-4)[1500]{}]{}
(10801,-5161)[( 5,-4)[1500]{}]{}
(10501,-5161)[( 5,-4)[1500]{}]{}
(11401,-5161)[( 1,-4)[300]{}]{}
(11101,-5161)[( 1,-4)[300]{}]{}
(10801,-5161)[( 1,-4)[300]{}]{}
(10501,-5161)[( 1,-4)[300]{}]{}
(11401,-5161)[(-3,-4)[900]{}]{}
(11101,-5161)[(-3,-4)[900]{}]{}
(10801,-5161)[(-3,-4)[900]{}]{}
(10501,-5161)[(-3,-4)[900]{}]{}
(11401,-5161)[(-5,-3)[1500]{}]{} (9901,-6061)[(-2,-1)[600]{}]{}
(11101,-5161)[(-5,-3)[1500]{}]{} (9601,-6061)[(-2,-1)[600]{}]{}
(10801,-5161)[(-5,-3)[1500]{}]{} (9301,-6061)[(-2,-1)[600]{}]{}
(10501,-5161)[(-5,-3)[1500]{}]{} (9001,-6061)[(-2,-1)[600]{}]{}
(12901,-6361)[(-3,-4)[900]{}]{}
(12601,-6361)[(-1,-2)[600]{}]{}
(12301,-6361)[(-1,-4)[300]{}]{}
(12001,-6361)[( 0,-1)[1200]{}]{}
(11701,-6361)[(-3,-4)[900]{}]{}
(11401,-6361)[(-1,-2)[600]{}]{}
(11101,-6361)[(-1,-4)[300]{}]{}
(10801,-6361)[( 0,-1)[1200]{}]{}
(10501,-6361)[(-3,-4)[900]{}]{}
(10201,-6361)[(-1,-2)[600]{}]{}
(9901,-6361)[(-1,-4)[300]{}]{}
(9601,-6361)[( 0,-1)[1200]{}]{}
(9301,-6361)[(-1,-2)[600]{}]{}
(9001,-6361)[(-1,-4)[300]{}]{}
(8701,-6361)[( 0,-1)[1200]{}]{}
(8401,-6361)[( 1,-4)[300]{}]{}
(8701,-7561)[(-6,-5)[1800]{}]{} (6901,-9061)[(-1,-1)[300]{}]{}
(9601,-7561)[(-5,-3)[3000]{}]{}
(10801,-7561)[(-5,-2)[3000]{}]{} (7801,-8761)[(-2,-1)[1200]{}]{}
(12001,-7561)[(-3,-1)[5400]{}]{}
(12766,9166)(218,-9444) (6001,-2761) (5701,-2761) (5401,-2761) (5101,-2761) (4801,-2761) (4501,-2761) (4201,-2761) (3901,-2761) (3601,-2761) (3301,-2761) (3001,-2761) (2701,-2761) (2401,-2761) (2101,-2761) (1801,-2761) (1501,-2761) (1201,-2761) (901,-2761) (6301,-2761) (3601,-961) (12001,-2761) (11701,-2761) (11401,-2761) (11101,-2761) (10801,-2761) (10501,-2761) (10201,-2761) (9901,-2761) (9601,-2761) (9301,-2761) (9001,-2761) (8701,-2761) (8401,-2761) (8101,-2761) (7801,-2761) (7501,-2761) (7201,-2761) (6901,-2761) (12301,-2761) (9601,-961) (6601,-361) (7201,-3961) (7501,-3961) (7801,-3961) (8101,-3961) (8401,-3961) (8701,-3961) (9001,-3961) (9301,-3961) (9601,-3961) (9901,-3961) (10201,-3961) (10501,-3961) (10801,-3961) (11101,-3961) (11401,-3961) (11701,-3961) (12001,-3961) (12301,-3961) (6001,-3961) (5701,-3961) (5401,-3961) (5101,-3961) (4801,-3961) (4501,-3961) (4201,-3961) (3901,-3961) (3601,-3961) (3301,-3961) (3001,-3961) (2701,-3961) (2401,-3961) (2101,-3961) (1801,-3961) (1501,-3961) (1201,-3961) (901,-3961) (12601,-3961) (601,-3961) (6001,-5161) (5701,-5161) (5401,-5161) (5101,-5161) (4801,-5161) (4201,-5161) (3901,-5161) (1201,-6361) (1501,-6361) (1801,-6361) (2101,-6361) (2401,-6361) (2701,-6361) (3001,-6361) (3301,-6361) (3601,-6361) (3901,-6361) (4201,-6361) (4501,-6361) (901,-6361) (601,-6361) (4501,-5161) (6301,-5161) (6601,-6361) (8101,-6361) (5101,-6361) (301,-6361) (4801,-6361) (2701,-5161) (2401,-5161) (2101,-5161) (1801,-5161) (1201,-7561) (2401,-7561) (3601,-7561) (4501,-7561) (6001,-7561) (6601,-7561) (7201,-7561) (6601,-9361) (6901,-5161) (7201,-5161) (7501,-5161) (7801,-5161) (8101,-5161) (8401,-5161) (8701,-5161) (8701,-5161) (9001,-5161) (9301,-5161) (10501,-5161) (10801,-5161) (11101,-5161) (11401,-5161) (8701,-7561) (9601,-7561) (10801,-7561) (12001,-7561) (8401,-6361) (8701,-6361) (9001,-6361) (9301,-6361) (9601,-6361) (9901,-6361) (10201,-6361) (10501,-6361) (10801,-6361) (11101,-6361) (11401,-6361) (11701,-6361) (12001,-6361) (12301,-6361) (12601,-6361) (12901,-6361)
(901,-2761)[( 3, 2)[2700]{}]{}
(1201,-2761)[( 4, 3)[2400]{}]{}
(1501,-2761)[( 6, 5)[1800]{}]{} (3301,-1261)[( 1, 1)[300]{}]{}
(1801,-2761)[( 1, 1)[1800]{}]{}
(2101,-2761)[( 5, 6)[1500]{}]{}
(2401,-2761)[( 2, 3)[1200]{}]{}
(2701,-2761)[( 1, 2)[900]{}]{}
(3001,-2761)[( 1, 3)[600]{}]{}
(3301,-2761)[( 1, 6)[300]{}]{}
(3601,-2761)[( 0, 1)[1800]{}]{}
(3601,-961)[( 1,-6)[300]{}]{}
(4201,-2761)[(-1, 3)[600]{}]{}
(4501,-2761)[(-1, 2)[900]{}]{}
(4801,-2761)[(-2, 3)[1200]{}]{}
(5101,-2761)[(-5, 6)[1500]{}]{}
(5401,-2761)[(-1, 1)[1800]{}]{}
(5701,-2761)[(-6, 5)[1800]{}]{} (3901,-1261)[(-1, 1)[300]{}]{}
(6001,-2761)[(-4, 3)[2400]{}]{}
(6301,-2761)[(-3, 2)[2700]{}]{}
(6901,-2761)[( 3, 2)[2700]{}]{}
(7201,-2761)[( 4, 3)[2400]{}]{}
(7501,-2761)[( 6, 5)[1800]{}]{} (9301,-1261)[( 1, 1)[300]{}]{}
(7801,-2761)[( 1, 1)[1800]{}]{}
(8101,-2761)[( 5, 6)[1500]{}]{}
(8401,-2761)[( 2, 3)[1200]{}]{}
(8701,-2761)[( 1, 2)[900]{}]{}
(9001,-2761)[( 1, 3)[600]{}]{}
(9301,-2761)[( 1, 6)[300]{}]{}
(9601,-2761)[( 0, 1)[1800]{}]{}
(9601,-961)[( 1,-6)[300]{}]{}
(10201,-2761)[(-1, 3)[600]{}]{}
(10501,-2761)[(-1, 2)[900]{}]{}
(10801,-2761)[(-2, 3)[1200]{}]{}
(11101,-2761)[(-5, 6)[1500]{}]{}
(11401,-2761)[(-1, 1)[1800]{}]{}
(11701,-2761)[(-6, 5)[1800]{}]{} (9901,-1261)[(-1, 1)[300]{}]{}
(12001,-2761)[(-4, 3)[2400]{}]{}
(12301,-2761)[(-3, 2)[2700]{}]{}
(3601,-961)[( 5, 1)[3000]{}]{} (6601,-361)[( 5,-1)[3000]{}]{}
(6301,-2761)[( 2,-1)[2400]{}]{}
(6001,-2761)[( 2,-1)[2400]{}]{}
(5701,-2761)[( 2,-1)[2400]{}]{}
(5401,-2761)[( 2,-1)[2400]{}]{}
(5101,-2761)[( 2,-1)[2400]{}]{}
(4801,-2761)[( 2,-1)[2400]{}]{}
(6901,-2761)[(-2,-1)[2400]{}]{}
(7201,-2761)[(-2,-1)[2400]{}]{}
(7501,-2761)[(-2,-1)[2400]{}]{}
(7801,-2761)[(-2,-1)[2400]{}]{}
(8101,-2761)[(-2,-1)[2400]{}]{}
(8401,-2761)[(-2,-1)[2400]{}]{}
(6001,-3961)[(-5, 4)[1500]{}]{}
(5701,-3961)[(-5, 4)[1500]{}]{}
(5401,-3961)[(-5, 4)[1500]{}]{}
(5101,-3961)[(-5, 4)[1500]{}]{}
(4801,-3961)[(-5, 4)[1500]{}]{}
(4501,-3961)[(-5, 4)[1500]{}]{}
(4201,-3961)[(-5, 4)[1500]{}]{}
(3901,-3961)[(-5, 4)[1500]{}]{}
(3601,-3961)[(-5, 4)[1500]{}]{}
(3301,-3961)[(-1, 2)[600]{}]{}
(3001,-3961)[(-1, 2)[600]{}]{}
(2701,-3961)[(-1, 2)[600]{}]{}
(2401,-3961)[( 1, 4)[300]{}]{}
(2101,-3961)[( 1, 4)[300]{}]{}
(1801,-3961)[( 1, 4)[300]{}]{}
(1801,-2761)[(-1,-4)[300]{}]{}
(1501,-2761)[(-1,-4)[300]{}]{}
(1201,-2761)[(-1,-4)[300]{}]{}
(901,-2761)[(-1,-4)[300]{}]{}
(7201,-3961)[( 5, 4)[1500]{}]{}
(7501,-3961)[( 5, 4)[1500]{}]{}
(7801,-3961)[( 5, 4)[1500]{}]{}
(8101,-3961)[( 5, 4)[1500]{}]{}
(8401,-3961)[( 5, 4)[1500]{}]{}
(8701,-3961)[( 5, 4)[1500]{}]{}
(9001,-3961)[( 5, 4)[1500]{}]{}
(9301,-3961)[( 5, 4)[1500]{}]{}
(9601,-3961)[( 5, 4)[1500]{}]{}
(9901,-3961)[( 1, 2)[600]{}]{}
(10201,-3961)[( 1, 2)[600]{}]{}
(10501,-3961)[( 1, 2)[600]{}]{}
(10801,-3961)[(-1, 4)[300]{}]{}
(11101,-3961)[(-1, 4)[300]{}]{}
(11401,-3961)[(-1, 4)[300]{}]{}
(11701,-3961)[(-1, 4)[300]{}]{}
(12001,-3961)[(-1, 4)[300]{}]{}
(12301,-3961)[(-1, 4)[300]{}]{}
(12601,-3961)[(-1, 4)[300]{}]{}
(6001,-3961)[( 1,-4)[300]{}]{}
(5701,-3961)[( 1,-4)[300]{}]{}
(5401,-3961)[( 1,-4)[300]{}]{}
(5101,-3961)[( 1,-4)[300]{}]{}
(4801,-3961)[( 1,-4)[300]{}]{}
(4501,-3961)[( 1,-4)[300]{}]{}
(4201,-3961)[( 1,-4)[300]{}]{}
(3901,-3961)[( 1,-2)[600]{}]{}
(3601,-3961)[( 3,-4)[900]{}]{}
(3301,-3961)[( 3,-4)[900]{}]{}
(3001,-3961)[( 1,-1)[1200]{}]{}
(2701,-3961)[( 5,-4)[1500]{}]{}
(1801,-3961)[( 5,-3)[1500]{}]{} (3301,-4861)[( 2,-1)[600]{}]{}
(2101,-3961)[( 3,-2)[1800]{}]{}
(2401,-3961)[( 5,-4)[1500]{}]{}
(4501,-5161)[( 3,-1)[3600]{}]{}
(4201,-5161)[( 3,-1)[2700]{}]{} (6901,-6061)[( 4,-1)[1200]{}]{}
(3901,-5161)[( 3,-1)[2700]{}]{} (6601,-6061)[( 5,-1)[1500]{}]{}
(3901,-5161)[( 1,-1)[1200]{}]{}
(4201,-5161)[( 3,-4)[900]{}]{}
(4501,-5161)[( 1,-2)[600]{}]{}
(4801,-5161)[( 1,-4)[300]{}]{}
(5701,-5161)[(-1,-2)[600]{}]{}
(5101,-5161)[( 5,-4)[1500]{}]{}
(6001,-5161)[( 1,-2)[600]{}]{}
(5401,-5161)[( 5,-2)[1500]{}]{} (6901,-5761)[( 2,-1)[1200]{}]{}
(6301,-5161)[( 3,-2)[1800]{}]{}
(601,-3961)[( 1,-1)[1200]{}]{} (1801,-5161)[(-5,-4)[1500]{}]{}
(901,-3961)[( 1,-1)[1200]{}]{} (2101,-5161)[(-5,-4)[1500]{}]{}
(1201,-3961)[( 1,-1)[1200]{}]{} (2401,-5161)[(-5,-4)[1500]{}]{}
(1501,-3961)[( 1,-1)[1200]{}]{} (2701,-5161)[(-5,-4)[1500]{}]{}
(1801,-5161)[(-1,-4)[300]{}]{}
(2101,-5161)[(-1,-4)[300]{}]{}
(2401,-5161)[(-1,-4)[300]{}]{}
(2701,-5161)[(-1,-4)[300]{}]{}
(1801,-5161)[( 3,-4)[900]{}]{}
(2101,-5161)[( 3,-4)[900]{}]{}
(2401,-5161)[( 3,-4)[900]{}]{}
(2701,-5161)[( 3,-4)[900]{}]{}
(1801,-5161)[( 5,-3)[1500]{}]{} (3301,-6061)[( 2,-1)[600]{}]{}
(2101,-5161)[( 5,-3)[1500]{}]{} (3601,-6061)[( 2,-1)[600]{}]{}
(2401,-5161)[( 5,-3)[1500]{}]{} (3901,-6061)[( 2,-1)[600]{}]{}
(2701,-5161)[( 5,-3)[1500]{}]{} (4201,-6061)[( 2,-1)[600]{}]{}
(301,-6361)[( 3,-4)[900]{}]{}
(601,-6361)[( 1,-2)[600]{}]{}
(901,-6361)[( 1,-4)[300]{}]{}
(1201,-6361)[( 0,-1)[1200]{}]{}
(1501,-6361)[( 3,-4)[900]{}]{}
(1801,-6361)[( 1,-2)[600]{}]{}
(2101,-6361)[( 1,-4)[300]{}]{}
(2401,-6361)[( 0,-1)[1200]{}]{}
(1201,-7561)[( 3,-1)[5400]{}]{}
(2401,-7561)[( 5,-2)[3000]{}]{} (5401,-8761)[( 2,-1)[1200]{}]{}
(3601,-7561)[( 5,-3)[3000]{}]{}
(4501,-7561)[( 6,-5)[1800]{}]{} (6301,-9061)[( 1,-1)[300]{}]{}
(6001,-7561)[( 1,-3)[600]{}]{}
(6601,-7561)[( 0,-1)[1800]{}]{}
(7201,-7561)[(-1,-3)[600]{}]{}
(2701,-6361)[( 3,-4)[900]{}]{}
(3001,-6361)[( 1,-2)[600]{}]{}
(3301,-6361)[( 1,-4)[300]{}]{}
(3601,-6361)[( 0,-1)[1200]{}]{}
(3901,-6361)[( 1,-2)[600]{}]{}
(4201,-6361)[( 1,-4)[300]{}]{}
(4501,-6361)[( 0,-1)[1200]{}]{}
(4801,-6361)[(-1,-4)[300]{}]{}
(5101,-6361)[( 3,-4)[900]{}]{}
(6601,-6361)[( 0,-1)[1200]{}]{}
(8101,-6361)[(-3,-4)[900]{}]{}
(7201,-3961)[(-1,-4)[300]{}]{}
(7501,-3961)[(-1,-4)[300]{}]{}
(7801,-3961)[(-1,-4)[300]{}]{}
(8101,-3961)[(-1,-4)[300]{}]{}
(8401,-3961)[(-1,-4)[300]{}]{}
(8701,-3961)[(-1,-4)[300]{}]{}
(9001,-3961)[(-1,-4)[300]{}]{}
(9301,-3961)[(-1,-2)[600]{}]{}
(9601,-3961)[(-3,-4)[900]{}]{}
(9901,-3961)[(-3,-4)[900]{}]{}
(10201,-3961)[(-1,-1)[1200]{}]{}
(10501,-3961)[(-5,-4)[1500]{}]{}
(10801,-3961)[(-5,-4)[1500]{}]{}
(11101,-3961)[(-3,-2)[1800]{}]{}
(11401,-3961)[(-5,-3)[1500]{}]{} (9901,-4861)[(-2,-1)[600]{}]{}
(11701,-3961)[(-1,-1)[1200]{}]{}
(12001,-3961)[(-1,-1)[1200]{}]{}
(12301,-3961)[(-1,-1)[1200]{}]{}
(12601,-3961)[(-1,-1)[1200]{}]{}
(11401,-5161)[( 5,-4)[1500]{}]{}
(11101,-5161)[( 5,-4)[1500]{}]{}
(10801,-5161)[( 5,-4)[1500]{}]{}
(10501,-5161)[( 5,-4)[1500]{}]{}
(11401,-5161)[( 1,-4)[300]{}]{}
(11101,-5161)[( 1,-4)[300]{}]{}
(10801,-5161)[( 1,-4)[300]{}]{}
(10501,-5161)[( 1,-4)[300]{}]{}
(11401,-5161)[(-3,-4)[900]{}]{}
(11101,-5161)[(-3,-4)[900]{}]{}
(10801,-5161)[(-3,-4)[900]{}]{}
(10501,-5161)[(-3,-4)[900]{}]{}
(11401,-5161)[(-5,-3)[1500]{}]{} (9901,-6061)[(-2,-1)[600]{}]{}
(11101,-5161)[(-5,-3)[1500]{}]{} (9601,-6061)[(-2,-1)[600]{}]{}
(10801,-5161)[(-5,-3)[1500]{}]{} (9301,-6061)[(-2,-1)[600]{}]{}
(10501,-5161)[(-5,-3)[1500]{}]{} (9001,-6061)[(-2,-1)[600]{}]{}
(12901,-6361)[(-3,-4)[900]{}]{}
(12601,-6361)[(-1,-2)[600]{}]{}
(12301,-6361)[(-1,-4)[300]{}]{}
(12001,-6361)[( 0,-1)[1200]{}]{}
(11701,-6361)[(-3,-4)[900]{}]{}
(11401,-6361)[(-1,-2)[600]{}]{}
(11101,-6361)[(-1,-4)[300]{}]{}
(10801,-6361)[( 0,-1)[1200]{}]{}
(10501,-6361)[(-3,-4)[900]{}]{}
(10201,-6361)[(-1,-2)[600]{}]{}
(9901,-6361)[(-1,-4)[300]{}]{}
(9601,-6361)[( 0,-1)[1200]{}]{}
(9301,-6361)[(-1,-2)[600]{}]{}
(9001,-6361)[(-1,-4)[300]{}]{}
(8701,-6361)[( 0,-1)[1200]{}]{}
(8401,-6361)[( 1,-4)[300]{}]{}
(8701,-7561)[(-6,-5)[1800]{}]{} (6901,-9061)[(-1,-1)[300]{}]{}
(9601,-7561)[(-5,-3)[3000]{}]{}
(10801,-7561)[(-5,-2)[3000]{}]{} (7801,-8761)[(-2,-1)[1200]{}]{}
(12001,-7561)[(-3,-1)[5400]{}]{}
(8101,-6361)[( 1, 1)[1200]{}]{}
(9001,-5161)[(-3,-4)[900]{}]{}
(8701,-5161)[(-1,-2)[600]{}]{}
(8401,-5161)[(-1,-4)[300]{}]{}
(7501,-5161)[( 1,-2)[600]{}]{}
(8101,-5161)[(-5,-4)[1500]{}]{}
(7201,-5161)[(-1,-2)[600]{}]{}
(7801,-5161)[(-5,-2)[1500]{}]{} (6301,-5761)[(-2,-1)[1200]{}]{}
(6901,-5161)[(-3,-2)[1800]{}]{}
(8701,-5161)[(-5,-3)[1500]{}]{} (7201,-6061)[(-2,-1)[600]{}]{}
(9001,-5161)[(-5,-3)[1500]{}]{} (7501,-6061)[(-3,-1)[900]{}]{}
(9301,-5161)[(-5,-3)[1500]{}]{} (7801,-6061)[(-4,-1)[1200]{}]{}
(3901,-5161)[( 5,-3)[1500]{}]{} (5401,-6061)[( 4,-1)[1200]{}]{}
(4201,-5161)[( 5,-3)[1500]{}]{} (5701,-6061)[( 3,-1)[900]{}]{}
(4501,-5161)[( 5,-3)[1500]{}]{} (6001,-6061)[( 2,-1)[600]{}]{}
The Billera-Liu ring of chain operators {#S_ring}
=======================================
As in Billera and Liu ([@Billera-Liu]) we view the flag $f$-vector as a vector of [*chain operators*]{} $\left(f_S^{n+1}\::\: S\subseteq
[1,n]\right)$; here $f_S^{n+1} (P)= f_S(P)$ if $P$ is a graded poset of rank $n+1$ and $0$ otherwise. The following multiplication of chain operators $f^n_S$ ($n\geq 1$, $S\subseteq [1,n-1]$) was introduced by Kalai in [@Kalai] and studied for Eulerian posets by Billera and Liu in [@Billera-Liu]: $$f^m_S f^n_T{:=}f^{m+n}_{S\cup \{m\}\cup (T+m)}.$$ It is straightforward that given a pair of valid linear inequalities $$F=\sum_{S\subseteq [1,m-1]} a_S f^{m}_S \geq 0 \quad\mbox{and}\quad
G=\sum_{T\subseteq [1,n-1]} b_S f^{n}_S \geq 0$$ that hold for a class of graded posets, the linear inequality $FG\geq 0$ is also valid for the same class. It was observed by Billera and Liu in [@Billera-Liu Proposition 1.3] that for the class of all graded posets the converse holds as well: if $FG\geq 0$ is a valid inequality, then either both $F\geq 0$ and $G\geq 0$ are valid inequalities, or both $-F\geq 0$ and $-G\geq 0$ are valid inequalities. According to [@Billera-Liu Theorem 2.1] the associative algebra generated by all chain operators (whose domain is taken to be the class of all graded posets) is the free polynomial ring in variables $\{f^i_{\emptyset}\::\: i\geq 1\}$. If we take the degree of the variable $f^i_{\emptyset}$ to be $i$, then linear combinations of the form $F=\sum_{S\subseteq [1,m-1]} a_S f^{m}_S$ become homogeneous polynomials. Hence, as noted by Billera and Hetyei in [@Billera-Hetyei], one can use a result of Cohn in [@Cohn Theorem 3] that the semigroup of homogeneous polynomials of a free graded associative algebra has unique factorization. Hence an inequality can be checked factor-by-factor. Billera and Hetyei also showed in [@Billera-Hetyei] that for the class of all graded posets the product of two facet inequalities is almost always a facet inequality, every exception being a consequence of the equalities $$f^m_{\emptyset}f^n_{\emptyset}=f^{m+n}_m=
\left(f^{m+n}_m-f^{m+n}_{\emptyset}\right)+f^{m+n}_{\emptyset}.$$
For Eulerian and half-Eulerian posets, it is advisable to convert our expressions into the flag-${\ensuremath{\ell}}$ or flag-$L$ forms respectively. Straighforward substitution into the definition shows $${\ensuremath{\ell}}_S^m{\ensuremath{\ell}}_T^n={\ensuremath{\ell}}_{S\cup (T+m)}^{m+n} \quad\mbox{and}
\quad {\mbox{$L$}}_S^m{\mbox{$L$}}_T^n=2 {\mbox{$L$}}_{S\cup (T+m)}^{m+n}$$ This means that when we write $[u_S]={\mbox{$L$}}^{n}_S$ as the coefficient of the $ce$-word $u_S$, the convolution of the forms $\sum_{S\subseteq [1,m-1]} a_S [u_S]$ and $\sum_{T\subseteq [1,n-1]} b_T [u_T]$ is a constant multiple of the form $\sum_{S\subseteq [1,m-1]}\sum_{T\subseteq [1,n-1]} a_S b_T [u_S c
u_T].$ In particular, if only monomials of $c$ and $ee$ occur in each factor, the same holds for the convolution. Hence the same result of Cohn [@Cohn Theorem 3] on unique homogeneous factorization proves the following.
Every homogeneous linear form $\sum_{S\subseteq [1,n]} a_S {\ensuremath{\ell}}^{n+1}_S$ or\
$\sum_{S\subseteq [1,n]} a_S {\mbox{$L$}}^{n+1}_S$, where $S$ ranges over only even sets, can be uniquely written as a product of irreducible expressions of the same kind.
Let us call such expressions [*even ${\ensuremath{\ell}}$-forms*]{} and [*even ${\mbox{$L$}}$-forms*]{}, respectively. The interest in this factorization stems from the following observation.
Let $F$ and $G$ both be even ${\ensuremath{\ell}}$-forms. Then $FG\geq 0$ holds for all half-Eulerian posets if and only if either both $F\geq 0$ and $G\geq 0$ or both $-F\geq 0$ and $-G\geq 0$ hold for all half-Eulerian posets. The analogous statement is true for even ${\mbox{$L$}}$-forms and Eulerian posets.
Only the “only if” implication is not completely trivial. In the half-Eulerian case, all we need to observe is that for a pair $(P,Q)$ of half-Eulerian posets the poset $P\circ Q$ obtained by putting all elements of $Q$ above all elements of $P$, and identifying the top element of $P$ with the bottom element of $Q$, is half-Eulerian. Moreover, if for posets $P_1,P_2$, and $Q$ and forms $F$ and $G$, $F(P_1)>0$, $F(P_2)<0$, and $G(Q)>0$, then $FG(P_1\circ Q)=F(P_1)
G(Q)>0$ and $FG(P_2\circ Q)=F(P_2)
G(Q)<0$. The same argument works for Eulerian posets using ${\mbox{$D$}}^2_{\{{\rho}(P)\}} (P\circ Q)$ instead of $P\circ Q$.
In terms of convolutions, Proposition \[ineqlemma\] states that the product of valid inequalities of the form $f^n_{\emptyset}\geq 0$ and $f^n_{i}-2f^n_{\emptyset}\geq 0$ is a valid inequality for all Eulerian posets. Theorem \[facets\] describes a subclass of these products that yield facet inequalities. Using ideas extracted from the proof, one can show the following, somewhat strengthened statements.
If $F\geq 0$ defines a facet of ${\cal C}^{n+1}_{\cal E}$, then $F(f^{k+1}_1-2 f^{k+1}_{\emptyset})\geq 0$ defines a facet of ${\cal C}^{n+k+2}_{\cal E}$.
If $F\geq 0$ defines a facet of ${\cal C}^{n+1}_{\cal E}$, and $F$ can be written as $$F=\sum_{S\subseteq [1,n]} a_S L^{n+1}_{S}$$ where $S$ ranges over only even sets that contain $n$, then $Ff^{k+1}_{\emptyset}\geq 0$ and $Ff^1_{\emptyset}f^1_{\emptyset}\geq 0$ define facets of ${\cal C}^{n+k+2}_{\cal E}$ and ${\cal C}^{n+3}_{\cal E}$, respectively.
It seems to be difficult, however, even in the case of these simple factors to predict which products yield facet inequalities. For example $(f^5_1-2f^5_{\emptyset})
f^1_{\emptyset}=(f^6_1-2f^6_{\emptyset})+{1\over 2}
(f^3_1-2f^3_{\emptyset})(f^3_1-2f^3_{\emptyset})\geq 0$ does not define a facet of ${\cal C}^{6}_{\cal E}$, while it can be shown that $(f^5_1-2f^5_{\emptyset})f^3_{\emptyset}\geq 0$ defines a facet of ${\cal C}^{8}_{\cal E}$.
[99]{}
M. M. Bayer and L. J. Billera, Generalized Dehn–Sommerville relations for polytopes, spheres and Eulerian partially ordered sets, [*Invent. Math.*]{} [**79**]{} (1985), 143–157.
M. M. Bayer and G. Hetyei, Half-Eulerian partially ordered sets, in preparation, 1999.
M. M. Bayer and A. Klapper, A new index for polytopes, [*Discrete Comput. Geom.*]{} [**6**]{} (1991), 33–47.
L. J. Billera, R. Ehrenborg, and M. Readdy, The [**c**]{}-2[**d**]{}-index of oriented matroids, [*J. Combinatorial Theory*]{}, Ser. A [**80**]{} (1997), 79–105.
L. J. Billera and G. Hetyei, Linear inequalities for flags in graded partially ordered sets, to appear in [*J. Combinatorial Theory*]{}, Ser A.
L. J. Billera and N. Liu, Noncommutative enumeration in graded posets, to appear in J. Algebraic Combinatorics. A. Björner, Posets, regular CW complexes and Bruhat order, [*Europ. J. Combinatorics*]{} [**5**]{} (1984), 7–16.
A. Björner, M. Las Vergnas, B. Sturmfels, N. White, G. Ziegler, “Oriented Matroids,” Cambridge University Press, Cambridge, 1993.
P. M. Cohn, On subsemigroups of free semigroups, [*Proc. Amer. Math. Soc.*]{} [**13**]{} (1962), 347–351.
T. Christof and A. Loebel, PORTA—A Polyhedron Representation Transformation Algorithm, elib.zib.de, 1997
G. Kalai, A new basis for polytopes, [*J. Combinatorial Theory*]{}, Ser. A [**49**]{} (1988), 191–209.
R. P. Stanley, Balanced Cohen-Macaulay complexes, [*Trans. Amer. Math. Soc.*]{}, [**249**]{} (1979), 139–157.
R. P. Stanley, Some aspects of groups acting on finite posets, [*J. Combinatorial Theory*]{}, Ser. A [**32**]{} (1982), 132–161.
R. P. Stanley, [*Enumerative Combinatorics*]{}, Vol. I, Wadsworth and Brooks/Cole, Monterey, 1986.
R. P. Stanley, Flag $f$-vectors and the $cd$-index, [*Math. Z.*]{} [**216**]{} (1994). 483–499.
R. P. Stanley, A survey of Eulerian posets, in: “Polytopes: Abstract, Convex, and Computational,” T. Bisztriczky, P. McMullen, R. Schneider, A. I. Weiss, eds., NATO ASI Series C, vol. 440, Kluwer Academic Publishers, 1994, pages 301–333.
[^1]: This research was supported by University of Kansas General Research allocation \#3552.
[^2]: On leave from the Alfréd Rényi Institute of Mathematics, Hungarian Academy of Sciences. Partially supported by Hungarian National Foundation for Scientific Research grant no. F 023436.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'We have measured the variation of the spontaneous emission rate with polarization for self-assembled single quantum dots in two-dimensional photonic crystal membranes. We observe a maximum anisotropy factor of $6$ between the decay rates of the two bright exciton states. This large anisotropy is attributed to the substantially different projected local density of optical states for differently oriented dipoles in the photonic crystal.'
author:
- 'Q. Wang'
- 'S. Stobbe'
- 'H. Thyrrestup'
- 'H. Hofmann'
- 'M. Kamp'
- 'T. Schlereth'
- 'S. Höfling'
- 'P. Lodahl'
title: 'Polarization dependent spontaneous-emission rate of single quantum dots in photonic crystal membranes'
---
In the past few decades, there has been considerable interest in applying photonic crystals (PCs) for controlling the spontaneous emission (SE) of embedded emitters, which may find applications in diverse areas such as quantum information science, efficient lasers and LEDs, and for energy harvesting. Originally proposed by Yablonovitch in 1987[@yabl], the experimental progress has been delayed due to the lack of sufficiently high quality emitters and PCs. The first experimental demonstrations of spontaneous emission control have appeared within the last five years using colloidal quantum dots or dye molecules in 3D opal PCs [@peter1; @niko1; @niko2] and self-assembled quantum dots (QDs) or quantum wells in 2D photonic crystal membranes (PCMs) [@fuji; @finl1; @finl2; @jeppe]. The latter technology has proven very successful due to the excellent optical properties of self-assembled QDs [@jeppe2], the ability to optically address single QDs [@finl2], and the strongly modified optical local density of states (LDOS) in PCMs [@koen1]. Recently it was theoretically proposed that the spontaneous emission (SE) rate in a PC can be highly anisotropic depending on the orientation of the transition dipole moment of the emitter [@koen3], which may be employed to enhance effects of quantum interference between the two radiating states [@agar] of relevance for quantum information applications. Here we experimentally demonstrate such a pronounced anisotropy by carrying out time- and polarization-resolved spontaneous emission measurements on a single QD addressing two orthogonally polarized bright exciton states. In this process, we probe the anisotropy of the vacuum electromagnetic field in the PCMs, which was not addressed experimentally previously.
When optically exciting a QD, choosing the sample growth direction $
[001]$ as the quantization axis (z) for angular momentum, one lifts an electron ($S_{e,z}=\pm \frac{1}{2}$) to the conduction band leaving a heavy hole ($J_{h,z}=\pm \frac{3}{2}$) in the valence band, which can form four possible exciton states ($\left \vert h,e\right \rangle $): $\left
\vert \frac{3}{2},-\frac{1}{2}\right \rangle $, $\left \vert -\frac{3}{2},
\frac{1}{2}\right \rangle $, $\left \vert \frac{3}{2},\frac{1}{2}\right
\rangle $,$\left \vert -\frac{3}{2},-\frac{1}{2}\right \rangle $. We note that the light holes ($J_{h,z}=\pm \frac{1}{2}$) can be neglected as the degeneracy of the light and heavy holes is lifted by the strain causing the QDs [@baye]. The four exciton states are categorized into two groups according to the values of their total angular momentum: bright states ($J_{z}=\pm 1$) and dark states ($J_{z}=\pm 2$), where only the bright states are optically active. Due to the reduced symmetry of self-assembled QDs and anisotropic exchange interactions, the two bright states are separated in energy (typically 0-30 $\micro$eV) [@stev] and usually denoted as X or Y states according to their dipole orientations ($[110]$ or \[$1\overset{-}{1}0$\]). The QD spontaneous emission decay curves are in general bi-exponential, where the fast component, which is considered here, is due to recombination of the bright exciton transitions while the slow component is due to dark state recombination mediated by spin-flip processes [@jeppe3]. Polarization resolved spontaneous-emission measurements enable addressing each of the orthogonally polarized bright exciton states individually and thereby to probe the anisotropy of the vacuum electromagnetic field in the PCM. We quantify the polarization dependence by defining the anisotropy factor $\eta ^{\gamma
}\equiv \frac{\gamma _{X}}{\gamma _{Y}}$, where $\gamma _{X}$ ($\gamma _{Y}$) represents the decay rate of the X (Y) states.
![(Color online) The schematic of the experimental setup. CCD: charge coupled device camera; APD: avalanche photodiode detector; HWP: half-wave plate; PBS: polarization beam-splitter. The inset shows a scanning electron micrograph of the sample, in which a layer of self-assembled InAs QDs (red color) is embedded in the center of a GaAs PCM.[]{data-label="Fig1"}](Fig1.eps)
The schematic of our experimental setup is illustrated in Fig. 1. The sample is a GaAs PCM with a layer of self-assembled InAs QDs of density $250$ $\micro$m$^{-2}$ embedded in the center of the membrane, see Fig. 1. It is mounted in a closed-cycle cryostat at a temperature of $10$ K and excited from the top by a pulsed diode laser at $780$ nm ($1.590$ eV, which is above the bandgap of GaAs) and a repetition rate of $20$ MHz. The photoluminescence (PL) is collected through a lens (NA = $0.65$), sent to a monochromator, and arrives either at a CCD camera for recording emission spectra or a silicon avalanche photodiode for the time-resolved measurements. In order to facilitate polarization resolved measurements, a polarizer consisting of a half-wave plate and a polarization beam-splitter is placed before the monochromator. The excitation intensity used in the measurements is about $300$ mW/cm$^{2}$, which is below the exciton saturation level so that only photon emission from the ground state is observed. The ground state emission wavelength is centered at $950$ nm ($1.305$ eV) with an inhomogeneous broadening of $70$ meV. The resolution of the monochromator is about $120$ $\micro$eV, which is larger than the energy splitting between the two bright states. However, they can still be separated by their different polarization.
During our experiments, we investigated about $30$ different QDs positioned in 7 different PCMs, with the lattice parameters ranging from $260$ nm to $320$ nm. For the sake of exploiting a pronounced 2D PC bandgap effect, we chose QDs in PCMs with $r/a=0.30$, where $r$ is the radius of the air holes and $a$ is the lattice constant. For comparison, we also measured decay curves of $4$ QDs positioned outside the PCMs. For each QD, the PL was projected onto different polarization directions by changing the orientation of the half-wave plate before the monochromator.
![(Color online) (a). PL spectra for QDs positioned in a PCM ($a=320$ nm) measured at H or V polarizations displaying single QD lines. The shaded area represents the resolution of the spectrometer. (b). Three decay curves for QD A (inside PCM, emission energy $1.274$ eV) corresponding to $0^{\circ }$ (blue, upper curve), $70^{\circ }$ (cyan, middle curve) or $90^{\circ }$ (green, lower curve) polarization. Also shown are two decay curves for QD B (outside PCM, emission energy $1.267$ eV) for $
0^{\circ }$ (blue curve) and $90^{\circ }$ (green curve) polarizations that are almost on top of each other. The red lines are bi-exponential fits to the decay curves.[]{data-label="Fig2ab"}](Fig2ab.eps)
Fig. 2(a) shows the PL spectrum of single QDs inside a PCM by recording either horizontal ($H$) or vertical ($V$) polarizations. The spectrum is composed of sharp emission lines originating from single QDs with linewidths limited by the resolution of the spectrometer, as indicated by the shaded area in Fig. 2(a). Fig. 2(b) displays typical decay curves for two QDs, where QD A is inside a PCM, and QD B is in the unpatterned substrate while being close in emission energy to QD A. Three decay curves for QD A are displayed corresponding to different polarization components $0^{\circ }$ (i.e. H), $70^{\circ }$, and $%
90^{\circ }$ (i.e. V). We clearly observe that the SE rate is strongly dependent on polarization illustrating that X and Y bright excitons decay significantly different in the PCM due to the anisotropic vacuum fluctuations experienced by the QD. For comparison, no such anisotropy is observed in the reference measurements on QD B. The SE rate is furthermore found to be strongly inhibited in the PCM with the inhibition factors differing for X and Y. By comparing QD A and B we derive an inhibition factor of $15.8$ for the X state and $6.5$ for the Y state.
The PL intensity and decay rate obtained when probing different polarizations for QD A are presented in Fig. 3. Polarizations H and V correspond to probing the two orthogonally polarized bright states X and Y, while intermediate directions probe a combination of the two bright states. Note that this implies that only in the former case are the decay curves strictly bi-exponential functions. However this model turns out to model the decay curves rather well also for intermediate polarization settings, and the goodness-of-fit ($\chi ^{2}$) varying between $1.0$ and $1.4$ is found for the complete data set. The PL intensity shows a maximum (minimum) value at H (V) polarization, which is opposite to the decay rate. This is expected since a strong suppression of the decay rate in the plane of the PCM results in a high emission vertically out of the membrane due to energy redistribution [@fuji]. The PL intensity variation with polarization $\theta$ is observed to follow the simple relation $I=\frac{I_X+I_Y}{2}+\frac{I_X-I_Y}{2}\cos (2\theta )$, where $I_X$ and $I_Y$ are the intensities of the $X$ and $Y$ exciton states, respectively, see Fig. 3. This can be easily understood as the result of applying polarization projection measurements on two orthogonal states.
![(Color online) PL intensities and decay rates versus polarization for QD A. The triangular points (square points) are experimental results for intensities (decay rates). The solid line is the fitted result with a cosine function, and the dashed line is a guide to the eye.[]{data-label="Fig3"}](Fig3.eps)
![(Color online) Measured anisotropy factor of decay rates between X and Y states. The triangular points represent QDs inside PCMs (with 7 different lattice parameters), the circular points represent QDs outside PCMs, and the dashed horizontal line separates regions $\protect \gamma _{X}>\protect \gamma _{Y}$ and $\protect \gamma _{X}<\protect \gamma _{Y}$.[]{data-label="Fig4"}](Fig4.eps)
Fig. 4 shows the anisotropy factor of decay rates for all the measured QDs measured on PCMs with various values of the lattice constant. Note that in all measurements presented in the present manuscript the QD emission was within the 2D photonic bandgap of the PCMs [@koen1]. Large variations are observed between the individual QDs in the PCM with a maximum value of about $6$. This directly demonstrates the large anisotropy of the vacuum electromagnetic field in a PC that was theoretically proposed in Ref. [@koen3]. This anisotropy gives rise to substantial differences in the projected LDOS leading to the different decay dynamics of X and Y exciton states. For comparison, reference QDs in a bulk substrate showed no anisotropy in the decay rates for the two orthogonally polarized states.
To conclude, we have systematically measured the polarization dependent SE rate for self-assembled single QDs inside PCMs and obtained a maximum anisotropy factor of decay rate between the X and Y states of $6$. Our measurement results demonstrate the large anisotropy of the vacuum electromagnetic field inside PCMs [@koen1; @koen3], which is a crucial condition for achieving quantum interference between two closely lying energy levels [@agar] that could enable demonstration of fascinating phenomena, such as lasing without inversion [@imam] or quantum beats [@koch]. Therefore, our experiment is not only vital in realizing complete control of the SE of single QDs with PCs, but also enables fundamental quantum optics experiments with practical systems.
We thank T. Lund-Hansen and M. L. Andersen for help during the experiment, and we gratefully acknowledge financial supports from the Danish Research Council (FTP grant 274-07-0459 and FTP/FNU grant 272-09-0159).
[99]{} E. Yablonovitch, Phys. Rev. Lett. **58**, 2059 (1987).
P. Lodahl *et al.*, Nature **430**, 654 (2004).
I. S. Nikolaev, P. Lodahl, A. F. Driel, A. F. Koenderink, and W. L. Vos, Phys. Rev. B **75**, 115302 (2007).
I. S. Nikolaev, P. Lodahl, and W. L. Vos, J. Phys. Chem. C **112**, 7250 (2008).
M. Fujita *et al.*, Science **308**, 1296 (2005).
M. Kaniber *et al.*, Appl. Phys. Lett. **91**, 061106 (2007).
M. Kaniber *et al.*, Phys. Rev. B **77**, 073312 (2008).
B. Julsgaard *et al.*, Appl. Phys. Lett. **93**, 094102 (2008).
J. Johansen *et al.*, Physical Review B **77**, 073303 (2008).
A. F. Koenderink, M. Kafesaki, C. M. Soukoulis and V. Sandoghdar, J. Opt. Soc. Am. B **23**, 6 (2006).
W. L. Vos, A. F. Koenderink and I. S. Nikolaev, Phys. Rev. A **80**, 053802 (2009).
G. S. Agarwal, Phys. Rev. Lett. **84**, 5500 (2000).
M. Bayer *et al.*, Phys. Rev. B **65**, 195315 (2002).
R. M. Stevenson *et al.*, Phys. Rev. B **73**, 033306 (2006).
J. Johansen, B. Julsgaard, S. Stobbe, J. M. Hvam, P. Lodahl, arXiv:0905.4493.
A. Imamoglu, Phys. Rev. A **40**, R2835 (1989).
O. Kocharovskaya et al., Found. Phys. **28**, 561 (1998).
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
address: |
University of Bern, Sidlerstrasse 5, 3012 Bern, Switzerland\
E-mail: [email protected]
author:
- Sonja Kabana
title: A new interpretation of the QCD phase transition and of strangeness as QGP signature
---
Introduction
============
It has been repeateadly demonstrated in the literature [@pbm; @redlich_qm2001; @mapping] that in many cases the final state of nuclear and particle collisions is compatible with the hypothesis of an equilibrated source. In the following the discussion will concern only colliding systems for which this finding holds (measured by the $\chi^2$ of thermal model fits). Much work is presently concentrated on identifying the QCD phase transition at a certain $\sqrt{s}$, separating the colliding systems which go through the phase transition from those which dont. This defines the problem that we address. One way to approach this problem is to compare the estimated thermodynamic parameters temperature ($T$) and baryon chemical potential ($\mu_B$) at the chemical freeze-out extracted by thermal models from the data, to the expected critical ($T_c$, $\mu_B(c)$) values of QCD (e.g. [@pbm]). However, the theoretical expectations for the critical temperature $T_c$ are uncertain. It is thus obviously interesting to try to extract informations on the critical parameters from the data, without any use of the QCD predictions. In this spirit, we approach this problem in the present study in a different way, namely by extrapolating the data to equivalent systems with zero $\mu_B$ and studying their $T$ [@sqm2000]. To see why this is interesting, we follow a line of arguments : If all measured colliding systems above a certain $\sqrt{s}$ are heated enough to go through the phase transition and back, they would appear to have the same $T$ (if they don’t freeze-out in a considerably different way) which we note as ’limiting’ hadronic temperature $T_{lim}$. The colliding systems which do not reach the $T_c$ at any time because of their small $\sqrt{s}$, will exhibit a final $T$ smaller than $T_{lim}$. Therefore, in order to separate the colliding systems which went through the transition from the ones which did not, one can investigate the $T$ of all colliding systems at conditions of $\mu_B$=0, as a function of $\sqrt{s}$ or $dN/dy$ at midrapidity, and search for an increase of $T$ followed by a saturation starting at the $\sqrt{s}$ where the $T_c$ is reached. This would work if we use always the same projectile and target combination. Otherwise, one should correct for the fact that different projectile/target combinations at the same $\sqrt{s}$, reach different initial energy densities. For this reason, we will investigate the $T$ as a function of the initial energy density ($\epsilon_i$) reached at each collision after 1 fm/c based on the Bjorken estimate [@bjorken] and also using other Ansätze especially at low $\sqrt{s}$ [@mapping]. For a discussion on the uncertainty on $\epsilon_i$ see [@charm; @mapping; @border]. The same conclusions can however be reached by the present analysis, while looking only at one projectile/target combination A+A, namely with $A \sim 200$ (that is at Pb+Pb and Au+Au central collisions), as a function of e.g. $\sqrt{s}$ instead of $\epsilon_i$.
Energy density dependence of temperature and $\lambda_s$
========================================================
We compare the ratios of experimentally measured hadron yields in nuclear collisions with the prediction of a thermal model of non interacting free hadron resonances (for details see [@mapping]). We extract the thermodynamic parameters describing best the particle source: temperature, baryochemical potential ($\mu_B$) and strangeness chemical potential ($\mu_S$), imposing exact strangeness conservation. We then extrapolate all thermodynamic states with parameters ($T$, $\mu_B$, $\mu_s$) to equivalent states at zero chemical potentials ($T(\mu_B=0)$, 0, 0) along isentropic paths. As demonstrated in figure \[t\_vs\_e\] the resulting temperature at $\mu_B=0$ rises and saturates above $\epsilon_i \sim$ 1 $GeV/fm^3$. We interpret the onset of saturation of $T$ as due to the reach of an initial temperature greater or equal to $T_c$. We therefore extract from the onset of saturation in fig. \[t\_vs\_e\] the ’critical’ $\sqrt{s}$ of $\sim$ 8.8 GeV/nucleon pair respectively the critical $\epsilon_c$ of $\sim$ 1 +0.3 -0.5 GeV/fm$^3$ of the QCD phase transition, independently of the QCD predictions. An important assumption needed to support this interpretation is that the cooling until freeze-out of the particle source, is not significantly (as compared to the errors) dependent on the $\sqrt{s}$, in the range between $\sqrt{s}$=2 and 9 GeV for A+A collisions. [^1] A discussion of $T$ rise and saturation can be found in [@horst], but not for the $\mu_B$=0 case. However, it is only when investigating systems with the same $\mu_B$, where this behaviour is expected to be exact. The importance of using a common $\mu_B$ [@sqm2000; @mapping; @border], becomes more apparent when investigating strangeness. This is the next topic.
\[t\_vs\_e\]
The so called ’strangeness suppression factor’ $
\lambda_s \ = \
\frac{ (2 \overline{s}) }
{ (\overline{u} + \overline{d}) }
$ at nonzero $\mu_B$, rises with $\epsilon_i$ up to 1 GeV/fm$^3$ (line (a) in fig. \[ls\_vs\_e\], left) and then decreases (line (b) in fig. \[ls\_vs\_e\], left). After extrapolating to $\mu_B=0$, $\lambda_s$ rises and saturates universally above $\epsilon_i$ $\sim$ 1 GeV/fm$^3$ (fig. \[ls\_vs\_e\], right). Therefore the peak of $\lambda_s$ at the 40 A GeV Pb+Pb (point at $\epsilon_i$ = 1 GeV/fm$^3$ in fig. \[ls\_vs\_e\], left) is due to the high $\mu_B$ reached there. Furthermore, the so called ’strangeness enhancement’ defined usually as double ratio of strange/pion ratio in A+A over p+p collisions at the same $\sqrt{s}$, can be traced back to the different $\mu_B$ and $\epsilon_i$ of those reactions. In particular this explains why the so defined strangeness enhancement increases with decreasing $\sqrt{s}$ (fig. 7 in [@redlich_qm2001]), since this corresponds to increasing $\mu_B$. If one eliminates the bias due to different $\mu_B$ by considering only $\lambda_s$ at $\mu_B=0$, the non-trivial -with respect to the phase transition- ’strangeness enhancement’ is revealed (fig. \[ls\_vs\_e\], right) as a consequence of the behaviour of $T$ seen in fig. \[t\_vs\_e\]. We therefore redefine the ’strangeness enhancement’, as an enhancement of $\lambda_s$ in all thermalised systems above $\epsilon_c$, as compared to all thermalised systems with $\epsilon_i < \epsilon_c$, at the same $\mu_B$. [^2] We also analysed for this conference ratios from $e^+ e^-$ collisions at $\sqrt{s}$=3.6 GeV [@dasp] and found a T of 124 +15 -20 MeV, and $\lambda_s$=0.25 +0.06 -0.09 with $\chi^2/DOF$=4 10$^{-3}$/1. This T is below $T_{lim}$, however the $\epsilon_i$ is difficult to extract due to lack of data. Extrapolating the A+A data in fig. \[ls\_vs\_e\], left, to higher $\epsilon_i$, we find that the approximately netbaryonfree limit of $\lambda_s$ in Au+Au collisions, is expected to be reached at LHC energies [@border].
Conclusions {#conclusions .unnumbered}
===========
The problem we address is how to identify the onset of the QCD phase transition using measured light (u,d,s)-flavoured hadrons, and separate the colliding systems which go through the QCD phase transition from the ones which dont, without any use of the QCD predictions, as well as the role of strangeness. We achieve this by estimating for the first time [@sqm2000; @mapping; @border] the equivalent temperature at zero baryochemical potential of the thermodynamic states describing several measured particle and nuclear collisions. We find at zero chemical potential ($\mu_B$=0) a universal rise and saturation of both the $T$ and of the strangeness suppression factor $\lambda_s$ (=$\frac { 2\overline{s} } { \overline{u} + \overline{d} } $) with increasing initial energy density ($\epsilon_i$) (fig. \[t\_vs\_e\] and fig. \[ls\_vs\_e\], right). The onset of saturation of both $T$ and $\lambda_s$ at $\mu_B=0$, allows to discriminate systems which go through the QCD phase transition from those which dont. The critical energy density is thus estimated at the onset of saturation to be $\epsilon_c$ $\sim$ 1 +0.3 -0.5 GeV/fm$^3$, corresponding approximately to a $\sqrt{s}$ of $\sim$ 8.8 GeV for central Pb+Pb collisions. Further, we identify trivial and non-trivial (with respect to the phase transition) sources of ’strangeness enhancement’: e.g. the maximum of $\lambda_s$ at the 40 A GeV Pb+Pb collisions [@border], and other phenomena of ’strangeness enhancement’, are ’trivially’ traced back to the different baryochemical potentials and $\epsilon_i$ of the compared systems (compare fig. \[ls\_vs\_e\] right with $\mu_B$=0 and left with nonzero $\mu_B$). This explaines why e.g. the strange/pion double ratios in A+A over p+p collisions both at the same $\sqrt{s}$, increase with decreasing $\sqrt{s}$, because $\mu_B$ increases and $\epsilon_i$ differs for A+A and p+p collisions at the same $\sqrt{s}$. It could explain also an enhancement of $\lambda_s$ in p+A over p+p collisions. We redefine the non-trivial ’strangeness enhancement’, as the increase of $\lambda_s$ in all (thermalised) systems which reached $\epsilon_c$ as compared to all (thermalised) systems which did not, at the same $\mu_B$. It is a consequence of the $\epsilon_i$ dependence of $T$. The netbaryonfree $\lambda_s$ limit is estimated to be approximately reached in Au+Au collisions at the LHC [@border]. In conclusion, in contrast to external non-equilibrium signatures like charmonia suppression or jet quenching which are important but may be overcritical, hadrons with $u, d, s, \overline{u}, \overline{d}, \overline{s}$ quarks play a special role as equilibrium signatures and part of the plasma itself, in allowing to extract the critical parameters of the QCD phase transition.
Acknowledgments {#acknowledgments .unnumbered}
===============
We thank the Schweizerischer Nationalfonds for their support, as well as K. Pretzl and P. Minkowski for fruitfull discussions. We also thank the organisers of ISMD2001 for creating an open and scientifically fruitfull atmosphere.
[99]{}
P. Braun-Munzinger, J. Stachel, Nucl. Phys. A606 (1996) 320.
K. Redlich, hep-ph/0105104, proceedings of QM2001.
S. Kabana, P. Minkowski, New J of Phys 3 (2001) 4.
S. Kabana, J. of Phys. G 27, 3, (2001), 497, S. Kabana, hep-ph/0010246.
J D Bjorken, Phys. Rev. D 27 (1983) 140. S. Kabana, New J. of Phys. 3 (2001) 16.
S. Kabana, Eur. Phys. J. C21, 3, (2001), 545, S. Kabana, hep-ph/0105152.
G. Ambrosini et al., (NA52 Coll.): J. of Phys. G27, 3, (2001), 495, hep-ph/0010045, New J. of Phys. 1 (1999) 22, New J. of Phys. 1 (1999) 23, Nucl. Phys. A661 (1999) 370c.
H. Stoecker et al., LBL-12971, (1981).
M. Gazdzicki, D. Röhrich, Z. Phys. C71, (1996), 55.
R. Brandelik et al., (DASP coll.), Nucl. Phys. B 148 (1979) 189.
[^1]: Note that a similar behaviour (rise and saturation) was found in the $\epsilon_i$ dependence of the kaon number density $\rho_k$ [@charm; @centr_na52], which may be related to the behaviour seen in fig. \[t\_vs\_e\].
[^2]: Discussions of the QCD phase transition appearing possibly between AGS and SPS energy can be found e.g. in [@dieter], however not at a common $\mu_B$.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'This paper reports the experimental determination of the [*phase-breaking length*]{} $L_{\phi}$ of conduction electrons in graphene using Raman spectroscopy. Based on the double-resonance model, we extract $L_{\phi}$ from the spatial dependence of the $D$ band susceptibility $\chi_D$ near the graphene edge. By using prior knowledge of sample properties and the excitation point-spread function we are able to determine the spatial variation of the Raman susceptibilities with high accuracy, and the results reveal a [*phase-breaking length*]{} $L_{\phi}$=40nm near the graphene edge.'
author:
- 'Luiz Gustavo Cançado, Ryan Beams, and Lukas Novotny.'
title: 'Optical Measurement of the Phase-Breaking Length in Graphene'
---
The unique electronic properties of graphene are attracting considerable scientific and technological interest [@geim]. Conduction electrons in graphene behave as massless and relativistic two-dimensional Dirac fermions that give rise to an unusual quantum hall effect [@zhang; @novo01; @novo02]. The very high electron mobility at room temperature detected in micron-sized graphene samples (5000-10000cm$^{2}$V$^{-1}$s$^{-1}$)[@novo04; @mean01; @berger02] is already being explored for novel ballistic-transport devices, including graphene-based field-effect transistors (FETs) and single-electron transistors (SETs) [@berger; @geim]. However, the physical properties of nanoscale graphene devices are predicted to be strongly affected by the edges, which act as defects in graphene’s crystal structure [@louis; @denis]. Recent studies of the transport properties of graphene nanoribbon based FET architectures revealed the strong impact of edge states for nanoribbons having widths below 50nm [@Avouris]. Therefore, it is important to understand the symmetry breaking nature of the edges and associated local electronic effects.
The Raman spectrum of graphite is known to exhibit several peaks that are only observed in the presence of structural defects in the hexagonal lattice [@tan; @thomrev; @pimenta]. Of particular interest is the $D$ band ($\sim$1350cm$^{-1}$) that appears close to the first-order allowed peak denoted as $G$ band ($\sim$1580cm$^{-1}$) [@tuinstra01]. Electrons involved in the double-resonance process giving rise to the $D$ band are inelastically scattered by phonons in the interior of the first Brillouin zone. However, because of momentum conservation these phonons can only become Raman active if the electrons involved in the process are elastically scattered by a defect [@thomsen; @saito]. The intensity ratio between the disorder-induced $D$ band and the $G$ band is widely used as a measure for the average crystallite size $L_{a}$ in graphitic systems [@tuinstra01]. Single layers, bilayers and few layers of graphene can be identified by the lineshape of the $G^{\prime}$ band (the second harmonic of the $D$ band occurring at $\sim$2700cm$^{-1}$) [@ferrari02; @eklund; @graf]. Recent Raman scattering studies have also revealed that the electron-phonon interaction in graphene can be affected by electric fields [@castro; @ferrari01; @yan] and that the adiabatic Born-Oppenheimer approximation can be violated [@ferrari01].
In this paper, we use confocal Raman spectroscopy to experimentally determine the [*phase-breaking length*]{} $L_{\phi}$ of conduction electrons in graphene. $L_{\phi}$ is the average distance traveled by an electron before undergoing inelastic scattering with a lattice phonon [@solid01]. We extract $L_{\phi}$ from the spatial dependence of the $D$ band susceptibility $\chi_D$ near the graphene edge. Based on the double-resonance model, $\chi_D$ is directly associated with inelastic scattering of transverse optical phonons near the vertices of the $1^{st}$ Brillouin zone of graphene. Magnetotransport measurements predict a [*phase-breaking length*]{} $L_{\phi}\,\approx$100nm for micron-sized graphene samples at room temperature [@mean01; @berger02]. To the best of our knowledge, a direct optical measurement has not been performed so far.
At first sight, length scales on the order of $L_{\phi}$ cannot be resolved with a diffraction-limited confocal imaging system. However, using prior knowledge of sample properties (edge position) and the excitation (point-spread function) we are able to determine the spatial variation of the Raman susceptibilities with an accuracy of $\approx 10\,$nm. The confocal measurements are further guided by simultaneous topographic imaging with a shear-force atomic-force microscope (AFM) [@shear], and the results reveal a [*phase-breaking length*]{} of 40nm near the graphene edge. This value is shorter than that predicted for micron-sized graphene samples at room temperature [@mean01; @berger02], and shows the influence of the edges on the transport properties of charge carriers in nanoscale graphene systems (such as nanoribbons), where edge defects play an important role [@louis; @denis].
The graphene samples used in this study were prepared by micromechanical cleavage of a piece of highly oriented pyrolytic graphite at the surface of a clean microscope cover glass, following the method reported in reference [@novo03]. Single layer graphene flakes were identified by optical microscopy and subsequent topographic characterization with shear-force microscopy using a chemically etched glass tip attached to a quartz tuning fork. The confocal Raman instrument used in this experiment is based on an inverted optical microscope equipped with an x,y-scan stage. A high numerical aperture objective (1.4NA) is used to focus a linearly polarized laser beam with wavelength 632.8nm on the sample surface. The scattered light is collected using the same objective and then detected with a single-photon counting avalanche photodiode (APD) or a combination of a spectrograph and a cooled charge-coupled device (CCD).
![(a): Topographic image of a graphene layer on a glass substrate. (b)-(d): Confocal Raman images showing the $G$, $G^{\prime}$ and $D$ band intensities over the same region. (e): Raman spectra of a single graphene layer. Top: spectrum recorded near the edge \[white square in (a)\]; bottom: spectrum recorded in the interior \[white circle in (a)\]. The spectra are scaled by a factor of 4 in the range $1250-1700\,cm^{-1}$ .\[fig1\]](fig01.eps){width="38.00000%"}
Figure \[fig1\](a) shows the topographic image of a single graphene layer on a glass substrate. Figures \[fig1\](b)-(d) are corresponding confocal Raman images showing the $G$, $G^{\prime}$ and $D$ band intensities, respectively. The $G$ band is a first-order scattering process that originates from the double degenerate vibrational mode $\Gamma_{6}^{+}$ (E$_{2g}$) that occurs at the crossing of the longitudinal optical (LO) and transverse optical (TO) phonon branches at the $\Gamma$ point in the 1$^{st}$ Brillouin zone of graphene [@tuinstra01]. Notice that the $G$ band intensity is roughly uniform along the graphene surface, as shown in Figure \[fig1\](b). A similar situation occurs for the $G^{\,\prime}$ band, which is the overtone of the $D$ band but does not require a disorder-induced process to become Raman active, since momentum conservation is guaranteed in two-phonon Raman processes [@stokes]. On the other hand, the $D$ band can be detected only near the graphene edges, which act as defects necessary for momentum conservation in the one-phonon double-resonance process involving phonons in the interior of the $1^{st}$ Brillouin zone [@edge].
All confocal Raman images shown in Figures \[fig1\] were recorded with the polarization vector ($\vec{P}_{0}$) of the excitation laser beam oriented parallel to the graphene edge \[$y$-direction in Fig. \[fig1\](b)\]. Polarizations perpendicular to the edge do not generate any $D$ band Raman scattering [@edge]. This is the reason why the $D$ band intensity associated with the top edge in Figure \[fig1\](d) (forming a relative angle of $\sim$60$^{0}$ with $\vec{P}_{0}$) is weaker than that obtained from the other edges. Figure \[fig1\](e) shows Raman scattering spectra acquired at two different locations \[indicated in Fig. \[fig1\](a)\]. The upper spectrum was acquired near the edge of the graphene layer whereas the lower spectrum was recorded $\approx 1\,\mu$m from the edge. The $D$ band appears only in the spectrum acquired near the edge indicating that the graphene sheet is free of structural defects. The Raman scattering spectra also reveal that the $G^{\prime}$ band is composed of a single peak, which confirms that the sample is a [*single*]{} graphene sheet [@ferrari02; @eklund; @graf].
In order to analyze the spatial dependence of the $D$, $G$ and $G^{\prime}$ band intensities near the graphene edge, Raman spectra similar to those shown in Figure \[fig1\](e) were recorded as the position of the incident laser focus was moved in steps of 30nm along a 1.2$\mu$m line perpendicular to the edge \[dotted line in Figure \[fig1\](a)\]. Figure \[fig2\] shows the measurement (dark triangles) of the $G$, $G^{\prime}$, and $D$ band intensities along the selected line. As the graphene edge is moved through the laser focus, the intensities of the $G$ and $G^{\prime}$ bands gradually transit from a minimum value (dark counts) to a maximum value. On the other hand, the $D$ band intensity achieves a maximum value when the graphene edge is in the laser focus. In order to extract the material-specific response functions (Raman susceptibilities $\chi_s$) the Raman intensity measurements have to be deconvolved with the field distribution $\vec{E}({\bf r})$ of the excitation laser (point spread function). This inverse scattering problem is nontrivial and can only be accomplished by multiple sets of measurements (tomography) or by use of prior information.
Because the size of the excitation field is much larger than the electron-phonon scattering length in graphene, the measured Raman intensities can be expressed as incoherent spatial sums, i.e. $$I_s(x,\omega_{s})\propto\,\int^{+\infty}_{-\infty}\!\!\left|\,\tensor{\chi}_s(x^{\prime};\omega_{s},\omega)\,\vec{E}(x^{\prime}\!-\!x,\omega)\right|^{\,2} dx^{\prime} \; ,
\label{eq03}$$ where $x$ designates the lateral position of the laser focus and $s \in \{G,G',D\}$. The $y$-dependence has been eliminated by integration (vertical edge). Using the fact that $\vec{E}$ has negligibly weak polarization components perpendicular to the graphene edge and that the phase distribution in the focal plane is uniform [@novobook] allows Eq. (\[eq03\]) to be expressed in scalar form. The Raman susceptibilities $\chi_s$ represent the local interaction strength between incident light and a specific phonon mode.
In order to solve for $\chi_s$ we require an accurate measurement of the excitation field $\vec{E}$. This task has been accomplished by spin-casting a 1nM solution of nile blue molecules onto the graphene sample and acquiring fluorescence rate images of single molecules with an attenuated laser power ($\approx 200\,$nW). Each molecule maps out the spatial distribution of the excitation intensity ($|\vec{E}|$) [@novobook] and hence renders the excitation profile for Eq. (\[eq03\]). Thus, the excitation field $\vec{E}$ is measured under the same experimental conditions as the intensities $I_s$ of the different Raman lines.
. Triangles indicate experimental data and the red curves are reconstructions according to Eq. (\[eq03\]) using the Raman susceptibilities $\chi_s$ shown as blue curves. (d): Fluorescence intensity profile of a single molecule (nile blue) deposited on the same sample. The inset shows the corresponding confocal fluorescence rate together with the line along which the intensity profile was evaluated. \[fig2\]](fig02.eps){width="38.00000%"}
Figure \[fig2\](d) shows the fluorescence emission intensity of a single molecule along the $x$-direction. The curve corresponds to a line cut through the center of the fluorescence rate image shown in the inset. The red curve is a Gaussian fit according to $\exp[-x^2/\gamma^{2}]$, with $\gamma=186.5\,$nm. It turns out that the fluorescence curve in Fig. \[fig2\](d) is narrower than the profile of the $D$ band intensity shown in Fig. \[fig2\](c), which indicates that the response function $\chi_D$ has finite extent.
In order to proceed we consider the fact that $\chi_s(x)$ vanishes for $x$ smaller than a certain value $x_e$ (the edge) and this value has to be the same for [*all*]{} Raman lines. According to this criterion we use a one-parameter model for the Raman susceptibilities. For the $G$ and $G^{\prime}$ bands we use $$\label{eq05}
\chi_{G,G^{\prime}} (x)= \left\{
\begin{array}{cc}
1- \exp[-(x-x_e)/x_{G,G^{\prime}}] & x \geq x_{e}\\
0\; & x<x_{e}
\end{array}\right.\;,$$ which we insert into Eq. (\[eq03\]), together with the experimentally determined excitation field $\vec{E}$. Numerical integration allows us to find a Raman intensity profile $I_{G,G^{\prime}}$ which can then be compared with the experimental data (dark triangles). Using this procedure, the parameters $x_{G,G^{\prime}}$ and $x_e$ can be solved in a least-squares sense. We find that $x_{G,G^{\prime}}\approx 0$ and hence the susceptibilities $\chi_{G}$ and $\chi_{G'}$ become step functions as indicated by the blue curves in Figs. \[fig2\](a) and (b). The red curves are the calculated fits for $I_{G}$ and $I_{G^{\prime}}$ according to relations (\[eq03\]) and (\[eq05\]). The best fit for the edge position turns out to be $x_{e}$=590nm, which is 10nm displaced from the maximum of the $D$ band intensity curve (Fig. \[fig2\]c) and provides further evidence for the finite width of the $D$ band response function.
For the $D$ band susceptibility we use the following one-parameter model $$\chi_{D}(x)= \left\{
\begin{array}{cc}
\exp[-(x-x_e)/x_D]\ & x \geq x_{e}\\
0 & x<x_{e}
\end{array}\right.\;,
\label{eq06}$$ and follow the same procedure as for the $G$ and $G'$ bands. The best fit is found for $x_{D}$=20nm and the calculated susceptibility according to Eq. (\[eq06\]) is shown as a blue curve in Fig. \[fig2\](c). The red curve is a calculated fit for $I_{D}$ according to relations (\[eq03\]) and (\[eq06\]). The finite value of $x_{D}$ is the central result of this study and proves that the $D$ band response function is not completely localized to the graphene edge.
We now discuss the origin of the spatial extent of the $D$ band susceptibility at the graphene edge. The double-resonance process giving rise to the $D$ band involves the inelastic scattering of a $\pi$ electron in the conduction band by a TO phonon [@ferrpho], whose wavevector $\vec{q}$ lies near the vertices of the first Brillouin zone of graphene ($K$ and $K^{\prime}$ points) [@thomsen; @saito]. Momentum conservation is only satisfied in the scattering process if the electron is elastically back-scattered by a defect providing a wavevector $\vec{d}$$\sim$-$\vec{q}$. In real space, the defect (edge) of our graphene sample is localized in the $x$ direction. Consequently, it is completely delocalized in the reciprocal space along the same direction, thereby providing the necessary condition for momentum conservation in the double-resonance mechanism giving rise to the $D$ band [@edge]. The closer an electron is located to the graphene edge the higher is its probability to be involved in $D$ band scattering.
The [*phase-breaking length*]{} $L_{\phi}$ of a conduction electron is defined as the average distance traveled before undergoing inelastic scattering [@solid01]. Electrons involved in $D$ band Raman scattering undergo a single inelastic scattering event with a lattice phonon. Therefore, $L_{\phi}$ corresponds to the average distance traveled by such electrons during the time interval in which the D band scattering process takes place. Considering that the electron can travel two distances (to the edge and away from it) $L_{\phi}$ is estimated as roughly twice the $1/e$ value of the response function $\chi_D$, that is, $L_{\phi}\,\approx\,40\,$nm \[see Figure \[fig2\](c)\]. Interestingly, this value is smaller than the average value of $\sim$100nm predicted for micron-sized samples at room temperatures [@mean01; @berger02]. We attribute this difference to the fact that the region near the edge presents a higher density of structural defects, causing a reduction of the [*phase-breaking length*]{} [@mean02].
Finally, it is worth noting that, since the frequency and wavevectors of phonons involved in the double-resonance process giving rise to the $D$ band are defined by the incident laser excitation [@thomsen; @saito], the [*phase-breaking length*]{} measured here is associated with a particular location in the phonon dispersion curve. This has clear advantages over standard transport measurements because contributions due to different scattering processes (electron-electron, electron-phonon, and electron-impurity) are not intermixed.
In summary, we have presented an optical measurement of the [*phase-breaking length*]{} of conduction electrons in graphene based on measuring the spatial dependence of the Raman spectrum near the edge of a single graphene layer. We have reconstructed the $D$ band susceptibility $\chi_D$ near the graphene edge by simultaneous measurements of Raman spectra and the point spread function of the incident laser beam. Based on the theory of double-resonance scattering, we associate the mean length of $\chi_D$ with the [*phase-breaking length*]{} of conduction electrons due to inelastic scattering with TO phonons near the vertices of the 1$^{st}$ Brillouin zone of graphene. Using this approach we determine $L_{\phi}$=40nm. This result is also valid for holes in the valence band, and shows the influence of the edges on the transport properties of charge carriers in nanoscale graphene systems where edge defects are inevitably present. An interesting extension of this work will be the simultaneous measurement of Stokes and anti-Stokes $D$ band profiles under different temperature conditions. These measurements would reveal the dependence of the [*phase-breaking length*]{} on the phonon population.
This work was supported by DOE (grant DE-FG02-05ER46207) and NSF (grant CHE-0454704).
[12]{}
A. K. Geim and K. S. Novoselov, Nature Mat. [**6**]{}, 183 (2007).
Y. Zhang, Y. W. Tan, H. L. Stormer, and Philip Kim, Nature [**438**]{}, 201 (2005).
K. S. Novoselov [*et al.*]{}, Science [**306**]{}, 5296 (2004).
K. S. Novoselov [*et al.*]{}, Nature [**438**]{}, 197 (2005).
K. S. Novoselov [*et al.*]{}, Science [**315**]{}, 1379 (2007).
S. V. Morozov [*et al.*]{}, Phys. Rev. Lett. [**97**]{}, 016801 (2006).
C. Berger [*et al.*]{}, Science [**312**]{}, 1191 (2006).
C. Berger [*et al.*]{}, J. Phys. Chem. B [**108**]{}, 016801 (2004).
E. Louis, J. A. Vergés, F. Guinea, and G. Chiappe, Phys. Rev. B [**75**]{}, 085440 (2007).
D. A. Areshkin [*et al.*]{}, Nanoletters [**7**]{}, 204 (2007).
Z. Chen, Y.-M. Lin, M. J. Rooks, and P. Avouris, Physica E, [*in press*]{} (2007).
P. Tan, S. Dimovski, and Y. Gogotsi, Phil. Trans. R. Soc. Lond. A [**362**]{}, 2289 (2004).
S. Reich and C. Thomsen, Phil. Trans. R. Soc. Lond. A [**362**]{}, 2271 (2004).
M. A. Pimenta [*et al.*]{}, Phys. Chem. Chem. Phys [**9**]{}, 1276 (2007).
F. Tuinstra and J. L. Koenig, J. Chem. Phys. [**53**]{}, 1126 (1970).
C. Thomsen and S. Reich, Phys. Rev. Lett. [**85**]{}, 5214 (2000).
R. Saito [*et al.*]{}, Phys. Rev. Lett. [**88**]{}, 027401 (2001).
A. C. Ferrari [*et al.*]{}, Phys. Rev. Lett. [**97**]{}, 187401 (2006).
A. Gupta [*et al.*]{}, Nanoletters [**6**]{}, 2667 (2006).
D. Graf [*et al.*]{}, Nanoletters [**7**]{}, 238 (2007).
A. H. Castro Neto and Fracisco Guinea, Phys. Rev. B [**75**]{}, 045404 (2007).
S. Pisana [*et al.*]{}, Nature Mat. [**6**]{}, 198 (2007).
J. Yan, Y. Zhang, P. Kim, and A. Pinczuk, Phys. Rev. Lett. [**98**]{}, 166802 (2007).
P. L. Taylor and O. Heinonen, [*A Quantum Approach to Condensed Matter Physics*]{} (Cambridge University Press, 2002).
K. Karrai and R. D. Grober, Appl. Phys. Lett. [**66**]{}, 1842 (1995).
K. S. Novoselov [*et al.*]{}, Proc. Natl Acad. Sci. USA [**102**]{}, 10451 (2005).
L. G. Cançado [*et al.*]{}, Phys. Rev. B [**66**]{}, 035415 (2002).
L. G. Cançado, M. A. Pimenta, B. R. A. Neves, M. S. S. Dantas, and A. Jorio, Phys. Rev. Lett. [**93**]{}, 247401 (2004).
L. Novotny and B. Hecht, [*Principles of Nano-Optics*]{} (Cambridge University Press, 2006).
A. C. Ferrari and J. Robertson, Phys. Rev. B [**64**]{}, 075414 (2001).
G. Moos, C. Gahl, R. Fasel, M. Wolf, and T. Hertel, Phys. Rev. Lett. [**87**]{}, 267402 (2001).
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'Entity linking – connecting entity mentions in a natural language utterance to knowledge graph [(KG)]{} entities is a crucial step for question answering over KGs. [It is often based on measuring the string similarity between entity label and its mention in the question.]{} The relation [referred to]{} in the question can help [to]{} disambiguate between entities with the same label. [This can be misleading if an incorrect relation has been identified in the relation linking step]{}. [However, an incorrect relation may still be semantically similar to the relation which the correct entity forms a triple with in the KG; which could be captured by a similarity of their KG embeddings. Based on this idea,]{} we propose the first end-to-end neural network approach that [employs KG as well as]{} word embeddings [to perform joint relation and entity classification of]{} simple questions while [implicitly performing]{} entity [dismabiguation with the help of]{} a novel gating mechanism. [An e]{}mpirical evaluation [shows]{} that the proposed approach achieves [a performance comparable]{} to state-of-the-art entity linking [while requiring less]{} [post-processing]{}.'
author:
- |
Rostislav Nedelchev\
University of Bonn\
Germany\
[[email protected]]{}\
Debanjan Chaudhuri\
University of Bonn & Fraunhofer IAIS\
Germany\
[[email protected]]{}\
Jens Lehmann\
University of Bonn & Fraunhofer IAIS\
Germany\
[[email protected]]{}\
Asja Fischer\
Ruhr University Bochum\
Germany\
[[email protected]]{}\
bibliography:
- 'emnlp-ijcnlp-2019.bib'
title: 'End-to-End Entity Linking and Disambiguation leveraging Word and Knowledge Graph Embeddings'
---
Introduction
============
Question answering is a scientific discipline which aims at automatically answering questions posed by humans in natural language. Simple question answering over knowledge graphs is a well researched topic. [@bordes2015large; @yin2016simpleqaattentivecnn; @mohammed2018strong; @Petrochuk2018SimpleQuestionsNS] A knowledge graph (KG) is a multi-relational graph which represents entities as nodes and relations between those entities as edges. Facts in a KG are stored in form of triples (*h*, *r*, *t*) where *h* and *t* denote the head (also called subject) and tail (also called object) entities, respectively, and *r* denotes their relation. A *simple* question is a [natural language]{} question [(NLQ)]{} that can be represented by a single (subject) entity and a relation. Answering the query then corresponds to identifying the correct entity and relation given in [NLQ]{} and returning the object entity of [the]{} matching triple. For example, for the question *Who is the producer of a beautiful mind?* [the corresponding KG fact is ]{} (a beautiful mind, produced by, Brian Grazer) [and the]{} question answering system should be able to link to the correct entity “a beautiful mind” [(of type]{} movie) and the relation “produced by” in the KG to answer the question [by “Brian Grazer”]{}. The tasks of [identifying]{} the [KG]{} entity [and relation]{} mentioned in the [NLQ]{} [are]{} called entity [and relation]{} linking, [respectively]{}. The former [is often decomposed into]{} two sub-tasks, firstly detecting the span of the entity mention in the [NLQ]{}, secondly to connect [the identified mention]{} to a single entity in the KG, which is usually solved by comparing the entity mention [to]{} the names of the KG entities based on string similarity measures. This [becomes]{} particularly challenging [if]{} there [exist]{} more than one entity in the KG with the same label (name). However, the context provided for the entity in the KG can be used [for disambiguation]{}. In our example, [correct relation linking would identify]{} the relation [“produced by”]{} [as being mentioned in the NLQ. Now one could make use of this information for entity linking by considering only entities which are connected to this specific relation. This would allow]{} to disambiguate and link to the movie “a beautiful mind” rather than the book. This [procedure]{} is called soft disambiguation. However, relation linking is [still]{} challenging since the number of relations in many KGs [is still large]{} (e.g. 6701 in [the]{} FB2M graph subset [of]{} the SimpleQuestions dataset [@bordes2015large]), while suffering from the problem of unbalanced classes [@xuquestion]. Furthermore, [some]{} relations [may be]{} semantically similar, for example, [“fb:film.film.executive\_produced\_by”]{} and [“fb:film.film.produced\_by”]{}; and hence can be confusing for the relation linker. [Therefore, relation linking may end up predicting the wrong relation, which would negatively effect relation based entity disambiguation. To encounter this effect,]{} it seems promising to leverage the relation specific information contained in the KG which is represented by the KG embedding of the relation. Semantically similar relations are closer to each other in KG embeddings vector space. So even if a model is not able to predict the correct relation , the semantic information provided by KG embeddings can be employed to perform soft disambiguation of the entity candidates.
Based on these line of thoughts, we propose a novel end-to-end neural network model for answering simple questions from knowledge graphs , that incorporates both word and KG embeddings. Specifically, the contributions of this paper are as follows:
- The proposal of a novel end-to-end model leveraging relatively simple architectures for entity and relation detection which is comparable to other state-of-the-art approaches for entity linking even without additional post-processing.
- The (to our knowledge) first investigation of incorporating KG embeddings for leveraging KG structures for the end task of entity linking in an end-to-end manner.
- A novel gating mechanism incorporated in the end-to-end architecture which can implicitly perform entity disambiguation if required, improving overall entity linking performance. The final prediction is based on vector similarities, which along with the gate’s output can be interpreted during prediction.
The rest of the paper is organized as follows: In Section 2, we summarize the related works on simple question answering. Sections 3 & 4 provide the background and preliminaries important to this work. The overall approach and the architecture is explained in Section 5. In Section 6, we describe the experiment conditions. Evaluation results are discussed in Section 7. We do an ablation study and result analysis in section 8. Finally, we conclude and state the planned future works in Section 9.
Related Work
============
The SimpleQuestions dataset, as proposed by [@bordes2015large] is the first large scale dataset for simple questions over Freebase. It consists of 108,442 questions split into train(70%), valid(10%), and test(20%). They also proposed an end-to-end architecture using memory networks along with the dataset.
The second end-to-end approach for simple question answering over Freebase was provided by [@golub2016character]. They proposed a character LSTM based question encoder for encoding the question, a CNN based encoder for encoding the [KG]{} entity and relation, and finally an attention-based LSTM decoder for predicting an entity-relation pair given the question. [A similar ent-to-end approach was suggested by]{} [@lukovnikov2017neural]. [It employs Gated Recurrent Unit (GRU) based encoders that work on character and word level and in]{} addition [encode]{} the hierarchical types of relations and entities [to provide further information]{}. [Furthermore, a growing set of]{} modular architectures [was]{} proposed on the [SimpleQuestions]{} dataset. [@yin2016simpleqaattentivecnn] proposed a character-level CNN [for identifying entity mentions]{} and a separate word-level CNN with attentive max-pooling to select knowledge graph tuples. [@yu2017improvedneural] utilized a hierarchical residual bidirectional LSTM for [for predicting a relation]{}, which is then used to re-rank the entity candidates. They replaced the topic entity in the question with a generic token <e> during relation prediction which helps in better distinguishing the relative position of each word compared to the entity.
[@dai2016cfo] proposed a conditional probabilistic framework with bidirectional GRUs that takes advantage of knowledge graph embeddings. [@mohammed2018strong] suggested [to use a]{} combination of relatively simple, component-based approaches that [build on bidirectional GRUs, bidirectional LSTMs (BiLSTMs), and conditional random fields (CRFs)]{} [as well as on]{} graph-based heuristics to select the most relevant entity [given a question from a candidate set. The resulting model]{} provides strong baselines for simple question-answering. More recently, [@huang2019knowledge] proposed an architecture based on [KG]{} embeddings, [@Petrochuk2018SimpleQuestionsNS] proposed a technique [combining]{} LSTM-CRF based entity detection [with Bi]{}LSTM based relation linking where they also replace the topic entity with generic tokens following [@yu2017improvedneural].
[Some other open-domain knowledge graphs are Wikidata[@vrandevcic2014wikidata] and DBpedia [@lehmann2015dbpedia]. In particular, there are two very recent efforts that provided adaptations of the SimpleQuestions dataset [@bordes2015large] to Wikidata [@diefenbach2017question] and DBpedia [@azmy2018farewell]. In addition, there is the Question Answering over Linked Data (QALD) [@unger20166th; @usbeck20177th; @usbeck20188th] series of challenges that use DBpedia as a knowledge base for QA. ]{}
Background
==========
Knowledge graphs, word embeddings, and KG embeddings are concepts that are fundamental to this work. Embeddings [provide]{} a numerical representation [of words and KG entities/relations]{} that facilitate [the incorporation of information provided by]{} KG and language [into]{} neural networks. [Their detailed description follows.]{}
Knowledge Graphs
----------------
In this work, a KG is a network of real world entities that are connected to [each other ]{} by means of relations. Those entities and their relations are represented as nodes and edges respectively in a multi-relational, directed graph. Knowledge graphs consist of ordered triples, also known as facts, of the form $(s, r, o)$, where [$s$ and $o$]{} are two entities connected by the [relation $r$]{}. Revisiting the example from the introduction, [the corresponding fact would be given by]{} (“abeautifulmind” “producedby”, “BrianGrazer”).
Knowledge Graph Embeddings
--------------------------
[Knowledge graphs are data structures ]{} that lack a default numerical representation that allows their straightforward application in a standard machine learning context. Statistical relation learning therefore relies beside other approaches on latent feature models for making prediction over KGs. These latent features usually correspond to embedding vectors of the KG entities and relations. Given the embeddings of the entities and relations of a fact $(h,r,t)$, a score function outputs the a value indicating the probability that the fact is existing in the KG. In this paper, we use TransE [@bordes2013transe] [to learn the KG embeddings used by our model]{}. Let the embedding vector of subject and object entity be given by $\vec h$ and $\vec t$ respectively, and that of the relation by a vector $\vec r$, then the score function of TransE is given by $f(h,r,t) = -\| \vec h+ \vec r- \vec t \|$.
Word Embeddings
---------------
[In recent times, various approaches were proposed that embed words on a vector space [@bengio2003neural; @collobert2008unified]. These methods create representations for each word as a vector of floating-point numbers. Words whose vectors are close to each other are demonstrated to be semantically similar or related. Especially, two works [@mikolov2013efficient; @mikolov2013distributed] showed the high potential of word embeddings in natural language processing (NLP) problems.]{}
[In this work, we use the GloVe [@pennington2014glove] vector embeddings. The method aims to create a $d$-dimensional vector representation of each word in the vocabulary, $w \in V_w$, and the context of each occurrence of a word in the corpus, $c \in V_C$ so that:]{} $$\vec{w} \cdot \vec{c} + b_w + b_c = \log(\# (w,c)) \quad \forall (w,c) \in D$$ [where $b_w$ and $b_c$ are biases that are learned together with $\vec{w}$ and $\vec{c}$ and $\# (w,c)$ is the word occurrence count of word $w$ in context $c$.]{}
$D$ being the documents from which the corpora is extracted from.
Preliminaries
=============
We [employ]{} two kinds of embeddings in the proposed model namely word embeddings and KG embeddings which are defined in the previous section. [For matching a question to the entities and relations of a KG, likely candidates are first selected in an reprocessing step to reduce the enormous number of candidates in the KG. This is described in the following sections.]{}
Entity Candidates Generation {#sub:candgen}
----------------------------
We first start with a simple [language based]{} candidate generation process [selecting potential candidate entities for a given question. That is,]{} given a question we generate candidates by matching the tf-idf vectors of the query with that of the entity labels of all the entities in the knowledge graph, [resulting in]{} a list of $n$ entity candidates $e_c^1, e_c^2.. e_c^n$ for [a given]{} question. This list is then re-ranked based on the tf-idf similarity score, whether the candidate label is present in the question, number of relation the candidate is connected to in the KG and whether it has a direct mapping to wikipedia or not. This is done to give importance to important entities (defined by connectivity in the KG) following previous works [@mohammed2018strong] [@Petrochuk2018SimpleQuestionsNS].
Relation Candidate Generation
-----------------------------
To generate a set of entity specific relation candidates, for each entity $e_c^j$ in the entity candidate set we extract the list of relations connected to this candidate at a 1-hop distance in the Freebase Knowledge Graph. For the $j^{th}$ entity candidate $e^j_c$, the relation candidates are $r_j^1...r_j^n$.
Model Description
=================
{width="65.00000%"}
The proposed neural network model is quintessentially composed of [three parts]{} as visualized in \[fig:Architecture\]:
1. A [word-embedding]{} based **entity span detection model**, which selects the probable [words of an]{} entity in a natural language question, [represented by a bi-LSTM]{}.
2. An [word-embedding]{} based **relation [prediction]{} model** which links the the question to one of the relations in the knowledge graph, [represented by a bi-LSTM with self-attention]{}.
3. [An **entity prediction model** which takes the predictions of the previous two submodels into account and employs a sentinel gating mechanism that performs disambiguation based on similarity measures.]{}
[The different model parts and the training objective of the resulting model will be described in more detail in the following.]{}
Entity Span Detection {#sec:e_span_detect}
---------------------
The span detection module is inspired by [@mohammed2018strong]. The given question is firstly passed through a bi-directional LSTM. [Its output]{} hidden states are then passed through a fully connected layer and a [sigmoid ($\sigma$)]{} activation function which [outputs the probability of]{} the word at time-step $t$ [corresponding to]{} an entity (or not). Mathematically, [this can be described as]{} $$\begin{split}
& h_t = f_{span}(x_t) \enspace ;
o_t = \sigma(Wh_t)
\end{split}$$ where, $o_t$ is the output [probability,]{} $W$ the weight [matrix of the fully connected layer,]{} and $h_t$ the hidden state [vector from]{} the applying the bi-directional LSTM $f_{span}$ on the input question $x_t$. We use I-O encoding for the output for training.
Relation Prediction {#sec:rel_detec}
-------------------
For relation prediction, the question is passed into a self-attention [based]{} bi-directional LSTM [which was]{} inspired by [@zhou2016attentionrl]. The [attention]{} weighted hidden states are then fed into a fully-connected [classification]{} layer [outputting a probability over the $r_n$ relations in the knowledge graph]{}. $$y_r = tanh(W_r(\alpha_r * f_{rel}(x_t)))$$ $W_r$ being a set of model parameters, $\alpha_r$ the self-attention weights and $f_{rel}$ is the Bi-LSTM function which produces a response for every time-step $t$ for the input query $x_t$.
Entity prediction
-----------------
### [Word-based]{} Entity Candidate Selection. {#sec:W_ent_cad_sel}
[With the help of the entity span identified by the span detection submodule described in \[sec:e\_span\_detect\], the questions are now compared to the]{} entity candidates based on vector similarity methods. [More specifically, the word embedding of each word of the question is multiplied with corresponding]{} output [probability]{} from the entity-span detection model [leading to an “entity-weighted” word representation]{} $$e_t = o_t * w_t^{emb}$$ where, $w_t^{emb}$ denotes the word embedding of the $t$-th word in the question and $o_t$ is the sigmoid output from \[sec:e\_span\_detect\]. We [then]{} take a simple average [of the entity-weighted representations of all words of the questions to yield the entity embedding of the question $e_q^{emb}$.]{} Similarly, the entity candidates [$e_c^1, e_c^2.. e_c^n$ generated in the preprocessing step described in \[sub:candgen\] are represented by the word embeddings of their labels $e^{emb1}_c$, $e^{emb2}_c$.. $e^{embn}_c $.]{} [If a label consists of multiple words, the word embeddings are averaged to yield a singe representation.]{} Finally, [to compute the similarity between a question and an entity candidate]{}, the cosine between the question embedding and the entity embedding is estimated. For the $j^{th}$ candidate, [that is]{} $$sim_c^j = cos(e_q^{emb}, e^{embj}_c) \enspace$$ [and the vector $sim_c=(sim_c^1, \dots sim_c^n)$ represents the word based similarity of the question to all entity candidates.]{}
### [KG-based Entity Candidate Selection.]{} {#sec:KG_ent_cad_sel}
[To leverage the relational information encoded in the KG, we firstly take the logits over $r_n$ from the relation prediction model and draw a categorical representation using gumbel softmax [@jang2016categorical]. This representation is multiplied with the KG embeddings over $r_n$ to get a KG embedding based representation of the query $r^{emb}_q$. This relation specific representation is then compared against the full relation candidate set of each candidate entity, where each candidate relation is as well represented by its KG embedding.]{} To match [the relation specific question representation to relation candidates for a given entity,]{} [we estimate the]{} cosine similarity of the [corresponding]{} KG embeddings followed by a max-pooling operation over all the candidate relations of an entity which produces an [entity specific]{} similarity metric $sim_{kg}^j$, [which indicates the degree of matching between the question and an entity candidate from a KG perspective, which specifically takes relation information into account]{}. Mathematically, for the $j$-th entity candidate, [let]{} the embedding [of]{} the $k$-th [relation]{} candidate $r_j^k$ [be]{} denoted by $r^{embk}_j$. [The KG based similarity $sim_{kg}^j$ between the question and the $j$-th entity then given by]{} $$\begin{split}
sim_{kg}^j = maxpool( & cos(r^{emb}_q, r^{emb1}_j), \\
& cos(r^{emb}_q, r^{emb2}_j) \\
& .. \\
& cos(r^{emb}_q, r^{embk}_j)) \enspace,
\end{split}$$ [and the vector $sim_{kg}=(sim_{kg}^1, \dots sim_{kg}^n)$ represents the KG based similarity of the question to all entity candidates.]{}
### [Disambiguation and final prediction.]{}
[The final entity prediction is based on the word- and KG-based similarity measures $sim_{c}$ and $sim_{kg}$ ]{} [First, for disambiguation, the word based similarity vector $sim_c$ is]{} passed into a gating mechanism $$g_{amb} = W_gsim_c \enspace$$ [with]{} $W_g \in \mathbf{R}^{nX1}$, [which aims at estimating if there is more than one single likely candidate in the entity candidate set based on word similarity. If so, the KG based similarity $sim_{kg}$ should also be taken into account, which is done by averaging $sim_{c}$ and $sim_{kg}$ and predicting the final entity candidate by]{} $$\begin{split}
y^e_p =
\sigma( & g_{amb} \space * mean(sim_{kg}, sim_c) \\
& + (1-g_{amb})*sim_c)
\end{split}$$ [Note that,]{} $y^e_p$ $\in$ $\mathbf{R}^n$ are the logits over [the set of candidate entities, from which the entity with the highest probability can be picked.]{} During inferencing, we perform an additional step for ensuring the entity and relation predicted from the model forms a pair in the KG. In order to achieve that, we take the top 5 probable relation from the relation linker and choose the one which is connected to the predicted entity at 1-hop.
[Training]{}
============
[Training objective]{}
----------------------
The model is trained based on a multi-task objective, where the total loss is the sum of the losses from the entity span detection, relation detection, entity candidate prediction, and disambiguation. The individual loss function are given below[^1]. [The loss function for the entity span detection model is the average binary cross entropy $ L_{span} $ over the words of the input question, with]{} $$\begin{split}
& L_{span} = \frac{1}{T}\sum_{t=1}^T {l_t} \\
& l_t = - [ y_t \cdot \log \sigma(o_t) + \\
& \qquad\quad (1 - y_t) \cdot \log (1 - \sigma(o_t)) ] \enspace,
\end{split}$$ where $y_t$ is the label denoting if the $t$-th word belongs to the entity span or not. For relation prediction, a weighted cross-entropy loss $L_{rel}$ is used (where the weights are given by the relative ratio of relations in the training set having the same class as the sample) [and for entity prediction a vanilla cross-entropy loss $L_{ent}$, which depends on the parameters of all sub-models.]{} Furthermore, an additional cross entropy loss function $L_{amb}$ is used to train the gating function. Last but not least, we add an regularization term for soft-parameter sharing following [@duong2015low] resulting in a total loss given by $$\begin{split}
L = & L_{span} + L_{rel} + L_{ent} + L_{amb} \\
& + ||W_{span}^1 - W_{rel}^1||^2 \enspace,
\end{split}$$ where, $W_{span}^1$ and $W_{rel}^1$ are the hidden layer weights of the entity span detection and relation detection module. Given $L$, all parameters of the model are [jointly]{} trained in an end-to-end manner.
[Training details]{}
--------------------
We use the pre-processed data [and word-embeddings]{} provided by [@mohammed2018strong] to train our models. [To obtain]{} KG embeddings, we train TransE [@bordes2013transe] on the provided Freebase KG of 2 million entities. The [size of the]{} word embedding vectors is 300, and that of the KG embeddings is 50. The KG embeddings are kept fixed but the word embeddings are [fine-tuned]{} during optimization. For training the disambiguation gate $g_{amb}$ we use [a label of]{} 1 if the correct entity label is present more than one times in the entity candidates, [and label of]{} 0 otherwise. For training, a batch-size of 100 is used and the model is trained for 100 epochs. We save the model with the best validation accuracy [of]{} entity prediction and evaluate [it on the test set.]{} We [apply Adam [@kingma2014adam] for optimization with a]{} learning rate of 1e-4. The [size of the hidden layer]{} of both the entity span and relation prediction Bi-LSTM is [set to]{} 300. The training process is [conducted]{} on a GPU with 3072 CUDA cores and a VRAM of 12GB.
Evaluation
==========
In this section, we compare our model with other state-of-the-art entity linking and question-answering systems, both end-to-end and modular approaches. [Resources are provided as supplementary materials to this paper that allow the reader to reproduce the final results reported in this section.]{}
Entity-linking
--------------
We compare our end-to-end entity-linking accuracy with other systems whose results are published, on the test set. The results are summarized in Table \[tab:ent\_linking\]. The number of candidates $n$ in the entity candidate set is varied from 100 to 300. The percentage of examples for which the correct entity candidate is present in the candidate set is reported in parenthesis. If the percentage is higher, model performance also increases. The model evaluated over wthe largest entity candidate set (i.e. $n$ = 300) gives the best performance, which is significantly better than the BiLSTM based model [@mohammed2018strong] (13.60% additional accuracy) and the Attentive CNN model (5% additional accuracy). It must be noted that our model cannot be compared directly to the one from [@mohammed2018strong] because they don’t use any candidate information for entity-linking, they do it during final question-answering as a post-processing step. The BiLSTM based model in combination with a n-gram based entity matching and relation based re-ranking suggested by [@yu2017improvedneural] is better than our proposed model by 0.40%.
**** ****
------------------------ ------------------
65.00 (-)
73.60 (-)
**79.00** (-)
Proposed model (n=100) 77.80(92.07)
Proposed model (n=200) 78.35(94.34)
Proposed model (n=300) **78.60**(95.49)
: Entity Linking Accuracy.[]{data-label="tab:ent_linking"}
Question Answering
------------------
The final metric for simple QAKG is defined by the number of correct entities and relations predicted by a given model. We are comparing the performance of our system with that of both end-to-end methods and modular approaches in Table \[tab:qa\]. The results show that the proposed architecture outperforms the state-of-the-art NN based model (GRU based) [@lukovnikov2017neural] by 2.0 %, and shows a performance competitive to simple modular baseline approaches like [@mohammed2018strong] and the KG embedding based approaches KEQA proposed by [@huang2019knowledge]. However, the best state-of-the-art approach on QAKG [@yu2017improvedneural] outperforms ours model by 5.50%. It should be noted here that although our entity-span detection and relation linking accuracy (82.01 %) is better than that of the model proposed by [@mohammed2018strong], the final question answering performance is worse by 1.7 %. This can be explained by the fact that their approach builds on additional string-matching heuristics along with the scores from the different models to re-rank the predicted entities and relations.
**Approach** **Model** **Accuracy(FB2M)**
--------------- ------------------------------------------------------------- --------------------
End-to-End NN Memory NN [@bordes2015large] 61.60
Attn. LSTM [@golub2016character] 70.90
GRU based [@lukovnikov2017neural] [71.20]{}
Proposed model (n=100) 72.29
Proposed model (n=200) 72.84
Proposed model (n=300) **73.20**
Modular BiLSTM & BiGRU [@mohammed2018strong] 74.90
KEQA [@huang2019knowledge] 75.40
CFO [@dai2016cfo] 75.70
CNN & Attn. CNN & BiLSTM-CRF [@yin2016simpleqaattentivecnn] 76.40
BiLSTM-CRF & BiLSTM [@Petrochuk2018SimpleQuestionsNS] 78.10
BiLSTM & Entity-reranking [@yu2017improvedneural] **78.70**
**** ****
---------------------------------------- -------
[Removing $L_{rel}$ from total loss]{} 67.55
74.63
78.17
[Without re-ranking candidates]{} 77.56
Our Best Model (n=300) 78.60
: Ablation Study[]{data-label="tab:ablat"}
Discussion
==========
Ablation Study
--------------
Finally, we do an ablation study where we remove some parts of the proposed model and observe the performance of entity linking for $n$ = 300. The results are in \[tab:ablat\]. As observed, the entity-linking accuracy from not training the relation linker are at par with [@mohammed2018strong] in Table \[tab:ent\_linking\]. The gating mechanism adds 3.97 %, because doing only a mean from the entity and relation prediction similarity scores would add in extra information overhead for the candidate selection for wrongly classified relation. The proposed soft-loss aids in 0.43 % increase in entity-linking accuracy and the candidate re-ranking improves it by 1.04%.
Quantitative and Error Analysis {#sec:heuristics}
-------------------------------
We do a quantitative analysis from the results of our best model with $n$=300. Percentage of questions with soft-disambiguity is 21.1 % and with hard-ambiguity is 18.51 %. Our model is able to predict 84.81 % of correct entity candidates for soft-disambiguation cases, out of which 75.02 % of times the correct relation was identified and 9.78 % the model predicted the wrong relation but the correct candidate is picked using our proposed KG embeddings based method; which proves that our intuition for using KG embeddings for the final task can be beneficial. For hard-ambiguity cases, the model was able to predict the correct candidate with an accuracy of 35.66 % (1432 out of 4015 cases), out of which the model predicted wrong relations 4.4 % of cases. But, it should be noted that there are no explicit linguistic signals to solve hard-disambiguity, following previous works we are predicting these cases based solely on candidate importance.
The model is able to predict the correct candidate 97.70 % of the times for cases where no disambiguation is required. Out of the 440 such wrongly classified candidates, 165 cases are because the true entity and correct relation are not connected in the KG at 1-hop, 162 because the entity span detector was not able to predict the correct span and the rest for wrong prediction in the disambiguate gating mechanism.
In general, some cases where the entity-span detector has failed to identify the correct entity is in table \[tab:span\_err\]. In some of these cases, there are more than 1 entity in the question. Hence, it is difficult for the entity span detector to detect the correct entity.
what ’s a that has been
---------------------------
who is a
what ’s the name of an in
which flight was in an
: Span Detection Error. Green - correct span, blue - detected span.[]{data-label="tab:span_err"}
For the final question-answering task, as mentioned previously, although the end-to-end accuracy for [@mohammed2018strong] is better than ours’, but the task of question answering is particularly challenging in this case because we don’t use any scores from string matching based methods such as Levenshtein distance for entity linking as done as an additional post-processing step by [@mohammed2018strong], especially in cases where the entity candidates and the entity mention in the question consists of out-of-vocabulary words. Also, for some cases, it is challenging to disambiguate between the predicted relations because there are no explicit linguistic signals available. To exemplify, let us consider the question *what county is sandy balls near ?*. The predicted relation relation for this question by our model is [“fb:location.location.containedby”]{} while the true relation in the dataset is [“fb:travel.tourist\_attraction.near\_travel\_destination”]{}.
Conclusion and Future Work
==========================
In this paper, we have proposed an end-to-end model for entity linking, leveraging KG embeddings along with word embeddings banking on relatively simple architectures for entity and relation detection. As reported, the proposed architecture performs better than other end-to-end models but modular architectures demonstrates better question answering performance. However, the purpose of this paper was to integrate KG and word embeddings in a single, end-to-end model for entity linking. Moreover, since the final prediction model is based on similarity scores, the final prediction (and gating) can be easily interpreted following equations 4, 5, 6 and 7.
Error analysis suggest that the model can gain from better entity span detection. As a future work we will experiment by integrating CRF-biLSTM for span-detection and also with more recent NLP models like BERT. The model will also improve with better relation linking and better handling of out-of-vocabulary words. We would also like to integrate more recent state-of-the-art KG embedding models [@dettmers2018conve; @schlichtkrull2018modeling], which can capture better relation semantics in the architecture as a future work.
[^1]: $y$ is used to denote the true label for all tasks here
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'In this paper we study the noncommutativity of a moving membrane with background fields. The open string variables are analyzed. Some scaling limits are studied. The equivalence of the magnetic and electric noncommutativities is investigated. The conditions for equivalence of noncommutativity of the $T$-dual theory in the rest frame and noncommutativity of the original theory in the moving frame are obtained.'
---
-15 mm
.5cm [Davoud Kamani]{} .1cm [*Institute for Studies in Theoretical Physics and Mathematics (IPM)\
P.O.Box: 19395-5531, Tehran, Iran*]{}\
[*e-mail: [email protected]*]{}\
.5cm [*PACS*]{}: 11.25.-w\
[*Keywords*]{}: Noncommutativity; D-brane; T-duality.
Introduction
============
Over the past years there have been attempts to explain noncommutativity on D-brane worldvolume through the study of open strings in the presence of background fields [@1; @2]. From the DBI action it is known that on a D-brane electric field cannot be stronger than a critical electric field, while the same is not true for magnetic fields. Also it is known that Lorentz boosts act on electromagnetic backgrounds. This affects the noncommutativity parameter and the effective open string metric. According to these facts, some properties of D-branes with background fields such as decoupling limits and light-like noncommutativity have been studied [@3].
Previously we have studied the noncommutativity of a moving $Dp$-brane, with the motion along itself [@4]. Now we study the noncommutativity of a moving membrane with electric and magnetic background fields. The motion is parallel or perpendicular to the membrane. For each case, the effective open string variables will be analyzed. We shall observe that for an appropriate magnetic field the open string metric is frame independent. In a special frame the open string metric is proportional to the closed string metric. For both electric and magnetic cases we find decoupling limits, which lead to the definite noncommutative theories. There are frames for the electric and magnetic membranes such that their noncommutativities are proportional to each other.
The effects of $T$-duality on the effective metric and noncommutativity parameter enable us to obtain equivalent noncommutativity structures. That is, we find speeds and background fields for the membrane such that the noncommutativity of the $T$-dual theory becomes equivalent to the noncommutativity of the original theory in the moving frame.
The analysis of Ref.[@2] leads to the definitions of the open string metric $G$, the noncommutativity parameter $\Theta$ and the effective open string coupling constant $G_s$, & & G\^ = (g+2 ’ B)\^[-1]{}g (g-2 ’ B)\^[-1]{} \^,\
& & G\_ = (g-2 ’ B)g\^[-1]{} (g+2 ’ B) \_,\
& & \^ = -(2 ’)\^2 (g+2 ’ B)\^[-1]{} B (g-2 ’ B)\^[-1]{} \^,\
& & G\_s = g\_s \^ , where $g_{\mu \nu}, B_{\mu \nu}$ and $g_s$ are closed string variables.
Note that the effective open string coupling $G_s$ does not change under the Lorentz boosts. Because $g_s$ is the exponential of the scalar field dilation, and the ratio of two determinants also is invariant.
In general a $D2$-brane parallel to the $X^1 X^2$-plane has the NS-NS background $B$-field as in the following B\_ = (
[ccc]{} 0 & E & E’\
-E & 0 & b\
-E’ & -b & 0
), where $\mu , \nu \in \{0,1,2 \}$. We shall discuss pure magnetic and pure electric cases. Let the closed string metric of the membrane worldvolume be g\_ = (
[ccc]{} -g\_0 & 0 & 0\
0 & g\_1 & g’\
0 & g’ & g\_2
).
The paper is organized as follows. In section 2, we study the behavior of effective variables of open string in terms of the background magnetic field and the speed of the membrane. In section 3, the same will be done in the presence of the electric field. In addition, some scaling limits and also equivalence of two noncommutativities will be obtained. In section 4, we study the $T$-duality of the theory and conditions for equivalence of boosted and T-dual noncommutativities.
Magnetic field noncommutativity
===============================
For the pure magnetic field (i.e., $E=E'=0$), the noncommutativity matrix is \^ = (
[ccc]{} 0 & 0 & 0\
0 & 0 & 1\
0 & -1 & 0
), where the parameter $\theta$ (i.e., the strength of the noncommutativity) is defined by & & = - ,\
& & g g\_1g\_2 -g’\^2 . This relation implies that the different magnetic fields $b_\pm = -\frac{1}{2\theta} \pm \frac{1}{2} \sqrt{\frac{1}{\theta^2}
-\frac{g}{\pi^2 \alpha'^2}}$ produce the same noncommutativity on the membrane. Since $g$ is positive, the function $\theta (b)$ has the maximum $\theta_0$ (the minimum $-\theta_0$) at $b=-b_0$, $(b=b_0)$ where b\_0 = , \_0 = . Therefore, to obtain a strong noncommutativity we should adopt $\pm b_0$ magnetic fields.
Now we proceed to study the expressions for the various geometrical quantities in different frames, appropriate to different states of motion of the brane.
Motion along the $X^1$-direction
--------------------------------
Consider the Lorentz transformations on the coordinates $X^0$ and $X^1$, & & X’\^1 = (X\^1 - v X\^0) ,\
& & X’\^0 = (X\^0 - X\^1) , where $\gamma = 1/\sqrt{1-v^2}$. The effect of these transformations on the matrix (4) is ’\^ = (
[ccc]{} 0 & 0 & -\
0 & 0 & 1\
& -1 & 0
). On the other hand, in the moving frame the noncommutativity parameter has non-zero time-like element. Since for a brane with electric field also there are time-like elements (e.g., see the equation (19)), this implies that the motion along the brane directions is equivalent to an appropriate electric field.
The transformation of the open string metric is G’\_ = (
[ccc]{} -\^2 g\_0(1-g\_1 av\^2) & -\^2 v (1-g\_1 a) & v g’a\
-\^2 v (1-g\_1 a) & \^2 (g\_1 a-v\^2) & g’a\
v g’a & g’a & g\_2 a
), where the parameter $a$ is defined by a=1+. Since $g$ is positive we have $a \geq 1$. For $v=0$, this metric reduces to the open string metric for the static membrane with magnetic field.
Let the off-diagonal element $g'$ vanish. In addition, consider the following relation between the magnetic field and the elements of the closed string metric b = . These give the diagonal open string metric G’\_ = [diag]{} (-g\_0 , 1 , g\_2/g\_1), which is independent of the speed of the membrane. Note that the membrane speed has not been hidden in the relation (11).
Under the above conditions we have $\theta = \pm 2\pi \alpha' \sqrt{\frac{1-g_1}{g_2}}$. Therefore, for $v \rightarrow 1$, $g_1 \rightarrow 1$ (i.e., $b \rightarrow 0$ ) and finite $g_0$ and $g_2$ we can introduce the scaling limit = , where $\sigma$ is a finite constant. This implies that & & ’\^ = 2’ (
[ccc]{} 0 & 0 & -\
0 & 0 & 1\
& -1 & 0
),\
& & G’\_ = [diag]{} (-g\_0 , 1 , g\_2). Since all elements of $\Theta'^{\mu \nu}$ and $G'_{\mu \nu}$ are finite, this is a definite noncommutativity.
Motion along the $X^3$-direction
--------------------------------
According to the Lorentz transformations & & X’\^3 = (X\^3 - v X\^0) ,\
& & X’\^0 = (X\^0 - X\^3) , the noncommutativity parameter $\Theta^{\mu \nu}$ does not change, while the open string metric $G_{\mu \nu}$ transforms to G’\_ = (
[ccc]{} -\^2 g\_0 & 0 & 0\
0 & g\_1 a & g’ a\
0 & g’ a & g\_2 a
). From the relativistic point of view, the motion perpendicular to the brane does not change the lengths along the brane. Furthermore, since the noncommutativity due to the magnetic field only has the space-space elements, it is an expected result that $\Theta^{\mu \nu}$ and the space-space elements of $G_{\mu \nu}$ remain unchanged.
For special speeds of the membrane there is $\gamma^2 = a$, or equivalently v = . These speeds are not greater than the speed of light. In these frames the open string metric reduces to G’\_ = a g\_. Therefore, the open string metric is the scaled closed string metric by the scale factor $a$.
Electric field noncommutativity
===============================
Now consider an electric field along the $X^1$-direction i.e., $E' = b=0$. The noncommutativity parameter of this system is \^ = T(E) (
[ccc]{} 0 & g\_2 & -g’\
-g\_2 & 0 & 0\
g’ & 0 & 0
), where the function $T(E)$ is & & T(E) = ,\
& & E\_0 . The electric field $E$ only admits space-time noncommutativity. Since all elements of the noncommutativity parameter are proportional to the function $T(E)$, this factor shows the strength of the noncommutativity.
The effective open string coupling is $G_s = g_s \sqrt{1- \bigg{(}\frac{E}{E_0}\bigg{)}^2}$. According to the root factor there is $-E_0 \leq E \leq E_0$. Therefore, unlike $\theta(b)$, the function $T(E)$ is one to one i.e., each noncommutativity strength only corresponds to one value of the electric field.
Near the electric field $E_0$, open strings do not interact and the strength of the noncommutativity is infinite. Although at the critical electric field $E_0$ the effective theory of the open string is singular and ill-defined, it is possible to define a space-time noncommutative theory by taking an appropriate scaling limit.
Motion in the $X^1$-direction
-----------------------------
The transformations (7) give the noncommutativity parameter as in the following ’\^ = T(E) (
[ccc]{} 0 & g\_2 & -g’\
-g\_2 & 0 & v g’\
g’ & -v g’ & 0
). The open string metric also has the transformation & & G’\_[00]{} = -\^2 g\_0 1-g\_1 v\^2 -1-v\^2 ,\
& & G’\_[01]{} = -\^2 v 1-g\_1 -1- ,\
& & G’\_[02]{} = v g’ ,\
& & G’\_[11]{} = \^2 g\_1- v\^2 --v\^2 ,\
& & G’\_[12]{} = g’ ,\
& & G’\_[22]{} = g\_2. For $v=0$ this metric reduces to the open string metric of the static membrane with electric field.
Near the electric field $E=E_0$ when there is $g' \rightarrow 0$, for any value of the speed $v$, all elements of the transformed metric $G'_{\mu \nu}$, except $G'_{22}$, go to zero G’\_ = 0 G’\_[22]{}, [for]{} g’ 0,E E\_0. To avoid this singularity, we can do the following scaling limit. For $E \rightarrow E_0$, $g' \rightarrow 0$ and finite $g_2$, we should have g’ = ,\^21- = , where $\kappa$ and $\rho$ are finite constants. These imply that the boost velocity approaches to the speed of light, $v \rightarrow \pm 1$. Therefore, the metric (22) takes the form G’\_ = (
[ccc]{} -g\_0(1-g\_1) & -(1-g\_1) &\
-(1-g\_1) & -(1-g\_1) &\
& & g\_2
). All elements of this metric are finite. To restore interactions of open strings, the string coupling $g_s$ can be scaled to infinity i.e., $g_s \sim \gamma$. This leads to a finite $G_s$. The noncommutativity parameter $\Theta'^{\mu\nu}$ near the critical field $E_0$ also should be finite. Therefore, we scale $\alpha'$ to zero as $\alpha' = \mu (1- \frac{E^2}{E_0^2})$, where $\mu$ is finite. In other words, after scaling we have ’\^ = (
[ccc]{} 0 & g\_2 & -\
-g\_2 & 0 &\
& - & 0
), which describes a well defined noncommutativity.
Motion in the $X^3$-direction
-----------------------------
From the transformations (15) we obtain ${\Theta'}^{\mu\nu} = \gamma {\Theta}^{\mu\nu}$, where ${\Theta}^{\mu\nu}$ has been given by the equation (19). The open string metric also becomes G’\_ = (
[ccc]{} -\^2 g\_0 (1- ) & 0 & 0\
0 & g\_1 (1-) & g’\
0 & g’ & g\_2
). Again in the limit $E \rightarrow E_0 $ but arbitrary $v$ and $g'$, we should introduce a scaling limit. Let the elements of the closed string metric $g_0$, $g'$ and $g_2$ be finite. In the limit $E \rightarrow E_0$ we can put $\gamma^2 (1-\frac{E^2}{E_0^2}) = \rho$, which gives G’\_ = (
[ccc]{} -g\_0 & 0 & 0\
0 & & g’\
0 & g’ & g\_2
). In this limit, the speed $v$ approaches to the speed of light such that $\rho$ to be finite.
To have a finite noncommutativity parameter, the parameter $\alpha'$ should go to zero like $\alpha' = \beta(1- \frac{E^2}{E_0^2})^{3/2}$, where $\beta$ is another finite constant. Therefore, we obtain ’\^ = 2 (
[ccc]{} 0 & g\_2 & -g’\
-g\_2 & 0 & 0\
g’ & 0 & 0
).
Since $\alpha'$ goes to zero, according to the definition (20), $E_0$ approaches to infinity. In this limit let $E$ be proportional to $E_0$ as in the following E = E\_0, where $q$ is an infinitesimal number from the interval $-\frac{1}{2 \rho^{1/2}}\leq q \leq \frac{1}{2\rho^{1/2}}$. Using the definition of $\rho$, this equation also can be written in the form $E = q \gamma E_0$. This implies that $q$ goes to zero such that $q\gamma \rightarrow \pm 1$. In other words, the field $E$ approaches to infinity like $\gamma^3$, E = . For this large electric field, the noncommutativity parameters (19) and (29) are related to each other through the equation ’\^\_[scaled]{} = \^\_E. Since there is $q \neq 1$, the above noncommutativities are equivalent but cannot be equal.
Equivalence of the electric and magnetic noncommutativities
-----------------------------------------------------------
Consider two parallel membranes which move along the $X^1$-direction. The first is a $D2$-brane with the speed $v$ and the magnetic field $b$. Its corresponding closed string metric has the elements $(-g_0 , g_1 , g_2=0 , g')$. The noncommutativity of this brane has been given by the equation (8). The second is $D2'$-brane with the speed $v'$, electric field $E$ and the closed string metric elements $(-g'_0 , g_1 , g_2=0 , g')$. This system has the noncommutativity parameter (21).
It is possible to have the relation ’\^\_E = ’\^\_b, where $\eta$ is any real number. This matrix equation leads to the conditions & & = ,\
& & vv’ = , where $\theta(b) = \frac{(2\pi \alpha')^2b}{g'^2-(2\pi \alpha' b)^2}$ and $T(E) = -(2\pi \alpha')^2 \frac{E}{g'_0 g'^2}$ with $-\infty < E< +\infty$. On the other hand, electric field also does not have critical value. The first condition compares the strengths of the noncommutativities. The second condition imposes that $g'_0 > g_0$ and the motions should be in the same direction. For the non-zero magnetic field when $v \rightarrow \pm 1$ we obtain $E \rightarrow \infty $, which is available. For $\eta = 1$ these different systems completely have the same noncommutativity structure.
$T$-duality and equivalence of noncommutativities
=================================================
In the case of toroidal compactification, when $d$-spatial coordinates are compactified on torus $T^d$, the $T$-duality group is $O(d,d; {\bf Z})$ [@5]. Assume that a $Dp$-brane is wrapped on torus $T^p$. Under the action of a particular element of $O(p,p; {\bf Z})$ $T$-duality group i.e., T = (
[cc]{} [**0**]{} & [**1**]{}\_[p p]{}\
[**1**]{}\_[p p]{} & [**0**]{}
), the background fields have the transformations [@6] (g+2’ B) ( [g]{}+ 2’ [B]{})=(g+2’ B)\^[-1]{}. According to the equations (1) and (36) we obtain & & G\^ = [g]{}\^,\
& & \^ = (2 ’ )\^2 [B]{}\^. That is, the open string metric and the noncommutativity parameter appear as the background fields of the $T$-dual theory of string theory. One can show that the effects of $T$-duality transformations on the open string metric $G^{\mu\nu}$ and on the noncommutativity parameter $\Theta^{\mu\nu}$ are as in the following & & [G]{}\^ = g\^,\
& & \^ = (2’)\^2 B\^. Therefore, the background fields of string theory appear as the effective metric and noncommutativity parameter of the effective theory of the $T$-dual theory [@7].
Now we find the background fields and the speed of the membrane such that noncommutativity in the moving frame be equivalent to the noncommutativity of the $T$-dual theory in the rest frame i.e., ’\^ = - \^, where $\lambda$ is a positive constant. This means that, $T$-duality also can act as Lorentz transformations and vice-versa. In other words, noncommutativity parameter in the moving frame is proportional to the background field $B^{\mu \nu}$ in the rest frame. The equations (38) and (39) give the following table for various values of $g'$, $v$ and $B_{\mu\nu}$,
$ $ ${\rm magnetic\;and }\;v_1$ $ {\rm magnetic\;and}\;v_3$ ${\rm electric\;and}\;v_1$ ${\rm electric\;and}\;v_3$
-------------- -------------------------------------------------------------- ------------------------------------------------------------- ---------------------------------------- ---------------------------------------------
$g'$ $g'$ $g'$ $0\;\;\;\;\;\bigg{|}\;\;\;\;g'$ $g'$
$v$ $0$ $v$ $v\;\;\;\;\bigg{|}\;\;\;\;0$ $v$
$B_{\mu\nu}$ $ b=\pm \frac{1}{2\pi \alpha'}\sqrt{(\frac{1}{\lambda}-1)g}$ $b=\pm \frac{1}{2\pi \alpha'}\sqrt{(\frac{1}{\lambda}-1)g}$ $E=\pm E_0 \sqrt{1-\frac{1}{\lambda}}$ $E=\pm E_0 \sqrt{1-\frac{\gamma}{\lambda}}$
$G_s$ $\frac{g_s}{\sqrt{\lambda}}$ $\frac{g_s}{\sqrt{\lambda}}$ $\frac{g_s}{\sqrt{\lambda }}$ $g_s \sqrt{\frac{\gamma}{\lambda }}$
where $v_i$ shows the membrane motion along the $X^i$-direction.
When the equation (39) holds, to find the corresponding $G'_{\mu \nu}$, we should use the equations (9), (16), (22) and (27) and the values of $g'$, $v$, $b$ and $E$ of this table. For example for the magnetic part of the table, the space-space elements of the metric $G'_{\mu \nu}$ are proportional to $g_{\mu \nu}$ with the scale factor $\frac{1}{\lambda}$. For the special case that has been given by the equation (18) we have $\lambda=\frac{1}{\gamma^2}$. According to the equation (38), the equation (18) can be written as G’\_ = \_. That is, on the open string metric, Lorentz transformations also act as $T$-duality.
Noncommutativity parameter in the moving frame is $\Theta'^{\mu \nu}=
-(2\pi \alpha')^2 \lambda B^{\mu \nu } $. For the magnetic case ${\Theta'}^{\mu \nu }$ is similar to the matrix (4), that $\theta$ should be replaced with the factor $\mp 2\pi \alpha' \sqrt{\frac{\lambda -\lambda^2}{g}}$. This factor is the strength of the noncommutativity, and for $\lambda = \frac{1}{2}$ has extremum $\mp \frac{\pi \alpha'}{\sqrt{g}}$. For the electric case $\Theta'^{\mu \nu}$ is given by the matrix (19) in which the function $T(E)$ should be changed with the factor $\pm \frac{1}{g_2 E_0} \sqrt{\lambda^2 - \eta \lambda}$. For the motion along the $X^1$ and $X^3$ directions $\eta$ is 1 and $\gamma$, respectively.
Note that the equation $\Theta^{\mu\nu} =
- \lambda ({\tilde \Theta})'^{\mu\nu}$, for the membrane with magnetic field, produces the results of the equation (39). The matrix $({\tilde \Theta})'^{\mu\nu}$ is noncommutativity parameter of the $T$-dual theory in the moving frame.
Conclusions
===========
For the moving membranes with electric or magnetic background fields, we studied the effective variables of open string. For the magnetic membrane we observed the followings. There are two values of the magnetic field that produce the same noncommutativity on the membrane. By choosing an appropriate magnetic field, the open string metric becomes independent of the speed of the membrane. When this magnetic field goes to zero and near the speed of light we obtained a well defined noncommutativity. For the special speeds perpendicular to the membrane, the open string metric is proportional to the closed string metric.
For the pure electric field on the membrane, the strength of the noncommutativity in terms of the electric field is one to one. When the electric field approaches to its critical value, we obtained some definite noncommutative theories from decoupling limits. We found the conditions that the noncommutativity structures of the moving electric and magnetic membranes to be equivalent. Similarly the equivalence of the electric noncommutativities in the rest frame and in the moving frame (with the scaling limit form) was obtained.
The background fields of string theory ($T$-dual of string theory), are effective metric and noncommutativity parameter of the effective $T$-dual theory (the effective theory of string theory). Therefore, we observed that for the special background fields and speeds of the membrane, the noncommutativity of the $T$-dual theory in the rest frame appears like the noncommutativity of the original theory in the moving frame. The open string metric in the moving frame also is equivalent to that one of the $T$-dual theory in the rest frame.
[99]{} A. Connes, M.R. Douglas and A. Schwarz, JHEP [**9802**]{}(1998)003, hep-th/9711162; M.R. Douglas and C. Hull, JHEP [**9802**]{}(1998)008, hep-th/9711165; F. Ardalan, H. Arfaei and M.M. Sheikh-Jabbari, JHEP [**9902**]{}(1999)016, hep-th/9810072; Nucl. Phys. [**B576**]{}(2000)578, hep-th/9906161; A. Fayyazuddin and M. Zabzine, Phys. Rev. [**D62**]{}(2000)046004, hep-th/9911018; P.M. Ho and Y.T. Yeh, Phys. Rev. Lett. [**85**]{}(2000)5523, hep-th/0005159; Y.E. Cheung and M. Krog, Nucl. Phys. [**B528**]{}(1998)185, hep-th/9803031; D. Bigatti and L. Susskind, Phys. Rev. [**D62**]{}(2000)066004, hep-th/9908056; A. Schwarz, Nucl. Phys. [**B534**]{}(1998)720, hep-th/9805034; C.S. Chu and P.M. Ho, Nucl. Phys. [**B550**]{}(1999)151, hep-th/9812219; Nucl. Phys. [**B568**]{}(2000)447, hep-th/9906192; V. Schomerus, JHEP [**9906**]{}(1999)030, hep-th/9903205; D. Kamani, Phys. Lett. [**B548**]{} (2002)231, hep-th/0210253; Mod. Phys. Lett. [**A17**]{}(2002)2443, hep-th/0212088. N. Seiberg and E. Witten, JHEP [**9909**]{}(1999)032, hep-th/9908142. N. Seiberg, L. Susskind and N. Toumbas, JHEP [**0006**]{}(2000)021, hep-th/0005040; R. Gopakumar, J. Maldacena, S. Minwalla and A. Strominger, JHEP [**0006**]{}(2000)036, hep-th/0005048; J.L.F. Barbon and E. Robinovici, Phys. Lett. [**B486**]{}(2000)202, hep-th/0005073; J. Gomis and T. Mehen, Nucl. Phys. [**B591**]{}(2000)265, hep-th/0005129; O. Aharony, J. Gomis and T. Mehen, JHEP [**0009**]{}(2000)023, hep-th/0006236; R.G. Cai and N. Ohta, JHEP [**0010**]{}(2000)036, hep-th/0008119; G.H. Chen and Y.S. Wu, Phys. Rev. [**D63**]{}(2001)086003, hep-th/0006013. D. Kamani, Europhys. Lett. [**57**]{}(2002)672, hep-th/0112153. K.S. Narain, H. Sarmadi and E. Witten, Nucl. Phys. [**B279**]{}(1987)369. A. Giveon, M. Porrati and E. Robinovici, Phys. Rep. [**C244**]{}(1994)77-202, hep-th/9401139; R.G. Leigh, Mod. Phys. Lett.[**A4**]{} 28(1989)2767. D. Kamani, Mod. Phys. Lett. [**A17**]{}(2002)237, hep-th/0107184.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'Repeated interaction between individuals is the main mechanism for maintaining cooperation in social dilemma situations. Variants of tit-for-tat (repeating the previous action of the opponent) and the win-stay lose-shift strategy are known as strong competitors in iterated social dilemma games. On the other hand, real repeated interaction generally allows plasticity (i.e., learning) of individuals based on the experience of the past. Although plasticity is relevant to various biological phenomena, its role in repeated social dilemma games is relatively unexplored. In particular, if experience-based learning plays a key role in promotion and maintenance of cooperation, learners should evolve in the contest with nonlearners under selection pressure. By modeling players using a simple reinforcement learning model, we numerically show that learning enables the evolution of cooperation. We also show that numerically estimated adaptive dynamics appositely predict the outcome of evolutionary simulations. The analysis of the adaptive dynamics enables us to capture the obtained results as an affirmative example of the Baldwin effect, where learning accelerates the evolution to optimality.'
author:
- |
Shoma Tanabe$^{1}$ and Naoki Masuda$^{2,3,*}$\
\
$^1$ Faculty of Engineering, University of Tokyo\
7-3-1 Hongo, Bunkyo, Tokyo 113-8656, Japan\
$^2$ Graduate School of Information Science and Technology\
University of Tokyo, 7-3-1 Hongo, Bunkyo, Tokyo 113-8656, Japan\
$^3$ PRESTO, Japan Science and Technology Agency\
4-1-8 Honcho, Kawaguchi, Saitama 332-0012, Japan\
\
$^*$ Corresponding author ([email protected])\
title: Evolution of cooperation facilitated by reinforcement learning with adaptive aspiration levels
---
Introduction
============
The mechanisms of cooperation in social dilemma situations are a central topic in interdisciplinary research fields including evolutionary biology, ecology, economics, and sociology. As analyzed by the prisoner’s dilemma (PD) game and its relatives, direct reciprocity is among the main known mechanisms underlying cooperative behavior [@Trivers1971; @Axelrod1984book]. In direct reciprocity, iterated interaction between the same individuals motivates them to continue cooperating (C) rather than to defecting (D) to obtain momentarily large payoffs; defection would be negatively rewarded by the opponent player’s retaliation in later rounds. Variants of the celebrated retaliatory strategy tit-for-tat (mimicking the opponent’s action in the previous round) [@Nowak1992Nature_gtft] and a win-stay lose-shift strategy [@Kraines1989; @Nowak1993Nature] are recognized as strong competitors in the iterated PD game.
In the iterated games concerning direct reciprocity, it is natural to assume that players modify their strategies in response to their experiences in past rounds. The tit-for-tat, its variants, and win-stay lose-shift strategies can be interpreted as examples of such learning strategies because the tit-for-tat, for example, implies that the player selects the action (i.e., C or D) depending on the result of the last round. A more sophisticated learning player of this kind exploits a longer history of the game for action selection (e.g., cooperate if the player and the opponent cooperated in the previous two rounds, and defect otherwise) [@Lindgren1991ALife]. Classes of other learning models include fictitious play and reinforcement learning [@Camerer2003book; @Fudenberg1998book]. Learning apparently seems beneficial in iterated games because learning players are more flexible than nonlearning players.
If learning is a key factor in promoting cooperation in real societies, the number of learning players should increase when a population evolves under selection pressure. However, the advantage of learners over nonlearners in evolutionary dynamics is elusive because a pair of learning players often results in mutual defection [@Macy1996SMR; @Sandholm1996BioS; @Posch1999RoyalB; @Taiji1999PhysicaD; @Macy2002pnas; @MasudaOhtsuki2009BMB] and learning may be costly.
The constructive roles of learning in the evolution of certain traits are collectively called the Baldwin effect (see Simpson, 1953; Turney *et al.*, 1996; Weber & D. J. Depew, 2003; Crispo, 2007; Badyaev, 2009 for reviews). Although earlier examples of the Baldwin effect are not necessarily founded on firm empirical evidence [@Simpson1953Evol; @Weber2003book], there exists a plethora of positive evidence of the Baldwin effect. Examples include fly’s morphological developments [@Waddington1942Nature], colonization of house finch in North America [@Badyaev2009PTRSB], and persistence of coastal juncos [@Yeh2004AmNat]. In fact, the concept of the Baldwin effect differs by authors (see Simpson, 1953; Downes, 2003; Turney *et al.*, 1996). Although earlier computational models suggest that learning accelerates evolution [@Hinton1987CS; @Ancel1999JTB; @Maynard-Smith1987Nature], later theoretical and numerical studies suggest that learning either accelerates or decelerates evolution toward the optimum depending on the details of the models [@Ancel2000TPB; @Dopazo2001BMB; @Borenstein2006; @Paenke2007AmNat; @Paenke2009AI]. The advantage of learning in evolution is also nontrivial in this broader context.
We numerically investigate the effect of learning on evolution in the iterated PD game. This question was explored in previous literature (Suzuki & Arita, 2004; also see Wang *et al.*, 2008 for discussion). Our emphasis in this study is to use a reinforcement learning model for the iterated PD game [@MasudaNakamura2011JTB] that is much simpler in terms of the number of plastic elements than the plastic look-up-table model adopted in [@Suzuki2004BioS]. In our model, players are satisfied with and persist in the current action when the obtained payoff is larger than a plastic threshold. Our model of players introduced in [@MasudaNakamura2011JTB] modifies those in [@Karandikar1998JET; @Posch1999RoyalB; @Macy2002pnas]. Via the stability analysis for nonlearning players, the numerical analysis of the discretized adaptive dynamics with nonlearning and learning players, and full evolutionary simulations, we show that learning is needed for a noncooperative population to evolve to be able to engage in mutual cooperation for wide parameter ranges. We also discuss our results in the context of the Baldwin effect.
Model
=====
Iterated PD game {#sec:iterated PD}
----------------
We assume that each player plays the PD game against each of the other players in a population. In each round $t\ (t=1,\ 2,\ \ldots)$ within a generation, a player selects C or D without knowing the action (i.e., C or D) of the opponent player. The payoff to the focal player is defined by $$\begin{aligned}
\bordermatrix{
&{\rm C}&{\rm D}\vspace{-2mm}\cr
{\rm C}& R & S \vspace{-2mm}\cr
{\rm D}& T & P
},
\label{eq:payoff matrix}\end{aligned}$$ where $T>R>P>S$ and $R>(T+S)/2$. Equation represents the row player’s payoff. The payoff to the opponent (column player) is defined likewise; the PD game is symmetric. Because $T>R$ and $P>S$, mutual defection is the only Nash equilibrium of the single-shot PD game.
However, players may continue mutual cooperation for their own benefits in the iterated PD game [@Trivers1971; @Axelrod1984book]. We denote the number of rounds per generation by $t_{\max}$. Technically, the Nash equilibrium of the iterated PD game is perpetual mutual defection if the players know $t_{\max}$ beforehand. The number of rounds is often randomized to avoid this effect [@Axelrod1984book]. To simplify the analysis, we assume that the players are unaware of the fixed value of $t_{\max}$.
Earlier studies identified tit-for-tat, which involves imitating the previous action of the opponent, as a strong strategy in the iterated PD game when various strategies coexist in a population [@Trivers1971; @Axelrod1984book]. However, later studies showed that tit-for-tat is not robust against error and that alternative strategies such as generous tit-for-tat [@Nowak1992Nature_gtft] and Pavlov [@Kraines1989; @Nowak1993Nature] are strong competitors in the iterated PD game with error. By definition, a Pavlov player receiving payoff $T$ or $R$ is satisfied and does not change the action in the next round, whereas the same player receiving payoff $P$ or $S$ is dissatisfied and flip the action. A population composed of Pavlov players, for example, realizes mutual cooperation such that a player gains approximately $R$ per round.
Reinforcement learning {#sec:reinforcement learning}
----------------------
Intuitively, the ability to learn may seem to be an advantageous trait in the iterated PD game if the cost of learning is negligible. However, this is generally not the case. A pair of learning players often ends up with mutual defection unless a learning algorithm is carefully designed [@Macy1996SMR; @Sandholm1996BioS; @Posch1999RoyalB; @Taiji1999PhysicaD; @Macy2002pnas; @MasudaOhtsuki2009BMB]. Learning requires trial and error, i.e., the exploration of unknown behavioral patterns as well as the exploitation of known advantageous behavioral patterns. Exploratory behavior of a learning player may look just random to opponents, and it is rational to defect against random-looking players.
To compromise the possibility of mutual cooperation, the simplicity of the learning algorithm, and the biological plausibility of the model as compared to some other learning algorithms, we use a variant of the Bush–Mosteller (BM) reinforcement learning model [@MasudaNakamura2011JTB]. This model modifies the models in the previous literature [@Karandikar1998JET; @Posch1999RoyalB; @Macy2002pnas] such that players learn to mutually cooperate for wide parameter ranges.
In round $t$, the cooperability of the learning player is given by the probability $p_t$. We update $p_t$ using the results of the single-shot PD game as follows: $$\begin{aligned}
p_{t+1}=\left \{
\begin{array}{ll}
p_t+(1-p_t)s_t & ({\rm action\ in\ round\ } t {\rm\ is\ C,\ and\ }s_t \geq 0),\vspace{-2mm}\cr
p_t+p_ts_t & ({\rm action\ in\ round\ } t {\rm\ is\ C,\ and\ }s_t < 0), \vspace{-2mm}\cr
p_t-p_ts_t & ({\rm action\ in\ round\ } t {\rm\ is\ D,\ and\ }s_t \geq 0),\vspace{-2mm}\cr
p_t-(1-p_t)s_t & ({\rm action\ in\ round\ } t {\rm\ is\ D,\ and\ }s_t < 0),
\end{array}
\right .
\label{eq:p_t update}\end{aligned}$$ where $$s_t=\tanh [\beta(r_t-A_t)],
\label{eq:s_t update}$$ and $r_t \in \{R,T,S,P\}$ is the payoff to the player in round $t$. $s_t$ stands for the degree of satisfaction in round $t$. When $s_t$ is large, the player increases the probability of taking the current action in round $t+1$. For example, the third line in Eq. indicates that the player decreases the probability of cooperation $p_t$ because selecting D in round $t$ has yielded a satisfactory outcome. In addition, we assume that the player misimplements the action with a small probability $\epsilon$ such that the player in fact cooperates with probability $(1-2\epsilon)p_t+\epsilon$ in round $t$. Equations and indicate that the player is satisfied with the current situation if the obtained payoff $r_t$ is larger than the so-called aspiration level $A_t$. Otherwise, the player is motivated to flip the action. $\beta$ controls the sensitivity in the plasticity of $p_t$. If $\beta=0$, $s_t=0$ for any $t$ such that $p_t$ is constant. If $\beta=\infty$, $s_t=1$ or $-1$ for any $t$ such that $p_{t+1}=1$ or 0.
Unless otherwise stated, we set the initial condition to $p_1=0$, i.e., the player defects in round 1. This value of $p_1$ is the most adverse to mutual cooperation. We will confirm in Secs. \[sec:Nash h=0\] and \[sec:evolutionary simulations\] that our main results are qualitatively the same if we set $p_1=1$.
The dynamics of the aspiration level are given by $$A_{t+1}=(1-h)A_t+hr_t,
\label{eq:A_t update}$$ where $h$ represents the learning rate. If $h=0$, $A_t$ is constant, and the model is equivalent to the classical BM model. If $h=1$, the player compares the current payoff and the payoff obtained in the last round to determine $s_t$. In our previous work, we showed that mutual cooperation is established among the players only after $t_{\max}=100$ rounds if $\beta$ is large and $h$ is small [@MasudaNakamura2011JTB]. In the numerical simulations, we set $\beta=3$, which is large enough to support mutual cooperation if other conditions, such as small $h$ and small $\epsilon$, are met.
We remark that the initial condition $A_1$ is a key parameter to characterize the player.
Evolutionary dynamics {#sec:evolutionary dynamics}
---------------------
We set the number of players in the population to $N=500$. In a single generation, each player $i$ plays the iterated PD game with $t_{\max}=200$ against all the other players. We always reset $p_t$ and $A_t$ to $p_1$ and $A_1$ when a player starts the iterated PD game with a new opponent. The single generation payoff $\overline {r}_i$ $(\in [S,T])$ is equal to the summation of the payoff obtained by playing against $N-1$ players, which is divided by $(N-1)t_{\max}$.
After the single generation payoffs to all the players are determined, we select two players $i$ and $j$ with equal probability for strategy update. We use the Fermi rule [@Szabo1998; @Traulsen2006] in which player $i$ adopts $j$’s $A_1$ and $h$ values in the next generation with probability $1/\left[1+\exp \left(\tilde \beta({\overline r}_i-{\overline r}_j) \right)\right]$, and player $j$ adopts $i$’s parameter values, otherwise. We set $\tilde \beta=1$. To account for mutation, we assume that after strategy update, $A_1$ and $h$ of the adopter are displaced by random small values obeying the uniform density on $[-\Delta_{A_1},\Delta_{A_1}]$ and $[-\Delta_h,\Delta_h]$, respectively. If the displaced $h$ exceeds 1 or is negative, we reset $h$ to 1 or 0, respectively. However, the resetting seldom occurs in our evolutionary simulations.
The phenotype of a player in round $t$ is specified by $p_t$ and $A_t$. It should be noted that $p_t$ and $A_t$ are not inherited over generations. In other words, the natural selection operates on the capacity to learn (i.e., $h$) but not on the acquired behavior (i.e., $p_t$ and $A_t$). Because we let $\beta$ in Eq. to be relatively large to realize mutual cooperation [@MasudaNakamura2011JTB], $p_t$ is sensitive to the excess payoff relative to $A_t$ in the sense that $p_t$ is close to $0$ or $1$ unless $r_t$ is close to $A_t$. Therefore, $p_t$ is similar to the probability of cooperation conditioned on the outcome of the PD game in the previous round. When we use the term learning in the following, we exclusively refer to that induced by $h$ in the iterated PD game. A positive value of $h$ directly raises the plasticity of $A_t$ and indirectly controls that of $p_t$.
Results
=======
Nash equilibria when without learning {#sec:Nash h=0}
-------------------------------------
To show that learning is necessary for the emergence of cooperation, we start by analyzing the competition between players that do not learn. With learning rate $h$ equal to zero, the aspiration level is fixed over rounds (i.e., $A_t=A_1$, $t \geq 1$). For the sake of analysis, we set $\beta = \infty$. Then, Eqs. and imply that the player persists in the current action (i.e., $(p_t,p_{t+1})= (0,0)$ or $(1,1)$) if $r_t-A_t \geq 0$ and flips the action (i.e., $(p_t,p_{t+1})=(0,1)$ or $(1,0)$) otherwise. When $A_t$ is fixed, there are five strategies:
- Strategy st1 is defined by $A_t<S$. Except for the action misimplemantation, an st1 player always cooperates or always defects, depending on the action in the first round.
- Strategy st2 is defined by $S<A_t<P$. An st2 player does not flip the action unless $r_t=S$.
- Strategy st3 is defined by $P<A_t<R$. An st3 player does not flip the action if mutual cooperation or unilateral defection is realized. It is equivalent to Pavlov, which is a strong competitor in the iterated PD game [@Kraines1989; @Nowak1993Nature].
- Strategy st4 is defined by $R<A_t<T$. An st4 player flips the action unless $r_t=T$.
- Strategy st5 is defined by $T<A_t$. An st5 player flips the action in every round except when the player misimplements the action.
In Table \[table:tab1\], the average payoff to a nonlearning (i.e., $h=0$) player (row player) playing against another nonlearning player (column player) is shown for $0<\epsilon \ll 1$ and $t_{\max} = \infty$. For example, st1 playing against st2 obtains $(R+3P+2S)/6$ per round on an average. The results shown in Table \[table:tab1\] are a subset of those obtained in [@Nowak1995JMB] (see Appendix A for details). Table \[table:tab1\] indicates that st3 is a Nash equilibrium when the five strategies are considered. In particular, st3 playing against another st3 realizes mutual cooperation and obtains the largest average payoff per round $R$. Therefore, a unanimous population composed of st3 players represents a eusocial situation. Mutual cooperation is not realized by any other combination of two players.
Table \[table:tab1\] indicates that st2 is also a Nash equilibrium when $3P>R+2S$. In addition, although st4 is not a Nash equilibrium, a homogeneous population composed of st4 players is resistant to invasion by st3 in evolutionary situations because st4 gains a larger payoff than st3 does when playing against an st4 opponent.
To test the robustness of the results shown in Table \[table:tab1\], we set $\beta=3,\ \epsilon=0.02$, and $h=0$, and numerically calculate the payoff averaged over $t_{\max}$ generations to different nonlearning players with different fixed aspiration levels $A_1$. We also set $R=4,\ T=5,\ S=0,$ and $P=2$ in this and the following numerical simulations. The average payoff to a nonlearner playing against another nonlearner is shown for $p_1=0$ and $t_{\max}=200$ in Fig. \[fig:nl vs nl\]. The presented values are averages over 100 trials for each pair of $A_1$ values. The results shown in Fig. \[fig:nl vs nl\] are qualitatively the same as those shown in Table \[table:tab1\]. We also confirmed that the results hardly change for $(p_1,t_{\max})=(0,2000),\ (1,200),$ and $(1,2000)$.
Possibility of mutual cooperation via learning {#sec:mutual C}
----------------------------------------------
If $h>0,$ players different from st3 may adjust $A_t$ until $P<A_t<R$ is satisfied such that they learn to behave as Pavlov. Therefore, learning may play a constructive role in the evolution of mutual cooperation. In fact, this is not always the case; $\epsilon>0$ is a necessary condition for mutual cooperation to evolve.
To explain this point, we set $\beta=3$ and $h=0.1$, and numerically examine the behavior of a pair of players. Typical time courses of the aspiration level for a pair of learning players over rounds without action misimplementation (i.e., $\epsilon=0$) are shown in Fig. \[fig:l vs l\](a). Each of the three pairs with close $A_1$ values represents a pair of st1 (thick lines), st3 (dotted lines), and st5 players (medium lines), respectively. We used different values of $A_1$ for each pair for the clarity of the figure; making $A_1$ equal for two players does not qualitatively change the results. The thick lines in Fig. \[fig:l vs l\](a) indicate that the two st1 players playing with each other are satisfied with payoff $P=2$ obtained by mutual defection. Therefore, their aspiration levels converge to $A_t=P$. The results would be the same if we start from a pair of st2 players or a combination of an st1 player and an st2 player. A pair of st3 players begin mutual cooperation from the second round, and their $A_t$ values converge to $R=4$ (dotted lines). Mutual cooperation is also realized if the two players are initially either st4 or st5, although some rounds are required before the players mutually cooperate (medium lines).
Although two learning players having $A_t<P$ do not end up with mutual cooperation when $\epsilon=0$, the action misimplementation (i.e., $\epsilon>0$) can trigger a shift from mutual defection to mutual cooperation. Artificially generated time courses in the presence of action implementation are shown in Fig. \[fig:l vs l\](b) for expository purposes. Until the intended action is misimplemented ($1 \leq t \leq 29$ in Fig. \[fig:l vs l\](b)), two players starting with $A_1<P$ keep mutual defection (thick lines). When $A_t$ has sufficiently approached $P$, we assume that one player misimplements the action ($t=30$). Then, the $A_t$ values of both players cross $P$ from below within a couple of rounds such that the players start to behave as Pavlov and mutually cooperate. The possibility of mutual cooperation through this mechanism is sensitive to the value of $h$. Two players starting with $A_1<P$ end up with $A_t>P$ owing to the action misimplementation when $(P-S)/(T-P)<(1-h)^2$ (see Appendix B for derivation). When $T=5,\ S=0$, and $P=2$, this condition yields $0<h<1-\sqrt{2/3} \approx 0.184$ for an arbitrary value of $R$. Then, the $A_t$ values of the two players converge to $R$.
The $A_t$ values also converge to $R$ when we start with a pair of st3 players (dotted lines in Fig. \[fig:l vs l\](b)) and a pair of st4 or st5 players (medium lines in Fig. \[fig:l vs l\](b)). This is because mutual cooperation is stable against action misimplementation; if one player turns into D by action misimplementation in round $t$, both players defect in round $t+1$ and cooperate in round $t+2$, if the actions are not misimplemented in rounds $t+1$ and $t+2$. This event sequence is likely unless $\epsilon$ is large.
Adaptive dynamics {#sec:adaptive dynamics}
-----------------
In the evolutionary numerical simulations that we will describe in Sec. \[sec:evolutionary simulations\], we allow the initial aspiration level $A_1$ and learning rate $h$ to mutate (Sec. \[sec:evolutionary dynamics\]). If the distribution of $A_1$ and that of $h$ for an evolving population are single peaked and sufficiently localized, we can grasp the evolutionary dynamics for a population by tracking the dynamics of the population averages of $A_1$ and $h$, denoted by $\overline {A_1}$ and $\overline h$, respectively. In the extreme case in which all the players share identical values of $A_1$ and $h$, the instantaneous dynamics of $\overline {A_1}$ and $\overline h$ are captured by adaptive dynamics [@Metz96chapter; @Hofbauer1998book; @Hofbauer2003BAMS; @Doebeli2004Science]. Adaptive dynamics reveal the possibility for mutants with a slightly deviated parameter value to invade a homogeneous resident population. In this section, we numerically examine two-dimensional adaptive dynamics with respect to $\overline {A_1}$ and $\overline h$ to foresee the evolutionary simulations carried out in Sec. \[sec:evolutionary simulations\].
In this and the following sections, we set $\beta=3,\ p_1=0,\ \epsilon=0.02$, and $t_{\max}=200$ unless otherwise stated. Consider a homogeneous population of players sharing the parameter values $\overline {A_1}=A_1$ and $\overline h=h$. A mutant player with aspiration level $A_1^{\prime}$ and learning rate $h^{\prime}$ can invade the population if $$\pi[s^\prime,s]-\pi[s,s]>0
\label{eq:pi ieq1}$$ or $$\pi[s^\prime,s]-\pi[s,s]=0\ {\rm and}\ \pi[s^\prime,s^\prime]-\pi[s,s^\prime]>0,
\label{eq:pi ieq2}$$ where $s=(A_1,h)$ and $s^\prime =(A_1^\prime,h^\prime)$ are the strategies of the resident and mutant players, respectively, and $\pi[s_1,s_2]$ represents the average payoff of strategy $s_1$ when playing with strategy $s_2$. $s$ is ESS if the converse of Eq. or the converse of Eq. is satisfied. If Eq. or is satisfied, the homogeneous population comprising strategy $s$ would evolve toward $s^\prime$. We numerically calculate $\pi[s^\prime,s]-\pi[s,s]$, where $s^\prime=(A_1+0.2,h),\ (A_1-0.2,h),\ (A_1,h+0.02),$ and $(A_1,h-0.02)$. We confine $s^\prime$ in the neighborhood of $s$ because the amount of mutation for $A_1$ and $h$ is assumed to be small. Examining $\pi[s^\prime,s]-\pi[s,s]$ corresponds to looking at the discretized adaptive dynamics, i.e., the discretized derivative of $\pi[s^\prime,s]$ with respect to $s^\prime$ at $s^\prime=s$.
For various values of $A_1$ and $h$, $\pi[s^\prime,s]-\pi[s,s]$ is shown in Fig. \[fig:adaptive dynamics\]. The plotted values are averages over $10^4$ runs for any $s$. In Fig. \[fig:adaptive dynamics\](a), $s^\prime=(A_1+0.2,h)$ obtains a larger payoff than $s=(A_1,h)$ in the red region. In this region, $s^\prime$ would invade a homogeneous resident population of $s$ such that $\overline{A_1}$ increases. In contrast, $s^\prime$ obtains a smaller payoff than $s$ does in the blue region. Figure \[fig:adaptive dynamics\](a) indicates that, if learning is prohibited (i.e., $h=0$), the population starting from $A_1=0$, for example, is expected to evolve such that $\overline{A_1}$ increases, but only up to $A_1 \approx P=2$. Therefore, a population does not evolve from st2 to st3 without learning.
Figure \[fig:adaptive dynamics\](b), which reveals the possibility of invasion by mutant $s^\prime=(A_1-0.2,h)$ in the resident population of $s$, is a sign flipped version of Fig. \[fig:adaptive dynamics\](a) in most parameter regions. Nevertheless, neither the mutants with $A_1^\prime=A_1+0.2$ nor the ones with $A_1^\prime=A_1-0.2$ invade the resident population (i.e., parameter regions colored in blue in both Figs. \[fig:adaptive dynamics\](a) and \[fig:adaptive dynamics\](b)) for $(A_1,h) \approx (2,0)$ and along a bent line passing through $(A_1,h) \approx (4,0)$ and $(A_1,h) \approx (4.3,0.2)$. These regions constitute singular points of the adaptive dynamics and serve as repellers. In other words, $\overline{A_1}$ does not pass through $\approx P=2$ for $h=0$ and $\approx R=4$ for various values of $h$ in adaptive dynamics. The observations for $h=0$ that the homogeneous population of st2 is not invaded by st3 mutants, that of st3 is not invaded by st2 or st4 mutants, and that of st4 is not invaded by st3 mutants, are consistent with the results obtained in Sec. \[sec:Nash h=0\].
The possibility of invasion by mutant $s^\prime=(A_1,h+0.02)$ in the homogeneous population of $s=(A_1,h)$ is shown in Fig. \[fig:adaptive dynamics\](c). The figure suggests that $\overline h$ would increase for a population of st1 players (i.e., $A_1<S=0$). Learning is preferred to nonlearning when $A_1<0$ for the following reason. As shown in Sec. \[sec:mutual C\], when $h>0$ and $\epsilon>0$, $A_t$ increases until the players behave as Pavlov to mutually cooperate within a relatively small number of rounds (Fig. \[fig:l vs l\](b)). In contrast, the players do not establish mutual cooperation when $h=0$ or $\epsilon=0$, as shown in Sec. \[sec:Nash h=0\] (Fig. \[fig:l vs l\](a)). Figure \[fig:adaptive dynamics\](c) indicates that $\overline h$ increases up to $\overline h \approx 0.15$. This value of $\overline h$ is consistent with the upper bound of $h$ for mutual cooperation to be possible, which was derived in Sec. \[sec:mutual C\]. Based on these results, $\overline h$ is expected to initially increase in evolutionary dynamics starting with a population of nonlearning st1 players. We refer to the stage of evolutionary dynamics in which $\overline h$ increases as stage 1. The existence of stage 1 is also supported by Fig. \[fig:adaptive dynamics\](d) in which the mutant has $s^\prime=(A_1,h-0.02)$.
After $\overline h$ has increased, Figs. \[fig:adaptive dynamics\](a) and \[fig:adaptive dynamics\](b) imply that $\overline{A_1}$ increases to cross $P=2$. When $h>0$, a larger value of $A_1\ (<P)$ is beneficial because fewer rounds are required for such players to turn to Pavlov (i.e., $A_t>P$). Once $A_1$ exceeds $P$ for a majority of players, they earn a large average payoff $\approx R$ through mutual cooperation. We refer to the transition for learning players from a small $A_1$ corresponding to st1 or st2 to a large $A_1$ corresponding to st3 as stage 2. Figures \[fig:adaptive dynamics\](a) and \[fig:adaptive dynamics\](b) indicate that the difference between $\pi[s^\prime,s]$ and $\pi[s,s]$ when $A_1$, $A_1+0.2<P$, $h>0$ is small, presumably because $s$ and $s^\prime$ are only slightly different in terms of the number of transient rounds before the entrance to $A_1>P$. Therefore, we expect that stage 2 occurs slowly in evolutionary dynamics.
Although it is a minor phenomenon as compared to stages 1 and 2, a smaller $h$ is more beneficial on the boundary between st2 and st3 (i.e., $A_t \approx P=2$), as shown in Figs. \[fig:adaptive dynamics\](c) and \[fig:adaptive dynamics\](d). For expository purposes, time courses of the iterated PD game between an st2 player and an st3 player are shown in Fig. \[fig:adaptive dynamics\](e). As shown by the solid lines, the initial st3 player flips to st2 before establishing mutual cooperation if $h>0$. In fact, a nonlearning st3 player (i.e., $h=0$) realizes mutual cooperation with a learning st2 player in earlier rounds (dotted lines) than a learning st3 player does (solid lines). Therefore, in evolutionary dynamics, $\overline h$ in the vicinity of $A_t \approx P$ is expected to decrease. We refer to this transition as stage 3. It should be noted that stage 3 occurs in a narrow range of $A_1$ (i.e., $A_1<P$ and $A_1+0.2>P$ in Fig. \[fig:adaptive dynamics\](a)).
Through stages 1, 2, and 3, evolution from a defective population of nonlearning st1 players to a cooperative population of st3 players is logically possible. In contrast, the emergence of mutual cooperation is hampered if learning is prohibited.
After stage 3, $\overline {A_1}$ would not evolve beyond $R$; $A_1 \approx R$ is a line of repellers in adaptive dynamics, as already explained in Figs. \[fig:adaptive dynamics\](a) and \[fig:adaptive dynamics\](b). When $P<A_1<R$ (i.e., $2<A_1<4$), the mutant’s payoff is indistinguishable from the resident’s payoff unless $h$ is large (Figs. \[fig:adaptive dynamics\](a)–(d)). Therefore, $\overline {A_1}$ and $\overline h$ would perform approximately unbiased diffusion. This implies that $\overline h$ that has decreased via stage 3 may increase again.
When at least one of the two players is st4 or st5, a player with a larger $A_1$ is more advantageous than the opponent with a smaller $A_1$. This is because the former exploits the latter in early rounds. Nevertheless, these players do not obtain the average payoff as large as that for a pair of st3 players, which would start to mutually cooperate from the second round. Therefore, st3 is stable against invasion by st4 and vice versa.
We predict that the learning rate would not eventually decrease to the small value in evolutionary simulations. In other words, the disadvantage of learning is too small to be evolutionarily relevant unless the cost of learning is explicitly incorporated.
To assess the robustness of the results obtained from the adaptive dynamics, we reproduced Figs. \[fig:adaptive dynamics\](a)–(d) with $\epsilon=0.05$ and $\epsilon=0.1$. The results for $\epsilon=0.05$ are qualitatively the same as those for $\epsilon=0.02$ (results not shown). The results for $\epsilon=0.1$ are different in some aspects from those for $\epsilon=0.02$ (Fig. \[fig:eps=0.1\]). Most notably, when $\epsilon=0.1$, st3 is no longer stable against invasion by st2 even without learning (i.e., $h=0$). Therefore, mutual cooperation would not be stable in evolutionary dynamics. In Figs. \[fig:tmax eps\](a) and \[fig:tmax eps\](b), $\pi[s^\prime,s]-\pi[s,s]$ is shown for $(s,s^\prime)=((1.9,0),(2.1,0))$ and $(s,s^\prime)=((2.1,0),(1.9,0))$, respectively, for a variety of values of $t_{\max}$ and $\epsilon$. Figure \[fig:tmax eps\](a) indicates that an st3 mutant does not invade the population of st2 residents for all the examined values of $t_{\max}$ and $\epsilon$. Figure \[fig:tmax eps\](b) indicates that a population of st3 residents is resistant to invasion by st2 mutants when $t_{\max}$ is large and $\epsilon$ $(>0)$ is small. Nevertheless, st3 is stable for various values of $t_{\max}$ and $\epsilon$. Because stage 2 is hampered when $\epsilon=0$, $\epsilon$ must take an intermediate value for the learning-mediated mutual cooperation to emerge.
Evolutionary simulations {#sec:evolutionary simulations}
------------------------
The results in Sec. \[sec:adaptive dynamics\] predict the presence of a learning mediated evolutionary route from a noncooperative population composed of st1 players to a cooperative population composed of st3 players. In this section, we carry out direct numerical simulations of the evolutionary dynamics using a population composed of $N=500$ players. We initially set $h=0$ and select $A_1$ for each player independently from the uniform density on $[S-1,S]$. Therefore, all the players are initially nonlearning st1. Refer to Sec. \[sec:evolutionary dynamics\] for details of the numerical setup.
The evolution of $\overline h$, the total amount of plasticity experienced in a generation, defined by $\sum_{t=1}^{t_{\max}-1} \left| A_{t+1}-A_t \right|$, $\overline r$, and the fraction of mutual cooperation for an example run with $\Delta_{A_1}=0.05$ and $\Delta_h=0.01$ are shown in Fig. \[fig:evolutionary dynamics\](a). The average learning rate $\overline h$ and the total amount of plasticity rapidly increase until the $\approx 3.9\times10^4$th round. The payoff and the fraction of mutual cooperation also increase during this period because st1 players learn to behave as Pavlov when $h>0$. This period corresponds to stage 1 described in Sec. \[sec:adaptive dynamics\]. Then, the fraction of mutual cooperation and the total amount of plasticity gradually increase until the $\approx 3.5\times10^5$th round, corresponding to stage 2. In the $\approx 3.5\times10^5$th round, an st3 mutant emerges in the population mostly composed of st2 players and gains a larger payoff than st2 residents do. Then, st3 players rapidly replace st2 players in the population such that $\overline r/R$ and the fraction of mutual cooperation suddenly increase (Fig. \[fig:evolutionary dynamics\](a)). This is because stable mutual cooperation between st3 players emerges in an early round, whereas that between st2 players emerges after $\approx 2/\epsilon$ rounds. The learning rate decreases almost at the same time, corresponding to stage 3. The time courses of the fractions of st1, st2, st3, st4, and st5 players corresponding to the run shown in Fig. \[fig:evolutionary dynamics\](a) are shown in Fig. \[fig:evolutionary dynamics\](b). For example, the fraction of the st1 player is defined by the fraction of players having $A_1<S$ and any value of $h$. Figure \[fig:evolutionary dynamics\](b) indicates that the population initially composed of st1 players evolves to that of st3 players. The trajectory of $\overline {A_1}$ and $\overline h$ corresponding to the same run is shown in Fig. \[fig:evolutionary dynamics\](c). Figure \[fig:evolutionary dynamics\](c) is consistent with the scenario of the evolution of cooperation described in Sec. \[sec:adaptive dynamics\]. The population evolves from no cooperation to mutual cooperation via the three stages involving learning. After stage 3, $\overline {A_1}$ and $\overline h$ diffuse without a recognizable bias, which is also consistent with the results obtained in Sec. \[sec:adaptive dynamics\] (white regions in Figs. \[fig:adaptive dynamics\](a)–(d)). However, it should be noted that the total amount of plasticity remains small after stage 3.
To examine the robustness of the results, we carry out five runs of numerical simulations for each of the different parameter sets; we could not carry out more extensive numerical simulations because of the computational cost. We measure two quantities in each run. The first quantity is the number of generations necessary for $\overline h$ to exceed $0.1$ for the first time. We call this number the end of stage 1. The second quantity is the number of generations necessary for $\overline {A_1}$ to exceed $P$ for the first time. We call this number the end of stage 2. The ends of stages 1 and 2 with $\epsilon=0.02$, $\Delta_{A_1}=0.05$, and $\Delta_h=0.01$ are equal to $257\pm28\ (\times10^2)$ and $3763\pm1514\ (\times10^2)$, respectively, where the mean $\pm$ standard deviation on the basis of the five runs are indicated. Those with $\epsilon=0.05$, $\Delta_{A_1}=0.05$, and $\Delta_h=0.01$ are equal to $299\pm55\ (\times10^2)$ and $6951\pm4231\ (\times10^2)$. Those with $\epsilon=0.01$, $\Delta_{A_1}=0.05$, and $\Delta_h=0.01$ are equal to $265\pm31\ (\times10^2)$ and $3780\pm1547\ (\times10^2)$. Those with $\epsilon=0.02$, $\Delta_{A_1}=0.02$, and $\Delta_h=0.01$ are equal to $239\pm18\ (\times10^2)$ and $9151\pm6836\ (\times10^2)$. For this parameter set, one out of the five runs did not reach the end of stage 2 within $2\times10^6$ generations, such that the statistics are based on the other four runs. Those with $\epsilon=0.02$, $\Delta_{A_1}=0.05$, and $\Delta_h=0.005$ are equal to $613\pm119\ (\times10^2)$ and $3504\pm999\ (\times10^2)$. Mutual cooperation evolves via learning (i.e., finite value of the end of stage 2 up to our numerical efforts) in most cases. When $\epsilon=0.05$, evolution to mutual cooperation is slower than when $\epsilon=0.02$. This may be because learning players having different values of $A_1$ turn into Pavlov (i.e., $A_t>P$) within a small number of rounds when $\epsilon$ is relatively large. Then, the payoff to different learning players would differ relatively little to weaken the selection pressure.
We perform another robustness test. For the original parameter values $\epsilon=0.02$, $\Delta_{A_1}=0.05$, and $\Delta_h=0.01$, the trajectory of $\overline {A_1}$ and $\overline h$ obtained from a single run with $p_1=1$ is shown in Fig. \[fig:p\_1=1\]. The results are qualitatively the same as those for $p_1=0$ (Fig. \[fig:evolutionary dynamics\](c)) although establishment of cooperation takes a considerably larger number of generations when $p_1=1$ than when $p_1=0$.
Baldwin effect {#sec:Baldwin effect}
--------------
If we assume an explicit cost of learning, the learning rate decreases after mutual cooperation is reached. An example time course of $\overline {A_1}$ and $\overline h$ when a linear cost $-ch$ is added to the single generation payoff to each player (Suzuki & Arita 2004; see Ancel 1999, 2000 for a different implementation of the explicit learning cost), where $c=1$, is shown in Fig. \[fig:cost\](a). The final value of $\overline h$ is smaller than that in the case without the learning cost (Fig. \[fig:evolutionary dynamics\](c)). The result shown in Fig. \[fig:evolutionary dynamics\](d) is an example of the standard Baldwin effect in which the learning rate initially increases and then decreases [@Ancel2000TPB; @Dopazo2001BMB; @Borenstein2006; @Paenke2007AmNat; @Paenke2009AI].
We examine the robustness of the observed Baldwin effect against the variation of $c$. An example time course of $\overline {A_1}$ and $\overline h$ when $c=10$ is shown in Fig. \[fig:cost\](b). As compared to when $c=1$ (Fig. \[fig:cost\](a)), $\overline h$ is smaller throughout the evolution, and $\overline {A_1}$ increases more slowly. Nevertheless, the population mostly consists of st3 players in the end. We carried out five runs for various values of $c$. The largest $\overline h$ value in $10^6$ generations is shown in Fig. \[fig:cost\](c) for each run. The largest $\overline h$ value decreases with $c$ because learning is costly for a large value of $c$. The final value of $\overline h$, calculated as the average over the last $10^4$ generations, is shown in Fig. \[fig:cost\](d). The final value of $\overline h$ is considerably smaller than the largest value (Fig. \[fig:cost\](c)) for each $c$, indicative of stage 3 of the Baldwin effect. The final value of $\overline {A_1}$, calculated as the average over the last $10^4$ generations, is plotted against $c$ in Fig. \[fig:cost\](e). If this value is larger than $P=2$ and smaller than $R=4$, we expect that the final population is mostly composed of st3 players and that the Baldwin effect is operative. Figure \[fig:cost\](e) suggests that the Baldwin effect occurs in the five runs when $c<15$. When $c\ge 15$, stage 1, i.e., the initial increase in $\overline h$, is often too small in magnitude such that stage 2 does not sometimes occur. We conclude that the Baldwin effect occurs for a wide range of $c$.
Discussion
==========
We have shown that reinforcement learning promotes the evolution of mutual cooperation in a population of players involved in the iterated PD game. Cooperation evolves under some conditions such as $3P>R+2S$, positive but not too large values of $\epsilon$, and $t_{\max}$ that is not too small. The present study is motivated by previous investigations of the Baldwin effect. Our results provide an example of the Baldwin effect in the form of a computational model of social behavior.
To understand the behavior of our model analytically, writing down the Fokker–Planck equation for the joint density of $A_1$ and $h$ may be useful. Starting from the singular density at a small value of $A_1$ and $h=0$, we may be able to solve the Fokker–Planck equation numerically to track the evolution of the joint density to find the Baldwin effect. Alternatively, discretizing $A_1$ and $h$ and then formulating a Markov chain on the discretized states may also be useful. Nevertheless, we refrained from such analyses because we consider that they eventually necessitates some numerical simulations and would not sufficiently advance the understanding of our numerical results.
The concept of the Baldwin effect is diverse [@Simpson1953Evol; @Downes2003chap2; @Turney1996EvolComput]. However, arguably, the most accepted variant of the Baldwin effect is formulated as a two-stage mechanism [@Simpson1953Evol; @Godfrey2003chap3; @Crispo2007Evol; @Turney1996EvolComput]. In stage 1, plasticity increases because plastic individuals are better at finding the optimal behavior than nonplastic individuals. In stage 2, mutation makes the optimal behavior innate and decreases the plasticity of individuals. Mutants that play optimally from the outset of their life without plasticity and resident individuals that acquire the optimal behavior through plasticity are eventually equally efficient. Nevertheless, because of the cost of learning, the mutants overwhelm the residents via natural selection. Stage 2 is often called genetic assimilation.
Stage 1 in our model corresponds to stage 1 of the standard Baldwin effect outlined above. In stage 2 in our model, $A_1$ increases such that the optimal behavior (i.e., mutual cooperation by turning into st3) becomes innate. Nevertheless, after stage 3 in our model in which the learning rate rapidly decreases, the learning rate starts to perform a random walk because the learning cost is marginal in our model (Fig. \[fig:evolutionary dynamics\](c)). Therefore, the behavior of our model in stages 2, 3, and onward does not qualify as stage 2 of the standard Baldwin effect in which the learning rate decreases. With a modified model with an explicit learning cost, we showed that the learning rate decreases after stage 3 (Fig. \[fig:evolutionary dynamics\](d)). In this case, our model naturally fits the framework of the Baldwin effect.
In a previous computational model of the Baldwin effect, learning rates remain large when the optimal behavior dynamically changes owing to environmental fluctuations [@Ancel1999JTB]. In our model without an explicit cost of learning, the learning rate remains large for a different reason. In our model, the optimal parameter set (i.e., $A_1$ and $h$) does not fluctuate after sufficient generations. Instead, approximate optimality is realized for various parameter sets, i.e., any $P<A_1<R$ and $h \geq 0$. Therefore, the learning rate performs a random walk to occasionally visit large values (Fig. \[fig:evolutionary dynamics\](c)).
Godfrey–Smith points out three alternative reasons why stage 1 cannot be skipped in the two-stage mechanism of the Baldwin effect [@Godfrey2003chap3]. First, learning may provide a breathing space by which a population can survive long enough to transit to stage 2. This reason is irrelevant to our model because our model is not concerned with the survival of the population. The population size is fixed in our model such that the population always survives. Second, the preferred state may be accessible for learners but not for nonlearners. Although not explicitly stated in Godfrey–Smith (2003), this mechanism seems to be relevant to cases in which the fitness landscape does not depend on the configuration of the population. In our case, however, the fitness landscape depends on the fractions of the different types of players because the payoff to a player is affected by the strategies of the other players. Third, evolution may change the “social ecology” of the population such that learners are more advantageous than nonlearners, a phenomenon called niche construction in a broad sense. The social ecology implies a fitness landscape that depends on the configuration of the population. In our model, the social ecology evolves via learning of players. This third mechanism seems to be relevant to our model. Suppose a hypothetical population comprising st1 nonlearners except two st1 learners. For a focal st1 learner, the social ecology is such that there is one st1 learner and $N-2$ st1 nonlearners. If $\epsilon>0$, the focal st1 learner is likely to gain a payoff that is larger than an st1 nonlearner because the focal player learns to mutually cooperate with the other st1 learner, whereas an st1 nonlearner does not. The focal st1 learner would not overwhelm st1 nonlearners if the other st1 learner is absent in the social ecology.
The main purpose of this study is to provide an evolutionary model of concrete social behavior in which learning plays a constructive role. We are not the first to achieve this end. Suzuki and Arita observed the Baldwin effect in the iterated PD game using different learning models [@Suzuki2004BioS]. In their model, the learning rate is assumed to be binary, and the player’s strategy is specified by a look-up table that associates the action to take (i.e., C or D) with the actions of the previous two rounds of the two players. The entries of the look-up table dynamically change when the plasticity is in operation. They also considered the effects of meta–learning in which the player adapts how to update each entry of the look-up table. The main contribution of the present work relative to theirs is to provide a much simpler model in terms of the number of plastic parameters. In contrast, the learning rates and the range of parameters are continuous in our model, whereas they are mostly binary in their model. Our model may be amenable to real animals and facilitates a mechanistic understanding of evolutionary dynamics by the numerically calculated adaptive dynamics. Apart from the fixed parameters common to all the individuals, our players only have two parameters that are plastic within a generation, $p_t$ and $A_t$, and two parameters inherited across generations, $A_1$ and $h$. The results obtained from the adaptive dynamics predict those of direct evolutionary numerical simulations and provide an intuitive reason why learning promotes the emergence of mutual cooperation. In particular, we showed the necessity of the evolution of learning ability for cooperation by explicitly comparing the cases with and without learning. The combination of adaptive dynamics and evolutionary simulations may also be useful for analyzing the Baldwin effect in different models.
Acknowledgements {#acknowledgements .unnumbered}
================
We thank Reiji Suzuki and Kohei Tamura for the valuable discussions. N.M. acknowledges the support provided through Grants-in-Aid for Scientific Research (Nos. 20760258 and 23681033, and Innovative Areas “Systems Molecular Ethology”(No. 20115009)) from MEXT, Japan.
Appendix A: Payoff to nonlearning players {#appendix-a-payoff-to-nonlearning-players .unnumbered}
=========================================
Nowak *et al.* (1995) analyzed iterated matrix games between a pair of players that select an action (i.e., C or D) in response to the actions of the two players in the previous round. There are four combinations of the actions of the two players in the previous round, i.e., (C, C), (C, D), (D, C), and (D, D). Because a player assigns C or D to each of these possible outcomes in the previous round, there are 16 strategies $S_i$ $(0 \leq i \leq 15)$. In fact, st1, st2, st3, st4, and st5 in the present study are equivalent to $S_{12},\ S_8,\ S_9,\ S_1,$ and $S_3$ in [@Nowak1995JMB], respectively.
By calculating the steady state of the Markov chain with four states $R$, $T$, $S$, and $P$, Nowak *et al.* (1995) calculated the average payoff to focal player $S_i$ playing against the opponent $S_j$ $(0 \leq i,\ j \leq 15)$ under a small probability of error in action implementation. Their assumption for the action misimplementation is slightly different from ours. We assumed that $\epsilon$ is the probability that each player independently misimplements the action, whereas only one of the two players may misimplement the action in a round in their model. Nevertheless, our model is equivalent to theirs in the limit $\epsilon \to 0$ if we set $\epsilon^{\prime}=2 \epsilon (1-\epsilon)$, where $\epsilon^{\prime}$ is the probability of action misimplementation in the sense of Nowak *et al.* (1995). Therefore, our results shown in Table \[table:tab1\] are a corollary of their results.
Appendix B: Upper bound of $h$ for st2 players to turn into Pavlov {#appendix-b-upper-bound-of-h-for-st2-players-to-turn-into-pavlov .unnumbered}
==================================================================
Given $\beta=\infty,\ 0<\epsilon \ll 1,\ h>0$, and $A_1<P$, $A_t$ of the two players, denoted by X and Y, are sufficiently close to $P$ when one player, which we assume to be Y without loss of generality, misimplements the action to select C for the first time in round $t \propto 2/\epsilon$. Without any further action misimplementation, X keeps D and Y flips to D in round $t+1$ because $A_{t+1}^{({\rm Y})}<P<A_{t+1}^{({\rm X})}$. In round $t+2$, X flips to C and Y keeps D. Therefore, we obtain $A_{t+3}^{({\rm X})}=hS+(1-h)A_{t+2}^{({\rm X})}$, $A_{t+2}^{({\rm X})}=hP+(1-h)A_{t+1}^{({\rm X})},$ $A_{t+1}^{({\rm X})}=hT+(1-h)A_t^{({\rm X})},$ and $A_t^{({\rm X})} \approx P$. Combining the four equations, we obtain $$A_{t+3}^{({\rm X})}=(T-P)h^3-2(T-P)h^2+(T+S-2P)h+P.$$ Using $T>P>S$ and $0<h \leq 1$, we obtain the condition for X to become Pavlov in round $t+3$ as $A_{t+3}^{({\rm X})} > P$, i.e., $$(1-h)^2 > \frac{P-S}{T-P}.
\label{eq:h iq1}$$ The condition for Y to become Pavlov in round $t+3$ is given by $A_{t+3}^{({\rm Y})} > P$, i.e., $$(1-h)^2 < \frac{T-P}{P-S}.
\label{eq:h iq2}$$ Equation implies Eq. because $(P-S)/(T-P)>0$ and $h>0$. Therefore, the two players become Pavlov in round $t+3$ if Eq. holds true.
We assume that Eq. is violated. If Eq. is also violated, we obtain $A_{t^\prime}^{({\rm X})},\ A_{t^\prime}^{({\rm Y})} \leq P\ (t^\prime \geq t+3)$ such that the two players mutually defect until the occurrence of another action misimplementation. If Eq. is satisfied, the two players mutually defect in round $t+3$. Because $r_{t+2}^{({\rm Y})}=r_{t}^{({\rm X})}=T,\ r_{t+3}^{({\rm Y})}=r_{t+1}^{({\rm X})}=P$, and $A_{t+2}^{({\rm Y})}<P$, we obtain $P<A_{t+4}^{({\rm Y})}<A_{t+2}^{({\rm X})}$. Because $r_{t+2}^{({\rm X})}=r_{t}^{({\rm Y})}=S,\ r_{t+3}^{({\rm X})}=r_{t+1}^{({\rm Y})}=P$, and $A_{t+2}^{({\rm X})}>P$, we obtain $A_{t+2}^{({\rm Y})}<A_{t+4}^{({\rm X})}<P$. These two inequalities indicate that X and Y behave as st3 and st2 in round $t+5$, respectively. By repeating the same procedure with X and Y swapped, we obtain $$A_{t+2}^{({\rm Y})}<A_{t+6}^{({\rm Y})}<P<A_{t+6}^{({\rm X})}<A_{t+2}^{({\rm X})}.$$ Therefore, we obtain $$A_{t+2}^{({\rm Y})}<A_{t+2+4i}^{({\rm Y})}<P<A_{t+2+4i}^{({\rm X})}<A_{t+2}^{({\rm X})}\ \ \ (i \geq 1)
\label{eq:At iq}$$ by induction. Equation implies that the two players do not realize mutual cooperation if Eq. is violated.
Therefore, an upper bound of $h$ for a pair of st2 players to turn into Pavlov is given by solving Eq. with equality.
Ancel, L. W. 1999[**]{}. A quantitative model of the [Simpson–Baldwin]{} effect. J. Theor. Biol. 196, 197–209.
Ancel, L. W. 2000[**]{}. Undermining the [Baldwin]{} expediting effect: does phenotypic plasticity accelerate evolution? Theor. Popul. Biol. 58, 307–319.
Axelrod, R. 1984[**]{}. . Basic Books, NY.
Badyaev, A. V. 2009[**]{}. Evolutionary significance of phenotypic accommodation in novel environments: an empirical test of the [Baldwin]{} effect. Philos. Trans. R. Soc. B 364, 1125–1141.
Borenstein, E., Meilijson, I. & Ruppin, E. 2006[**]{}. The effect of phenotypic plasticity on evolution in multipeaked fitness landscapes. J. Evol. Biol. 19, 1555–1570.
Camerer, C. F. 2003[**]{}. . Princeton University Press, NJ.
Crispo, E. 2007[**]{}. The [Baldwin]{} effect and genetic assimilation: revisiting two mechanisms of evolutionary change mediated by phenotypic plasticity. Evolution 61, 2469–2479.
Doebeli, M., Hauert, C. & Killingback, T. 2004[**]{}. The evolutionary origin of cooperators and defectors. Science 306, 859–862.
Dopazo, H., Gordon, M. B., Perazzo, R. & Risau-Gusman, S. 2001[**]{}. A model for the interaction of learning and evolution. Bull. Math. Biol. 63, 117–134.
Downes, S. M. 2003[**]{}. pp. 33–51. MIT Press, Cambridge, UK.
Fudenberg, D. & Levine, D. K. 1998[**]{}. . MIT Press, Cambridge, UK.
Godfrey-Smith, P. 2003[**]{}. pp. 53–67. MIT Press, Cambridge, UK.
Hinton, G. E. & Nowlan, S. J. 1987[**]{}. How learning can guide evolution. Complex Syst. 1, 495–502.
Hofbauer, J. & Sigmund, K. 1998[**]{}. . Cambridge University Press, Cambridge, UK.
Hofbauer, J. & Sigmund, K. 2003[**]{}. Evolutionary game dynamics. Bull. Am. Math. Soc. 40, 479–519.
Karandikar, R., Mookherjee, D., Ray, D. & Vega-Redondo, F. 1998[**]{}. Evolving aspirations and cooperation. J. Econ. Theory 80, 292–331.
Kraines, D. & Kraines, V. 1989[**]{}. Pavlov and the prisoner’s dilemma. Theory Decis. 26, 47–79.
Lindgren, K. 1991[**]{}. pp. 295–312.
Macy, M. 1996[**]{}. Natural selection and social learning in prisoner’s dilemma: co-adaptation with genetic algorithms and artificial neural networks. Sociol. Methods Res. 25, 103–137.
Macy, M. W. & Flache, A. 2002[**]{}. Learning dynamics in social dilemmas. Proc. Natl. Acad. Sci. USA 99, 7229–7236.
Masuda, N. & Nakamura, M. 2011[**]{}. Numerical analysis of a reinforcement learning model with the dynamic aspiration level in the iterated prisoner’s dilemma. J. Theor. Biol. 278, 55–62.
Masuda, N. & Ohtsuki, H. 2009[**]{}. A theoretical analysis of temporal difference learning in the iterated prisoner’s dilemma game. Bull. Math. Biol. 71, 1818–1850.
Maynard Smith, J. 1987[**]{}. Natural selection: when learning guides evolution. Nature 329, 761–762.
Metz, J. A. J., Geritz, S. A. H., Meszena, G., Jacobs, F. J. A. & van Heerwarden, J. S. 1996[**]{}. pp. 188–231. North Holland, Amsterdam, The Netherlands.
Nowak, M. & Sigmund, K. 1993[**]{}. A strategy of win-stay, lose-shift that outperforms tit-for-tat in the prisoner’s dilemma game. Nature 364, 56–58.
Nowak, M. A. & Sigmund, K. 1992[**]{}. Tit for tat in heterogeneous populations. Nature 355, 250–253.
Nowak, M. A., Sigmund, K. & El-Sedy, E. 1995[**]{}. Automata, repeated games and noise. J. Math. Biol. 33, 703–722.
Paenke, I., Kawecki, T. J. & Sendhoff, B. 2009[**]{}. The influence of learning on evolution: a mathematical framework. Artif. Life 15, 227–245.
Paenke, I., Sendhoff, B. & Kawecki, T. J. 2007[**]{}. Influence of plasticity and learning on evolution under directional selection. Am. Nat. 170, E47–E58.
Posch, M., Pichler, A. & Sigmund, K. 1999[**]{}. The efficiency of adapting aspiration levels. Proc. R. Soc. Lond. B 266, 1427–1435.
Sandholm, T. W. & Crites, R. H. 1996[**]{}. Multiagent reinforcement learning in the iterated prisoner’s dilemma. Biosystems 37, 147–166.
Simpson, G. G. 1953[**]{}. The [Baldwin]{} effect. Evolution 7, 110–117.
Suzuki, R. & Arita, T. 2004[**]{}. Interactions between learning and evolution: outstanding strategy generated by the [Baldwin]{} effect. Biosystems 77, 57–71.
Szabo, G. & Toke, C. 1998[**]{}. Evolutionary prisoner’s dilemma game on a square lattice. Phys. Rev. E 58, 69–73.
Taiji, M. & Ikegami, T. 1999[**]{}. Dynamics of internal models in game players. Physica D 134, 253–266.
Traulsen, A., Nowak, M. A. & Pacheco., J. M. 2006[**]{}. Stochastic dynamics of invasion and fixation. Phys. Rev. E 74, 011909.
Trivers, R. L. 1971[**]{}. The evolution of reciprocal altruism. Q. Rev. Biol. 46, 35–57.
Turney, P., Whitley, D. & Anderson, R. W. 1996[**]{}. Evolution, learning, and instinct: 100 years of the [Baldwin]{} effect. Evol. Comput. 4, 4–8.
Waddington, C. H. 1942[**]{}. Canalization of development and the inheritance of acquired characters. Nature 150, 563–565.
Wang, S., Szalay, M. S., Zhang, C. & Csermely, P. 2008[**]{}. Learning and innovative elements of strategy adoption rules expand cooperative network topologies. PLoS One, 3, e1997.
Weber, B. H. & Depew, D. J. eds. 2003[**]{}. . MIT Press, Cambridge, UK.
Yeh, P. J. & Price, T. D. 2004[**]{}. Adaptive phenotypic plasticity and the successful colonization of a novel environment. Am. Nat. 164, 531–542.
![ Average payoff to a nonlearning player (row player) playing against an opponent nonlearning player (column player). We set $\beta=3$, $p_1=0$, $\epsilon=0.02$, $h=0$, $t_{\max}=200$, $R=4,\ T=5,\ S=0,$ and $P=2$. []{data-label="fig:nl vs nl"}](nl_vs_nl.eps){width="\linewidth"}
![ Behavior of a pair of learning players. We set $\beta=3$, $p_1=0$, and $h=0.1$. (a) Example time courses of the aspiration level for a pair of players when $\epsilon=0$. The horizontal lines represent $A_t=P=2$ and $A_t=R=4$. We set $A_1$ for the two players to $-1$ and $-0.5$ (thick lines), $2.5$ and $3$ (dotted lines), and $5.5$ and $6$ (medium lines). (b) Example time courses of the aspiration level when $\epsilon=0.02$. We set $A_1$ as in (a). For each pair, one of the two players is assumed to misimplement the action to cooperate in round $30$ (thick lines), to defect in round $30$ (dotted lines), or to cooperate in round $7$ (medium lines). []{data-label="fig:l vs l"}](l_vs_l_a.eps){width="\linewidth"}
![ Behavior of a pair of learning players. We set $\beta=3$, $p_1=0$, and $h=0.1$. (a) Example time courses of the aspiration level for a pair of players when $\epsilon=0$. The horizontal lines represent $A_t=P=2$ and $A_t=R=4$. We set $A_1$ for the two players to $-1$ and $-0.5$ (thick lines), $2.5$ and $3$ (dotted lines), and $5.5$ and $6$ (medium lines). (b) Example time courses of the aspiration level when $\epsilon=0.02$. We set $A_1$ as in (a). For each pair, one of the two players is assumed to misimplement the action to cooperate in round $30$ (thick lines), to defect in round $30$ (dotted lines), or to cooperate in round $7$ (medium lines). []{data-label="fig:l vs l"}](l_vs_l_b.eps){width="\linewidth"}
![ Discretized adaptive dynamics when $\epsilon=0.02$. Plotted is $\pi[s^\prime,s]-\pi[s,s]$, where $s=(A_1,h)$ and (a) $s^\prime=(A_1+0.2,h)$, (b) $s^\prime=(A_1-0.2,h)$, (c) $s^\prime=(A_1,h+0.02)$, and (d) $s^\prime=(A_1,h-0.02)$. (e) Example time courses of a pair of players having $(A_1,h)=(1.7,0.1)$ and $(A_1,h)=(2.2,0.1)$ (solid lines), and $(A_1,h)=(1.7,0.1)$ and $(A_1,h)=(2.2,0)$ (dotted lines). The horizontal line indicates $A_t=P=2$. We set $\beta=3$, $p_1=0$, and $t_{\max}=200$. []{data-label="fig:adaptive dynamics"}](adaptive_dynamics_abcd.eps){width="\linewidth"}
![ Discretized adaptive dynamics when $\epsilon=0.02$. Plotted is $\pi[s^\prime,s]-\pi[s,s]$, where $s=(A_1,h)$ and (a) $s^\prime=(A_1+0.2,h)$, (b) $s^\prime=(A_1-0.2,h)$, (c) $s^\prime=(A_1,h+0.02)$, and (d) $s^\prime=(A_1,h-0.02)$. (e) Example time courses of a pair of players having $(A_1,h)=(1.7,0.1)$ and $(A_1,h)=(2.2,0.1)$ (solid lines), and $(A_1,h)=(1.7,0.1)$ and $(A_1,h)=(2.2,0)$ (dotted lines). The horizontal line indicates $A_t=P=2$. We set $\beta=3$, $p_1=0$, and $t_{\max}=200$. []{data-label="fig:adaptive dynamics"}](adaptive_dynamics_e.eps){width="\linewidth"}
![ Discretized adaptive dynamics when $\epsilon=0.1$. Plotted is $\pi[s^\prime,s]-\pi[s,s]$, where $s=(A_1,h)$ and (a) $s^\prime=(A_1+0.2,h)$, (b) $s^\prime=(A_1-0.2,h)$, (c) $s^\prime=(A_1,h+0.02)$, and (d) $s^\prime=(A_1,h-0.02)$. []{data-label="fig:eps=0.1"}](eps=0.1_abcd.eps){width="\linewidth"}
![ Effects of $t_{\max}$ and $\epsilon$ on the adaptive dynamics when $A_1 \approx P$ and $h=0$. We set $\beta=3$ and $p_1=0$. (a) $\pi[(2.1,0),(1.9,0)]-\pi[(1.9,0),(1.9,0)]$. (b) $\pi[(1.9,0),(2.1,0)]-\pi[(2.1,0),(2.1,0)]$. []{data-label="fig:tmax eps"}](tmax_eps_ab.eps){width="\linewidth"}
![ Evolutionary dynamics in a population composed of learning players. We set $\beta=3$, $p_1=0$, $\epsilon=0.02$, $t_{\max}=200$, $\Delta_{A_1}=0.05$, and $\Delta_h=0.01$. (a) Time course of ${\overline h},\ {\sum_{t=1}^{t_{\max}-1}\left|A_{t+1}-A_t \right|}/20,\ {\overline r}/R$, and the fraction of mutual cooperation. (b) Time course of the fraction of initially st1, st2, st3, st4, and st5 players in the run shown in (a). (c) Sample trajectory of the population averages $\overline {A_1}$ and $\overline h$ in the run shown in (a). []{data-label="fig:evolutionary dynamics"}](evolutionary_dynamics_a.eps){width="\linewidth"}
![ Evolutionary dynamics in a population composed of learning players. We set $\beta=3$, $p_1=0$, $\epsilon=0.02$, $t_{\max}=200$, $\Delta_{A_1}=0.05$, and $\Delta_h=0.01$. (a) Time course of ${\overline h},\ {\sum_{t=1}^{t_{\max}-1}\left|A_{t+1}-A_t \right|}/20,\ {\overline r}/R$, and the fraction of mutual cooperation. (b) Time course of the fraction of initially st1, st2, st3, st4, and st5 players in the run shown in (a). (c) Sample trajectory of the population averages $\overline {A_1}$ and $\overline h$ in the run shown in (a). []{data-label="fig:evolutionary dynamics"}](evolutionary_dynamics_b.eps){width="\linewidth"}
![ Evolutionary dynamics in a population composed of learning players. We set $\beta=3$, $p_1=0$, $\epsilon=0.02$, $t_{\max}=200$, $\Delta_{A_1}=0.05$, and $\Delta_h=0.01$. (a) Time course of ${\overline h},\ {\sum_{t=1}^{t_{\max}-1}\left|A_{t+1}-A_t \right|}/20,\ {\overline r}/R$, and the fraction of mutual cooperation. (b) Time course of the fraction of initially st1, st2, st3, st4, and st5 players in the run shown in (a). (c) Sample trajectory of the population averages $\overline {A_1}$ and $\overline h$ in the run shown in (a). []{data-label="fig:evolutionary dynamics"}](evolutionary_dynamics_c.eps){width="\linewidth"}
![Sample trajectory of $\overline {A_1}$ and $\overline h$ when $\beta=3$, $p_1=1$, $\epsilon=0.02$, $t_{\max}=200$, $\Delta_{A_1}=0.05$, and $\Delta_h=0.01$. []{data-label="fig:p_1=1"}](p1=1.eps){width="\linewidth"}
![ (a, b) Sample trajectory of $\overline {A_1}$ and $\overline h$ under a linear cost of learning. We set (a) $c=1$ and (b) $c=10$. (c) Largest $\overline h$ in each of the five runs of $10^6$ generations for various values of $c$. A cross corresponds to a single run. (d) Average of $\overline h$ over the last $10^4$ generations in each run. (e) Average of $\overline {A_1}$ over the last $10^4$ generations in each run. []{data-label="fig:cost"}](cost_a.eps){width="\linewidth"}
![ (a, b) Sample trajectory of $\overline {A_1}$ and $\overline h$ under a linear cost of learning. We set (a) $c=1$ and (b) $c=10$. (c) Largest $\overline h$ in each of the five runs of $10^6$ generations for various values of $c$. A cross corresponds to a single run. (d) Average of $\overline h$ over the last $10^4$ generations in each run. (e) Average of $\overline {A_1}$ over the last $10^4$ generations in each run. []{data-label="fig:cost"}](cost_b.eps){width="\linewidth"}
![ (a, b) Sample trajectory of $\overline {A_1}$ and $\overline h$ under a linear cost of learning. We set (a) $c=1$ and (b) $c=10$. (c) Largest $\overline h$ in each of the five runs of $10^6$ generations for various values of $c$. A cross corresponds to a single run. (d) Average of $\overline h$ over the last $10^4$ generations in each run. (e) Average of $\overline {A_1}$ over the last $10^4$ generations in each run. []{data-label="fig:cost"}](cost_c.eps){width="\linewidth"}
![ (a, b) Sample trajectory of $\overline {A_1}$ and $\overline h$ under a linear cost of learning. We set (a) $c=1$ and (b) $c=10$. (c) Largest $\overline h$ in each of the five runs of $10^6$ generations for various values of $c$. A cross corresponds to a single run. (d) Average of $\overline h$ over the last $10^4$ generations in each run. (e) Average of $\overline {A_1}$ over the last $10^4$ generations in each run. []{data-label="fig:cost"}](cost_d.eps){width="\linewidth"}
![ (a, b) Sample trajectory of $\overline {A_1}$ and $\overline h$ under a linear cost of learning. We set (a) $c=1$ and (b) $c=10$. (c) Largest $\overline h$ in each of the five runs of $10^6$ generations for various values of $c$. A cross corresponds to a single run. (d) Average of $\overline h$ over the last $10^4$ generations in each run. (e) Average of $\overline {A_1}$ over the last $10^4$ generations in each run. []{data-label="fig:cost"}](cost_e.eps){width="\linewidth"}
----- --------------------- --------------------- --------------------- -------------------- ---------------------
st1 $\frac{R+T+S+P}{4}$ $\frac{R+2S+3P}{6}$ $\frac{R+T+S+P}{4}$ $\frac{T+2S+P}{4}$ $\frac{R+T+S+P}{4}$
st2 $\frac{R+2T+3P}{6}$ $P$ $\frac{R+2T+2P}{5}$ $\frac{T+P}{2}$ $\frac{T+P}{2}$
st3 $\frac{R+T+S+P}{4}$ $\frac{R+2S+2P}{5}$ $R$ $\frac{R+S+P}{3}$ $\frac{R+T+S+P}{4}$
st4 $\frac{2T+S+P}{4}$ $\frac{S+P}{2}$ $\frac{R+T+P}{3}$ $\frac{R+P}{2}$ $\frac{R+P}{2}$
st5 $\frac{R+T+S+P}{4}$ $\frac{S+P}{2}$ $\frac{R+T+S+P}{4}$ $\frac{R+P}{2}$ $\frac{R+T+S+P}{4}$
----- --------------------- --------------------- --------------------- -------------------- ---------------------
: Average payoff to a nonlearning player (row player) playing against an opponent nonlearning player (column player). We set $\beta=\infty,\ 0<\epsilon\ll 1,$ and $t_{\max}=\infty$.[]{data-label="table:tab1"}
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
author:
- |
Mohsen Ghasempour$^\dagger$, Aamer Jaleel$^{\star}$, Jim Garside$^{\dagger}$ and Mikel Luj[á]{}n$^{\dagger}$\
School of Computer Science, University of Manchester$^{\dagger}$\
NVidia Research$^{\star}$
bibliography:
- 'references.bib'
title: 'HAPPY: Hybrid Address-based Page Policy in DRAMs'
---
Acknowledgements
================
The research leading to these results has received funding from the European Union’s Seventh Framework Programme (FP7/2007-2013) under grant agreement n$^{\circ}$ 318633; AXLE project http://axleproject.eu/. Mikel Luj[á]{}n is funded by a Royal Society University Research Fellowship and further supported by UK EPSRC grants DOME EP/J016330/1 and PAMELA EP/K008730/1.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'We show that there is a $\beta$-model of second-order arithmetic in which the choice scheme holds, but the dependent choice scheme fails for a $\Pi^1_2$-assertion, confirming a conjecture of Stephen Simpson. We obtain as a corollary that the Reflection Principle, stating that every formula reflects to a transitive set, can fail in models of ${{\rm ZFC}}^-$. This work is a rediscovery by the first two authors of a result obtained by the third author in [@kanovei:ACnotDC].'
address:
- 'Kurt Research Center, University of Vienna, Währinger Strasse 25 A-1090 Vienna, Austria'
- 'The City University of New York, CUNY Graduate Center, Mathematics Program, 365 Fifth Avenue, New York, NY 10016'
- 'Institute for the Information Transmission Problems (IITP), 19, build.1, Bolshoy Karetny per. Moscow 127051, Russia and Russian University of Transport (RUT-MIIT), Moscow, Russia'
author:
- 'Sy-David Friedman'
- Victoria Gitman
- Vladimir Kanovei
bibliography:
- 'ACnotDC.bib'
title: 'A model of second-order arithmetic satisfying AC but not DC'
---
Introduction
============
Models of arithmetic are two-sorted structures, having two types of objects, which we think of as numbers and sets of numbers. Their properties are formalized using a two-sorted logic with separate variables and quantifiers for numbers and sets. By convention, we will denote number variables by lower-case letters and sets variables by upper-case letters. The language of second-order arithmetic is the language of first-order arithmetic $\mathcal L_A=\{+,\cdot,<,0,1\}$ together with a membership relation $\in$ between numbers and sets. A multitude of second-order arithmetic theories, as well as the relationships between them, have been extensively studied (see [@simpson:second-orderArithmetic]).
An example of a weak second-order arithmetic theory is ${\rm ACA_0}$, whose axioms consist of the modified Peano axioms, where instead of the induction scheme we have the single second-order induction axiom $$\forall X [(0\in X\wedge \forall n(n\in X\rightarrow n+1\in X))\rightarrow \forall n (n\in X)],$$ and the *comprehension scheme* for first-order formulas. The latter is a scheme of assertions stating for every first-order formula, possibly with set parameters, that there is a set whose elements are exactly the numbers satisfying the formula. One of the strongest second-order arithmetic theories is ${{\rm Z}}_2$, often referred to as *full second-order arithmetic*, which strengthens comprehension for first-order formulas in ${\rm ACA}_0$ to full comprehension for all second-order assertions. This means that for a formula with any number of second-order quantifiers, there is a set whose elements are exactly the numbers satisfying the formula. For example, the reals of any model of ${{\rm ZF}}$ give a model of ${{\rm Z}}_2$. We can further strengthen the theory ${{\rm Z}}_2$ by adding choice principles for sets: the choice scheme and the dependent choice scheme.
The *choice scheme* is a scheme of assertions, which states for every second-order formula $\varphi(n,X,A)$ with a set parameter $A$ that if for every number $n$, there is a set $X$ witnessing $\varphi(n,X,A)$, then there is a single set $Y$ collecting witnesses for every $n$, in the sense that $\varphi(n,Y_n,A)$ holds, where $Y_n=\{m\mid {\langle}n,m{\rangle}\in Y\}$ and ${\langle}n,m{\rangle}$ is any standard coding of pairs. More precisely, an instance of the choice scheme for the formula $\varphi(n,X,A)$ is $$\forall n\exists X\,\varphi(n,X,A)\rightarrow \exists Y\forall n\,\varphi(n,Y_n,A).$$ We will denote by $\Sigma^1_n$-${{\rm AC}}$ the fragment of the choice scheme for $\Sigma^1_n$-assertions, making an analogous definition for $\Pi^1_n$, and we will denote the full choice scheme by $\Sigma^1_\infty$-${{\rm AC}}$. The reals of any model of ${{\rm ZF}}+{{\rm AC}}_\omega$ (countable choice) satisfy ${{\rm Z}}_2+\Sigma^1_\infty$-${{\rm AC}}$. It is a folklore result, going back possibly to Mostowski, that the theory ${{\rm Z}}_2+\Sigma^1_\infty$-${{\rm AC}}$ is bi-interpretable with the theory ${{\rm ZFC}}^-$ (${{\rm ZFC}}$ without the powerset axiom, with Collection instead of Replacement) together with the statement that every set is countable.
The *dependent choice scheme* is a scheme of assertions, which states for every second-order formula $\varphi(X,Y,A)$ with set parameter $A$ that if for every set $X$, there is a set $Y$ witnessing $\varphi(X,Y,A)$, then there is a single set $Z$ making infinitely many dependent choices according to $\varphi$. More precisely, an instance of the dependent choice scheme for the formula $\varphi(X,Y,A)$ is $$\forall X\exists Y\,\varphi(X,Y,A)\rightarrow \exists Z\forall n\,\varphi(Z_n,Z_{n+1},A).$$ We will denote by $\Sigma^1_n$-${{\rm DC}}$ the dependent choice scheme for $\Sigma^1_n$-assertions, with an analogous definition for $\Pi^1_n$, and we will denote the full dependent choice scheme by $\Sigma^1_\infty$-${{\rm DC}}$. The reals of a model of ${{\rm ZF}}+{{\rm DC}}$ (dependent choice) satisfy ${{\rm Z}}_2+\Sigma^1_\infty$-${{\rm DC}}$.
It is not difficult to see that the theory ${{\rm Z}}_2$ implies $\Sigma^1_2$-${{\rm AC}}$, the choice scheme for $\Sigma^1_2$-assertions. Models of ${{\rm Z}}_2$ can build their own version of [Gödel]{}’s constructible universe $L$. If a model of ${{\rm Z}}_2$ believes that a set $\Gamma$ is a well-order, then it has a set coding a set-theoretic structure constructed like $L$ along the well-order $\Gamma$. It turns out that models of ${{\rm Z}}_2$ satisfy a version of Shoenfield’s absoluteness with respect to their constructible universes. For every $\Sigma^1_2$-assertion $\varphi$, a model of ${{\rm Z}}_2$ satisfies $\varphi$ if and only its constructible universe satisfies $\varphi$ with set quantifiers naturally interpreted as ranging over the reals. All of the above generalizes to constructible universes $L[A]$ relativized to a set parameter $A$.[^1] Thus, given a $\Sigma^1_2$-assertion $\varphi(n,X,A)$ for which the model satisfies $\forall n\exists X\,\varphi(n,X,A)$, the model can go to its constructible universe $L[A]$ to pick the least witness $X$ for $\varphi(n,X,A)$ for every $n$, because $L[A]$ agrees when $\varphi$ is satisfied, and then put the witnesses together into a single set using comprehension. So long as the unique witnessing set can be obtained for each $n$, comprehension suffices to obtain a single set of witnesses. How much more of the choice scheme follows from ${{\rm Z}}_2$? The reals of the classical Feferman-[Lévy]{} model of ${{\rm ZF}}$ (see [@levy:choicescheme], Theorem 8), in which $\aleph_1$ is a countable union of countable sets, is a $\beta$-model of ${{\rm Z}}_2$ in which $\Pi^1_2$-${{\rm AC}}$ fails. This is a particulary strong failure of the choice scheme because, as we explain below, $\beta$-models are meant to strongly resemble the full standard model given by $P(\omega)$.
There are two ways in which a model of second-order arithmetic can resemble the full standard model given by $P(\omega)$. A model of second-order arithmetic is called an $\omega$-*model* if its first-order part is $\omega$, and it follows that its second-order part is some subset of $P(\omega)$. But even an $\omega$-model can poorly resemble $P(\omega)$ because it may be wrong about well-foundedness by missing $\omega$-sequences. An $\omega$-model of second-order arithmetic which is correct about well-foundedness is called a $\beta$-*model*. The reals of any transitive ${{\rm ZF}}$-model is a $\beta$-model of ${{\rm Z}}_2$. One advantage to having a $\beta$-model of ${{\rm Z}}_2$ is that the constructible universe it builds internally is isomorphic to an initial segment $L_\alpha$ of the actual constructible universe $L$.
The theory ${{\rm Z}}_2$ also implies $\Sigma^1_2$-${{\rm DC}}$ (see [@simpson:second-orderArithmetic], Theorem VII.9.2), the dependent choice scheme for $\Sigma^1_2$-assertions. In this article, we construct a symmetric submodel of a forcing extension of $L$ whose reals form a model of second-order arithmetic in which ${{\rm Z}}_2$ together with $\Sigma^1_\infty$-${{\rm AC}}$ holds, but $\Pi^1_2$-${{\rm DC}}$ fails. The forcing notion we use is a tree iteration of Jensen’s forcing for adding a unique generic real.
Jensen’s forcing, which we will call here ${\mathbb{P}}^J$, introduced by Jensen in [@jensen:real], is a subposet of Sacks forcing constructed in $L$ using the $\diamondsuit$ principle. The poset ${\mathbb{P}}^J$ has the ccc and adds a unique generic real over $L$. The collection of all $L$-generic reals for ${\mathbb{P}}^J$ in any model is $\Pi^1_2$-definable. Jensen used his forcing to show that it is consistent with ${{\rm ZFC}}$ that there is a $\Pi^1_2$-definable non-constructible real singleton [@jensen:real]. Recently Lyubetsky and the third author extended the “uniqueness of generic filters" property of Jensen’s forcing to finite-support products of ${\mathbb{P}}^J$ [@kanovei:productOfJensenReals]. They showed that in a forcing extension $L[G]$ by the $\omega$-length finite support-product of ${\mathbb{P}}^J$, the only $L$-generic reals for ${\mathbb{P}}^J$ are the slices of the generic filter $G$. The result easily extends to $\omega_1$-length finite support-products as well.
We in turn extend the “uniqueness of generic filters" property to tree iterations of Jensen’s forcing. We first define finite iterations ${\mathbb{P}}^J_n$ of Jensen’s forcing ${\mathbb{P}}^J$, and then define an iteration of ${\mathbb{P}}^J$ along a tree ${\mathcal T}$ to be a forcing whose conditions are functions from a finite subtree of ${\mathcal T}$ into ${\bigcup}_{n<\omega}{\mathbb{P}}_n^J$ such that nodes on level $n$ get mapped to elements of the $n$-length iteration ${\mathbb{P}}_n^J$ and conditions on higher nodes extend conditions on lower nodes. The functions are ordered by extension of domain and strengthening on each coordinate. We show that in a forcing extension $L[G]$ by the tree iteration of ${\mathbb{P}}^J$ along the tree ${{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1$ (or the tree ${{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega$) the only $L$-generic filters for ${\mathbb{P}}_n^J$ are the restrictions of $G$ to level $n$ nodes of tree. We proceed to construct a symmetric submodel of $L[G]$ which has the tree of ${\mathbb{P}}_n^J$-generic filters added by $G$ but no branch through it. The symmetric model we construct satisfies ${{\rm AC}}_\omega$ and the tree of ${\mathbb{P}}_n^J$-generic filters is $\Pi^1_2$-definable in it. The reals of this model thus provide the desired $\beta$-model of ${{\rm Z}}_2$ in which $\Sigma^1_\infty$-${{\rm AC}}$ holds, but $\Pi^1_2$-${{\rm DC}}$ fails.
\[th:main1\] There is a $\beta$-model of second-order arithmetic ${{\rm Z}}_2$ together with $\Sigma^1_\infty$-${{\rm AC}}$ in which $\Pi^1_2$-${{\rm DC}}$ fails.
It should be noted that in our model the instance of $\Pi^1_2$-${{\rm DC}}$ failure is parameter-free.
Theorem \[th:main1\] has a long and complicated history. In 1973, Simpson submitted an abstract claiming a proof of the result [@simpson:ACnotDC], but he didn’t follow it up with a publication because, as he kindly shared with the first author, he never worked out the details of the argument, which involved Jensen’s forcing, but did not have all the parts needed to solve the problem. In 1979, the third author published, in Russian, a technical report with the result [@kanovei:ACnotDC]. Finally, after the third author’s (joint with Lyubetsky) recent publications on uniqueness properties of Jensen’s forcing, the first and second author independently rediscovered the third author’s result, leading to this joint publication.
Our results also answer a long-standing open question of Zarach from [@Zarach1996:ReplacmentDoesNotImplyCollection] about whether the Reflection Principle holds in models of ${{\rm ZFC}}^-$. The *Reflection Principle* states that every formula can be reflected to a transitive set, and holds in ${{\rm ZFC}}$ by the [Lévy]{}-Montague reflection because every formula is reflected by some $V_\alpha$. In the absence of the von Neumann hierarchy, it is not clear how to realize reflection, and indeed we show that it fails in $H_{\omega_1}\models{{\rm ZFC}}^-$ of the symmetric model we construct.
\[th:main2\] The theory ${{\rm ZFC}}^-$ does not imply the Reflection Principle.
Jensen’s forcing {#sec:PerfectPosets}
================
Jensen’s poset ${\mathbb{P}}^J$, introduced in [@jensen:real], is a subposet of Sacks forcing in $L$ with the countable chain condition and the property that it adds a unique generic real. The poset ${\mathbb{P}}^J$ is constructed in $L$ as the union of a continuous chain ${\langle}{\mathbb{P}}_\alpha\mid\alpha<\omega_1{\rangle}$ of length $\omega_1$ of countable perfect posets. At successor stages $\alpha+1$, the principle $\diamondsuit$ is used to seal a certain countable collection of maximal antichains of ${\mathbb{P}}_\alpha$, so that by the end of the construction every maximal antichain of ${\mathbb{P}}^J$ is sealed and therefore countable.
Recall that a tree $T\subseteq {{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}2$ is *perfect* if every node in $T$ has a splitting node above it. Given two perfect trees $T$ and $S$, let’s define their *meet* $T\wedge S$ as follows. Let $U_0=T\cap S$, and, assuming $U_\alpha$ has been defined, let $U_{\alpha+1}$ be the set of all nodes in $U_\alpha$ which have a splitting node in $U_\alpha$ above them. At limit stages, take intersections. Since $U_0$ is countable, there must be a countable stage $\alpha$ in the construction for which $U_\alpha=U_{\alpha+1}$. We define $T\wedge S=U_\alpha$. It is not difficult to see that either $T\wedge S=\emptyset$ or $T\wedge S$ is a perfect tree. In the latter case, it is the maximal perfect tree contained in $T\cap S$, so that every perfect tree $U\subseteq T\cap S$ is contained in $T\wedge S$. Recall the standard terminology that if $T\subseteq {{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}2$ is a tree and $s$ is a node in $T$, then $T_s$ denotes the subtree of $T$ consisting of all nodes compatible with $s$. The following proposition is straightforward to check.
If $T,S$, and $R$ are perfect trees, then
1. $(T\wedge S)_s=T_s\wedge S_s$ for every node $s\in T\wedge S$,
2. $(T\wedge S)\wedge R=T\wedge (S\wedge R)$ ($\wedge$ is associative),
3. $(T{\cup}S)\wedge R=(T\wedge R){\cup}(S\wedge R)$ ($\wedge$ distributes over ${\cup}$).
Let’s recall now some facts about perfect posets, which are subposets of Sacks forcing closed under certain basic operations.
We say that a collection ${\mathbb{P}}$ of perfect trees ordered by inclusion is a *perfect poset*[^2] if
1. $({{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}2)_s\in{\mathbb{P}}$ for every $s\in {{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}2$, and\
for every $T,S\in {\mathbb{P}}$,
2. $T{\cup}S\in {\mathbb{P}}$ (closed under unions),
3. $T\wedge S\in {\mathbb{P}}$, if $T\wedge S\neq\emptyset$ (closed under meets).
The smallest perfect poset, which we will denote by ${\mathbb{P}}^{\text{min}}$, is the closure under finite unions of the collection $\{({{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}2)_s\mid s\in{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}2\}$. Note that two perfect trees $T$ and $S$ are compatible under the inclusion ordering precisely when $T\wedge S\neq\emptyset$. So if ${\mathbb{P}}\subseteq {\mathbb{Q}}$ are perfect posets, then $T,S\in{\mathbb{P}}$ are compatible in ${\mathbb{Q}}$ if and only if they are already compatible in ${\mathbb{P}}$. Standard arguments show that a generic filter $G$ for a perfect poset ${\mathbb{P}}$ is determined by the *generic real* $r$ that is a branch through $T\in{\mathbb{P}}$ if and only if $T\in G$.
Given a perfect poset ${\mathbb{P}}$, we will denote by ${\mathbb{P}}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$ the $\omega$-length finite-support product of ${\mathbb{P}}$. Conditions in ${\mathbb{P}}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$ are functions $p:\omega\to {\mathbb{P}}$ such that for all but finitely many $n$, $p(n)={{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}2$. We will sometimes abuse notation by writing $p={\langle}T_0,\ldots,T_{n-1}{\rangle}$ for conditions in ${\mathbb{P}}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$, meaning that all remaining coordinates are trivial.
Following [@abraham:jensenRealsIterations], we associate to a perfect poset ${\mathbb{P}}$, the poset ${\mathbb{Q}}({\mathbb{P}})$ whose conditions are pairs $(T,n)$ with $T\in{\mathbb{P}}$ and $n\in\omega$ ordered so that $(T_2,n_2)\leq (T_1,n_1)$ whenever $T_2\subseteq T_1$, $n_2\geq n_1$, and $T_1\cap {{}^{n_1}}2=T_2\cap {{}^{n_1}}2$. We will refer to ${\mathbb{Q}}({\mathbb{P}})$ as the *fusion poset* for ${\mathbb{P}}$ because fusion arguments involving ${\mathbb{P}}$ amount to producing a filter for ${\mathbb{Q}}({\mathbb{P}})$ meeting sufficiently many dense sets. If $G\subseteq {\mathbb{Q}}({\mathbb{P}})$ is $V$-generic, then the union of $T\cap {{}^{n}}2$ for $(T,n)\in G$ is the generic perfect tree ${\mathcal T}$ added by ${\mathbb{Q}}({\mathbb{P}})$. Note that ${\mathcal T}\leq T$ for every $T$ that appears in some condition in $G$.
Let ${\mathbb{Q}}({\mathbb{P}})^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$ denote the $\omega$-length finite-support product of the ${\mathbb{Q}}({\mathbb{P}})$. Conditions in ${\mathbb{Q}}({\mathbb{P}})^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$ are function $q:\omega\to {\mathbb{Q}}({\mathbb{P}})$ such that for all but finitely many $n$, $q(n)=({{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}2,0)$, but we will sometimes abuse notation by writing the conditions as finite tuples.
Suppose that $G\subseteq {\mathbb{Q}}({\mathbb{P}})^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$ is $V$-generic. Let ${\langle}{\mathcal T}_n\mid n<\omega{\rangle}\in V[G]$ be the $\omega$-length sequence of generic perfect trees derived from $G$ and let $\dot{{\mathcal T}}_n$ be the canonical ${\mathbb{Q}}({\mathbb{P}})^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$-names for the trees ${\mathcal T}_n$. In $V[G]$, let $${\mathbb{U}}=\{{\mathcal T}_n\wedge S\mid S\in{\mathbb{P}},\,n\in\omega,\,{\mathcal T}_n\wedge S\neq\emptyset\}.$$ We will denote by ${\mathbb{P}}^*$ the collection of perfect trees that is the closure under finite unions of ${\mathbb{P}}$ and ${\mathbb{U}}$ and argue that ${\mathbb{P}}^*$ is a perfect poset extending ${\mathbb{P}}$. This will be the case even when $G$ is only generic for a countable transitive model $M$ containing ${\mathbb{P}}$. At successor stages $\alpha+1$ in the construction of Jensen’s poset ${\mathbb{P}}^J$, if $\diamondsuit$ codes a certain countable transitive model $M$ with ${\mathbb{P}}_\alpha\in M$, we will extend ${\mathbb{P}}_\alpha$ to ${\mathbb{P}}_{\alpha+1}={\mathbb{P}}_\alpha^*$ constructed in a generic extension $M[G]$ by ${\mathbb{Q}}({\mathbb{P}}_\alpha)^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$.
Suppose that $M$ is a countable transitive model of ${{\rm ZFC}}^-+``\mathcal P(\omega)$ exists" and ${\mathbb{P}}\in M$ is a perfect poset. Clearly we have ${\mathbb{Q}}({\mathbb{P}})\in M$. We will argue that if $G\subseteq{\mathbb{Q}}({\mathbb{P}})^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$ is $M$-generic, then the poset ${\mathbb{P}}^*$ constructed in $M[G]$ is a perfect poset with the property that every maximal antichain of ${\mathbb{P}}$ that is an element of $M$ remains maximal in ${\mathbb{P}}^*$. First, we need the following easy proposition.
\[prop:subtreeContainedInS\] Suppose that ${\mathbb{P}}$ is a perfect poset and a condition $q\in{\mathbb{Q}}({\mathbb{P}})^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$ forces that $\dot{{\mathcal T}}_n\wedge \check S\neq\emptyset$ with $S\in {\mathbb{P}}$. Then there is a condition $\bar q\leq q$ with $\bar q(n)=(\bar T,m)$ and a node $s\in \bar T\cap {{}^{m}}2$ for which $(\bar T)_s\leq S$.
Let $q(n)=(T,m)$. Since $q{\Vdash}\dot{{\mathcal T}_n}\wedge \check S\neq\emptyset$, there must be some $s\in T\cap {{}^{m}}2$ such that $U=T_s\wedge S\neq\emptyset$. Let $\bar T$ be the perfect tree we get by replacing $T_s$ with $U$ in $T$ and let $\bar q\leq q$ be such that $\bar q(n)=(\bar T,m)$ and $\bar q(i)=q(i)$ for all $i\neq n$. Note that the closure under unions property of perfect posets is needed to conclude that $\bar T\in{\mathbb{P}}$.
\[prop:propertiesOfP\*\] Suppose $M$ is a countable transitive model of ${{\rm ZFC}}^-+``\mathcal P(\omega)$ exists" and ${\mathbb{P}}\in M$ is a perfect poset. If $G\subseteq {\mathbb{Q}}({\mathbb{P}})^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$ is $M$-generic and ${\mathbb{P}}^*$ is constructed in $M[G]$ as above, then:
1. ${\mathbb{P}}^*$ is a perfect poset.
2. ${\mathbb{U}}$ is dense in ${\mathbb{P}}^*$ and $\{{\mathcal T}_n\mid n<\omega\}$ is a maximal antichain of ${\mathbb{P}}^*$.
3. Every maximal antichain of ${\mathbb{P}}$ from $M$ remains maximal in ${\mathbb{P}}^*$.
First, let’s prove (2). We argue that every $T\in{\mathbb{P}}$ has some generic tree ${\mathcal T}_n$ below it. Fix $T\in{\mathbb{P}}$. Let $q\in{\mathbb{Q}}({\mathbb{P}})^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$. Since $q$ has finite support, we can choose some $n$ such that $q(n)=({{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}2,0)$ and strengthen $q$ to a condition $\bar q$ such that $q(m)=\bar q(m)$ for all $m\neq n$ and $\bar q(n)=(T,0)$. Clearly $\bar q{\Vdash}\dot{{\mathcal T}}_n\leq T$. So by density, there must be some such $\bar q\in G$.
Next, let’s show that for $i\neq j$, ${\mathcal T}_i\wedge {\mathcal T}_j=\emptyset$, and so the generic trees ${\mathcal T}_n$ form an antichain. Fix any condition $p\in{\mathbb{Q}}({\mathbb{P}})^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$ with $p(n)=(T_n,m_n)$. By strengthening further, we can assume that $m_i=m_j=m$. For each node $s$ that is on level $m$ of both $T_i$ and $T_j$, we can choose disjoint perfect trees $U^{s,i}\subseteq (T_i)_s$ and $U^{s,j}\subseteq (T_j)_s$. Let $S_i$ be the perfect tree which we get by replacing, for each $s$ on level $m$ of both $T_i$ and $T_j$, $(T_i)_s$ with $U^{s,i}$ in $T_i$, and let $S_j$ be obtained similarly. Now let $\bar p\leq p$ be the condition where on coordinate $i$, we put $S_i$ instead of $T_i$ and on coordinate $j$, we put $S_j$ instead of $T_j$. Clearly $\bar p$ forces that ${\mathcal T}_i\wedge{\mathcal T}_j=\emptyset$. It remains to show that every tree $T$ in ${\mathbb{P}}^*$ is compatible with some ${\mathcal T}_n$. If $T$ is a finite union of trees one of which is in $U$, then this is clear, and otherwise $T\in{\mathbb{P}}$, and we just argued that then some ${\mathcal T}_n\leq T$.
To prove (1), it suffices to show that ${\mathbb{P}}^*$ is closed under meets. So suppose that $R={\bigcup}_{i<n} R_i$ and $R'={\bigcup}_{j<m} R'_j$, where $R_i,R'_j\in{\mathbb{P}}{\cup}{\mathbb{U}}$, and consider $R\wedge R'={\bigcup}_{i<n,j<m}R_i\wedge R'_j$. We will argue that each $R_i\wedge R'_j$ is either empty or in ${\mathbb{P}}{\cup}{\mathbb{U}}$. If both $R_i$ and $R'_j$ are in ${\mathbb{P}}$, then the conclusion follows since ${\mathbb{P}}$ is a perfect poset. If one of the $R_i$ or $R'_j$ is ${\mathcal T}_k\wedge S$ with $S\in{\mathbb{P}}$ and the other is $S'\in{\mathbb{P}}$, then $R_i\wedge R'_j={\mathcal T}_k\wedge (S\wedge S')$ and the conclusion follows. Finally if $R_i={\mathcal T}_k\wedge S$ and $R'_j={\mathcal T}_{k'}\wedge S'$ with $S,S'\in{\mathbb{P}}$, then $R_i\wedge R'_j=({\mathcal T}_k\wedge {\mathcal T}_{k'})\wedge (S\wedge S')$. If $k=k'$, then the conclusion follows and if $k\neq k'$, we just argued above that ${\mathcal T}_k\wedge{\mathcal T}_{k'}=\emptyset$.
Let’s now prove (3). Fix a maximal antichain $\mathcal A\in M$ of ${\mathbb{P}}$. It suffices to show that every tree in ${\mathbb{U}}$ is compatible with some element of $\mathcal A$. So suppose that ${\mathcal T}_n\wedge S\in {\mathbb{U}}$ with $S\in{\mathbb{P}}$, and fix a condition $q\in G$ forcing that $\dot{{\mathcal T}}_n\wedge \check S\neq \emptyset$. We will argue that there is a a condition $q'\leq q$ with $q'(n)=(T',m)$ such that there is a node $s\in T'\cap {{}^{m}}2$ and $A\in\mathcal A$ for which $T'_s\leq S,A$. By Proposition \[prop:subtreeContainedInS\], there is $\bar q\leq q$ with $\bar q(n)=(\bar T,m)$ such that there is $s\in {{}^{m}}2$ with $\bar T_s\leq S$. Since $\mathcal A$ is maximal in ${\mathbb{P}}$, there is $A\in\mathcal A$ and $U\in{\mathbb{P}}$ such that $U\leq \bar T_s,A$. Let $T'$ be the tree we get by replacing $\bar T_s$ with $U$ in $\bar T$ and let $q'\leq q$ be the condition where $q'(k)=\bar q(k)$ for all $k\neq n$ and $q'(n)=(T',m)$. By density, there is some such condition $q'\in G$. But this means that $({\mathcal T}_n)_s\leq {\mathcal T}_n\wedge S$ and $({\mathcal T}_n)_s\leq A$, so that ${\mathcal T}_n\wedge S$ is compatible with $A\in\mathcal A$.
We let ${\mathbb{U}}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$ denote the collection of elements $p\in{\mathbb{P}}^{*{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$ such that for all $n$, $p(n)\in {\mathbb{U}}$ or $p(n)={{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}2$. Note that it follows from Proposition \[prop:propertiesOfP\*\] that ${\mathbb{U}}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$ is dense in ${\mathbb{P}}^{*{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$.
\[prop:propertiesOfProductP\*\] Suppose that $M$ is a countable transitive model of ${{\rm ZFC}}^-+``\mathcal P(\omega)$ exists" and ${\mathbb{P}}\in M$ is a perfect poset. If $G\subseteq {\mathbb{Q}}({\mathbb{P}})^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$ is $M$-generic and ${\mathbb{P}}^*$ is constructed in $M[G]$ as above, then every maximal antichain of ${\mathbb{P}}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$ from $M$ remains maximal in ${\mathbb{P}}^{*{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$.
Fix a maximal antichain $\mathcal A\in M$ of ${\mathbb{P}}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$. It suffices to show that every condition $p\in{\mathbb{U}}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$ is compatible with some element of $\mathcal A$. Let $$p={\langle}{\mathcal T}_{i_0}\wedge S_0,{\mathcal T}_{i_1}\wedge S_1,\ldots,{\mathcal T}_{i_m}\wedge S_m{\rangle}.$$ Note that the values $i_n$ do not have to be distinct and we need to carefully address this possibility. Fix a condition $q\in G$ forcing that $\dot {\mathcal T}_{i_n}\wedge \check S_n\neq \emptyset$ for every $n\leq m$. Repeatedly using the construction in the proof of Proposition \[prop:subtreeContainedInS\] and going to a large enough level, we can find a condition $\bar q\leq q$ with $\bar q(n)=(\bar T_n,k_n)$ such that for every $n\leq m$, there is a node $s_n$ on level $k_{i_n}$ of $\bar T_{i_n}$ such that $(\bar T_{i_n})_{s_n}\leq S_n$, and for $n\neq n'$, $s_n\neq s_{n'}$. Let $\bar p={\langle}(\bar T_{i_0})_{s_0}, (\bar T_{i_1})_{s_1},\ldots,(\bar T_{i_m})_{s_m}{\rangle}$. Since $\mathcal A$ is maximal in ${\mathbb{P}}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$, the condition $\bar p$ is compatible with some $a\in\mathcal A$, and so for every $n\leq m$, we can let $U_n=\bar p(n)\wedge a(n)\in {\mathbb{P}}$. Now, for every $i\in \{i_n\mid n\leq m\}$, let $R_i$ be the tree we get by replacing $(\bar T_i)_{s_n}$ with $U_n$ in $\bar T_i$ whenever $i_n=i$. Let $r$ be the condition such that $r(i)=(R_i,k_i)$ for $i\in \{i_n\mid n\leq m\}$ and $r(i)=\bar q(i)$ otherwise. By density, there is some such $r\in G$. But this means that $({\mathcal T}_{i_n})_{s_n}\leq S_n\wedge a(n)$, so that $p$ is compatible with $a$.
All our constructions will take place inside the constructible universe $L$. For reasons that will become obvious during the course of the constructions, we will restrict ourselves to countable models of ${{\rm ZFC}}^-+``\mathcal P(\omega)$ exists“ which happen to be initial segments $L_\alpha$ of $L$. So let’s call countable $L_\alpha$ satisfying ${{\rm ZFC}}^-+``\mathcal P(\omega)$ exists” *suitable* models. Relevant examples of suitable models for us will be transitive collapses of countable $M\prec L_{\omega_2}$.
We are now ready to review the construction of Jensen’s perfect poset ${\mathbb{P}}^J$, which will use the $\diamondsuit$-principle to anticipate and seal maximal antichains. So let’s start by fixing a canonically defined $\diamondsuit$-sequence ${\langle}S_\alpha\mid\alpha<\omega_1{\rangle}$. Note that if $M$ is a suitable model and $\delta=\omega_1^M$, then ${\langle}S_\alpha\mid \alpha<\delta{\rangle}$ is an element of $M$.
Jensen’s poset ${\mathbb{P}}^J$ will be the union of the following increasing sequence ${\langle}{\mathbb{P}}_\alpha\mid\alpha<\omega_1{\rangle}$ of perfect posets. Let ${\mathbb{P}}_0={\mathbb{P}}^{\text{min}}$. At limit stages, we will take unions. Suppose ${\mathbb{P}}_\alpha$ has been defined. We let ${\mathbb{P}}_{\alpha+1}={\mathbb{P}}_\alpha$, unless the following happens. Suppose $S_\alpha$ codes a well-founded and extensional binary relation $E\subseteq \alpha\times\alpha$ such that the collapse of $E$ is a suitable model $M_\alpha$ with ${\mathbb{P}}_\alpha\in M_\alpha$ and $\alpha=\omega_1^{M_\alpha}$. In this case, we take the $L$-least $M_\alpha$-generic filter $G\subseteq {\mathbb{Q}}({\mathbb{P}}_\alpha)^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$ and let ${\mathbb{P}}_{\alpha+1}={\mathbb{P}}^*_\alpha$ as constructed in $M_\alpha[G]$.
As we observed in Proposition \[prop:propertiesOfP\*\], ${\mathbb{P}}_{\alpha+1}$ is a perfect poset with the property that every maximal antichain of ${\mathbb{P}}_\alpha$ in $M_\alpha$ remains maximal in ${\mathbb{P}}_{\alpha+1}$. Also, by Proposition \[prop:propertiesOfProductP\*\], every maximal antichain of ${\mathbb{P}}_\alpha^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$ in $M_\alpha$ remains maximal in ${\mathbb{P}}_{\alpha+1}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$. Now let’s argue that every maximal antichain of ${\mathbb{P}}_\alpha^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$ in $M_\alpha$ remains maximal in the final poset ${\mathbb{P}}^{J{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$. It suffices to argue that the models $M_\alpha$ form an increasing sequence, which follows because if $\beta>\alpha$, then $\beta=\omega_1^{M_\beta}$, and therefore $M_\beta$ has $S_\alpha$ as an element and can collapse it to obtain $M_\alpha$. This shows that every maximal antichain of ${\mathbb{P}}_\alpha^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$ that is an element of $M_\alpha$ is sealed.
\[th:JensenForcingCCC\] The finite-support product ${\mathbb{P}}^{J{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$ of Jensen’s poset ${\mathbb{P}}^J$ has the ccc.
Fix a maximal antichain $\mathcal A$ of ${\mathbb{P}}^{J{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$. Choose some transitive $M\prec L_{\omega_2}$ of size $\omega_1$ with $\mathcal A\in M$. We can decompose $M$ as the union of a continuous elementary chain of countable substructures $$X_0\prec X_1\prec\cdots\prec X_\alpha\prec\cdots\prec M$$ with $\mathcal A\in X_0$. By properties of $\diamondsuit$, there is some $\alpha$ such that $\alpha=\omega_1\cap {X_\alpha}$, ${\mathbb{P}}_\alpha={\mathbb{P}}^J\cap X_\alpha$, and $S_\alpha$ codes $X_\alpha$. Let $M_\alpha$ be the transitive collapse of $X_\alpha$. Then ${\mathbb{P}}_\alpha$ is the image of ${\mathbb{P}}^J$ under the collapse and $\alpha$ is the image of $\omega_1$. Let $\bar {\mathcal A}=\mathcal A\cap X_\alpha$ be the image of $\mathcal A$ under the collapse. So at stage $\alpha$ in the construction of ${\mathbb{P}}^J$, we chose a forcing extension $M_\alpha[G]$ of $M_\alpha$ by ${\mathbb{Q}}({\mathbb{P}}_\alpha)^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$ and let ${\mathbb{P}}_{\alpha+1}={\mathbb{P}}^*_\alpha$ as constructed in $M_\alpha[G]$. Thus, by our observation above, $\bar {\mathcal A}$ remains maximal in ${\mathbb{P}}^{J{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$, and hence $\bar {\mathcal A}=\mathcal A$ is countable.
Finally, we would like to observe that every maximal antichain of ${\mathbb{P}}_{\alpha+1}$ from $M_\alpha[G]$, in particular the antichain ${\langle}{\mathcal T}_n\mid n<\omega{\rangle}$ of generic perfect trees, remains maximal in ${\mathbb{P}}^J$.
\[prop:MaxAntiMForcingExtension\] Suppose that $\mathcal A\in M_\alpha[G]$ is a maximal antichain of ${\mathbb{P}}_{\alpha+1}$ (or ${\mathbb{P}}_{\alpha+1}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}})$. Then $\mathcal A$ remains maximal in ${\mathbb{P}}^J$ (or ${\mathbb{P}}^{J{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}})$.
Fix $\beta>\alpha$. We already argued that $M_\alpha\in M_\beta$. Since $M_\beta$ sees that $M_\alpha$ is countable, it has some $M$-generic filter for ${\mathbb{Q}}({\mathbb{P}}_\alpha)^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$. But since $G_\alpha$ was chosen to be the $L$-least such filter, then by suitability, $M_\beta$ must contain $G_\alpha$. So $M_\alpha[G_\alpha]\subseteq M_\beta$.
The Kanovei-Lyubetsky Theorem {#sec:KanoveiLyubetsky}
=============================
We will reprove here the Kanovei-Lyubetsky theorem from [@kanovei:productOfJensenReals] showing that Jensen’s poset ${\mathbb{P}}^J$ from Section \[sec:PerfectPosets\], has the property that in a forcing extension $L[G]$ of $L$ by the finite support $\omega$-length product ${\mathbb{P}}^{J{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$, the only $L$-generic reals for ${\mathbb{P}}$ are the $\omega$-many slices of $G$. This is a generalization of Jensen’s uniqueness of generic filters property to products.
Suppose that ${\mathbb{P}}$ is a perfect poset and $H$ is a generic filter for the product ${\mathbb{P}}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$. We will call $x_n$ the real on the $n$-th coordinate of $H$ and let $\dot x_n$ be its canonical name.
For the next lemma, we suppose that ${\mathbb{P}}$ is a countable perfect poset that is an element of a suitable model $M$. We should think of ${\mathbb{P}}$ as one of the perfect posets ${\mathbb{P}}_\alpha$ arising at stage $\alpha$ in the construction of Jensen’s poset ${\mathbb{P}}^J$ and of $M$ as the model $M_\alpha$ from that stage.
\[th:denseToAvoidBranchesInProduct\] In $M$, suppose that $\dot r$ is a ${\mathbb{P}}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$-name for a real such that for all $n\in\omega$, ${\mathop{1\hskip-2.5pt {\rm l}}}{\Vdash}_{{\mathbb{P}}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}} \dot r\neq \dot x_n$. Then in a forcing extension $M[G]$ by ${\mathbb{Q}}({\mathbb{P}})^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$, for every generic perfect tree ${\mathcal T}_n$, conditions forcing that $\dot r\notin [{\mathcal T}_n]$ are dense in ${\mathbb{P}}^{*{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$.
Fix a condition $p\in {\mathbb{P}}^{*{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$ and $d\in\omega$. We need to find $p'\leq p$ such that $p'{\Vdash}\dot r\not\in [{\mathcal T}_d]$. Since ${\mathbb{U}}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$ is dense in ${\mathbb{P}}^{*{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$, we can assume without loss that $p\in {\mathbb{U}}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$. So let $$p={\langle}{\mathcal T}_{i_0}\wedge S_0,{\mathcal T}_{i_1}\wedge S_1,\ldots,{\mathcal T}_{i_m}\wedge S_m{\rangle}.$$ By strengthening if necessary, we can assume that for some $n\leq m$, $i_n=d$. As we argued in the proof of Proposition \[prop:propertiesOfProductP\*\], there is a condition $q\in G$ with $q(n)=(T_n,k_n)$ such that for every $n\leq m$, there is a node $s_n$ on level $k_{i_n}$ of $T_{i_n}$ such that $(T_{i_n})_{s_n}\leq S_n$ and for $n\neq n'$, $s_n\neq s_{n'}$.
Now we are going to construct a condition $a_q={\langle}W_0,\ldots,W_{l'},\ldots,W_l{\rangle}\in{\mathbb{P}}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$ associated to $q$ and satisfying the following properties.
1. For every $n\leq m\leq l'$, $W_n\leq (T_{i_n})_{s_n}$.
2. For every node $s$ on level $k_d$ of $T_d$, there is $i\leq l'$ such that $W_i\leq (T_d)_s$.
3. For every $n\leq l'$, $a_q{\Vdash}_{{\mathbb{P}}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\dot r\notin[W_n]$.
Let $a={\langle}(T_{i_0})_{s_0},(T_{i_1})_{s_1},\ldots,(T_{i_m})_{s_m}, (T_d)_{r_0},\ldots,(T_d)_{r_k}{\rangle}$, where $r_0,\ldots,r_k$ are nodes on level $k_d$ of $T_d$ excluding those amongst the $s_n$. Since ${\mathop{1\hskip-2.5pt {\rm l}}}_{{\mathbb{P}}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}{\Vdash}\dot r\neq \dot x_n$ for all $n\in\omega$, we can strengthen $a$ to a condition $a_q={\langle}W_0,W_1,\ldots,W_{m+k+2},\ldots,W_l{\rangle}$ forcing that $\dot r\notin [W_n]$ for all $n\leq m+k+2$.
Using $a_q$, we construct the following condition $\bar q\leq q$. For $i\in\{i_n\mid n\leq m\}$, let $R_i$ be the tree we get by replacing $(T_i)_{s_n}$ with $W_n$ in $T_i$ whenever $i_n=i$ and if $i=d$, then we also replace all $T_{r_n}$ with the appropriate $W_j$ as well. Let $\bar q(i)=(R_i, k_i)$ for $i\in \{i_n\mid n\leq m\}$ and $\bar q(i)=q(i)$ otherwise. By density, some such condition $\bar q$, constructed from $a_q$, is in $G$. It follows that each $({\mathcal T}_{i_n})_{s_n}\leq W_n\wedge S_n$ and ${\mathcal T}_d$ has a level $k_d$ such that for every node $s$ on level $k_d$, ${\mathcal T}_d\leq W_j$ for some $j$. The first part gives us that $p={\langle}{\mathcal T}_{i_0}\wedge S_0,{\mathcal T}_{i_1}\wedge S_1,\ldots,{\mathcal T}_{i_m}\wedge S_m{\rangle}$ is compatible with $a_q$. Finally, we would like to argue that $a_q{\Vdash}_{{\mathbb{P}}^{*{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}} \dot r\notin[{\mathcal T}_d]$.
Let $H^*$ be any $V$-generic filter for ${\mathbb{P}}^{*{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$ containing $a_q$ ($V=M[G]$). Since every maximal antichain of ${\mathbb{P}}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$ in $M$ remains maximal in ${\mathbb{P}}^{*{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$, it follows that $H^*$ restricts to an $M$-generic filter $H$ for ${\mathbb{P}}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$ and $a_q\in H$. Thus in $M[H]$, $\dot r_H$ is not a branch through any $W_n$. But this is absolute, and so $\dot r_H$ is not a branch through ${\mathcal T}_d$ in $V[H^*]$.
\[th:uniquenessOfGenericsProduct\] Suppose $H\subseteq {\mathbb{P}}^{J{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$ is $L$-generic. If $r\in L[H]$ is $L$-generic for ${\mathbb{P}}^J$, then $r=x_n$ for some $n<\omega$.
Let’s suppose that $r\in L[H]$ is a real which is not one of the $x_n$. Let $\dot r$ be a nice ${\mathbb{P}}^{J{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$-name for $r$ such that for all $n\in\omega$, ${\mathop{1\hskip-2.5pt {\rm l}}}_{{\mathbb{P}}^{j{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}{\Vdash}\dot r\neq \dot x_n$.
Choose some transitive model $M\prec L_{\omega_2}$ of size $\omega_1$ with $\dot r\in N$. We can decompose $M$ as the union of a continuous elementary chain of countable substructures $$X_0\prec X_1\prec\cdots\prec X_\alpha\prec\cdots\prec M$$ with $\dot r\in X_0$. By properties of $\diamondsuit$, there is some $\alpha$ such that $\alpha=\omega_1\cap X_\alpha$, ${\mathbb{P}}_\alpha={\mathbb{P}}^J\cap X_\alpha$, and $S_\alpha$ codes $X_\alpha$. Let $M_\alpha$ be the collapse of $X_\alpha$. Then ${\mathbb{P}}_\alpha$ is the image of ${\mathbb{P}}^J$ under the collapse and $\alpha$ is the collapse of $\omega_1$. Clearly $\dot r$ is fixed by the collapse because it is a nice name and all antichains of ${\mathbb{P}}^{J{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$ are countable (by Theorem \[th:JensenForcingCCC\]). So at stage $\alpha$ in the construction of ${\mathbb{P}}^J$, we chose a forcing extension $M_\alpha[G]$ by ${\mathbb{Q}}({\mathbb{P}}_\alpha)^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$ and let ${\mathbb{P}}_{\alpha+1}={\mathbb{P}}_\alpha^*$ as constructed in $M_\alpha[G]$. By elementarity, $M_\alpha$ satisfies that ${\mathop{1\hskip-2.5pt {\rm l}}}_{{\mathbb{P}}_\alpha^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}{\Vdash}\dot r\neq \dot x_n$ for all $n\in\omega$. Thus, by Theorem \[th:denseToAvoidBranchesInProduct\], for every $n<\omega$, ${\mathbb{P}}_{\alpha+1}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$ has a maximal antichain $\mathcal A_n\in M_\alpha[G]$ consisting of conditions $q$ such that $q{\Vdash}_{{\mathbb{P}}_{\alpha+1}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\dot r\notin [{\mathcal T}_n]$. It follows, using Proposition \[prop:MaxAntiMForcingExtension\], that all the antichains $\mathcal A_n$ remain maximal in ${\mathbb{P}}^{J{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$. Also, by Proposition \[prop:MaxAntiMForcingExtension\], ${\langle}{\mathcal T}_n\mid n<\omega{\rangle}$ remains maximal in ${\mathbb{P}}^J$.
So let’s argue that if $q\in \mathcal A_n$, then $q{\Vdash}_{{\mathbb{P}}^{J{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\dot r\notin [{\mathcal T}_n]$. Let $\bar H^*\subseteq {\mathbb{P}}^{J{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$ be an $L$-generic filter containing $q$ and let $\bar H$ be the restriction of $\bar H^*$ to ${\mathbb{P}}_{\alpha+1}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$. Since $q\in \bar H$, it follows that $\dot r_{\bar H}\notin [{\mathcal T}_n]$ holds in $M_\alpha[G][\bar H]$, but this statement is absolute and so also holds in $L[\bar H^*]$. Since $H$ must meet every $\mathcal A_n$, it holds in $L[H]$ that $\dot r_H=r$ is not a branch through any ${\mathcal T}_n$. So $r$ cannot be $L$-generic for ${\mathbb{P}}^J$.
Finite iterations of perfect posets
===================================
An iteration of perfect posets is an iteration of forcing notions in which every initial segment forces that the next poset is perfect. Here we will only be dealing with finite iterations of perfect posets, so we are not concerned with issues of support.
A *finite iteration of perfect posets* is a finite iteration $${\mathbb{P}}_n={\mathbb{Q}}_0*\dot {\mathbb{Q}}_1*\cdots*\dot {\mathbb{Q}}_{n-1}$$ such that ${\mathbb{Q}}_0$ is a perfect poset and for $1\leq i<n$,
${\mathop{1\hskip-2.5pt {\rm l}}}_{{\mathbb{P}}_i}{\Vdash}``\dot {\mathbb{Q}}_i$ is a perfect poset".
Suppose $G\subseteq {\mathbb{P}}_n$ is $V$-generic. For $1\leq i<n$, let $G_i$ be the restriction of $G$ to ${\mathbb{P}}_i$. Let $G(0)=G_1$ and for $1\leq i<n$, let $G(i)=\{p(i)_{G_i}\mid p\in G\}$. Let $r_i$ be the unique real determined by $G(i)$. It is not difficult to see that the sequence of reals $\vec r={\langle}r_0,\ldots, r_{n-1}{\rangle}$ determines $G$. Elements of $G_1$ are trees with $r_0$ as a branch, and inductively, elements of $G_{i+1}$ are conditions $p\in {\mathbb{P}}_i$ such that $p{\upharpoonright}i\in G_i$ and $r_i\in p(i)_{G_i}$.
The analogue of the fusion poset ${\mathbb{Q}}({\mathbb{P}})$ for a finite iteration ${\mathbb{P}}_n$ of perfect posets is the fusion poset ${\mathbb{Q}}({\mathbb{P}}_n)$ whose conditions are pairs $(p,F)$ with $p\in{\mathbb{P}}_n$ and $F:n\to\omega$, ordered so that $(p_2,F_2)\leq (p_1,F_1)$ whenever $p_2\leq p_1$ and for every $i<n$, we have $F_2(i)\geq F_1(i)$ and $$p_2{\upharpoonright}i{\Vdash}p_1(i)\cap {{}^{F_1(i)}}2=p_2(i)\cap {{}^{F_1(i)}}2.$$
Fusion arguments with names for perfect trees require that we have some information about a fixed level $n$ of the tree. We will now argue that there are densely many conditions in ${\mathbb{Q}}({\mathbb{P}}_n)$ where this is the case.
Suppose $p\in {\mathbb{P}}_n$ and $\sigma:n\to {{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}2$. Following [@abraham:jensenRealsIterations], let’s define, by induction on $n$, what it means for $\sigma$ to *lie* on $p$.[^3] For $n=1$, we shall say that $\sigma$ *lies* on $p$ whenever $\sigma(0)\in p(0)$. If $\sigma$ lies on $p$, we shall denote by $p\mid\sigma$ the condition $p(0)_{\sigma(0)}$. Note that $p\mid\sigma\leq p$. So suppose that we have defined when $\sigma$ lies on $p$ for $p\in {\mathbb{P}}_n$, and for $\sigma$ which lies on $p$, we have defined $p\mid\sigma$ so that $p\mid\sigma\leq p$. Let $p\in {\mathbb{P}}_{n+1}$. We define that $\sigma$ lies on $p$ if $\sigma{\upharpoonright}n$ lies on $p{\upharpoonright}n$ and $(p{\upharpoonright}n)\mid (\sigma{\upharpoonright}n){\Vdash}\sigma(n)\in p(n)$. If $\sigma$ lies $p$, we shall denote by $p\mid\sigma$ the condition $\bar p$ such that $\bar p{\upharpoonright}n=(p{\upharpoonright}n)\mid (\sigma{\upharpoonright}n)$ and $\bar p(n)=\dot T$, where $\dot T$ is a ${\mathbb{P}}_n$-name that is interpreted as $p(n)_{\sigma(n)}$ by any ${\mathbb{P}}_n$-generic filter containing $(p{\upharpoonright}n)\mid (\sigma{\upharpoonright}n)$ and as $p(n)$ otherwise. Clearly this gives that $p\mid\sigma\leq p$.
Let $F:n\to\omega$ and $\sigma:n\to{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}2$. We shall say that $\sigma$ *lies on levels* $F$ if $\sigma(i)\in {{}^{F(i)}}2$ for all $i<n$.[^4] We shall say that a pair $(p,F)$ with $p\in{\mathbb{P}}_n$ is *determined* if for every $\sigma$ lying on levels $F$, either $\sigma$ lies on $p$ or there is some $i<n$ such that $\sigma{\upharpoonright}i$ lies on $p{\upharpoonright}i$ and $(p{\upharpoonright}i)\mid (\sigma{\upharpoonright}i){\Vdash}\sigma(i)\notin p(i)$. If $\sigma$ lies on levels $F$ and lies on $p$, we shall say that $\sigma$ *lies* on $(p,F)$.
Observe that whenever a pair $(p,F)$ is determined, for every $\sigma$ which lies on $p$, $(p\mid\sigma){\upharpoonright}i$ decides the $F(i)$-th level of $p(i)$, and indeed $(p\mid\sigma){\upharpoonright}i$ forces that $p(i)$ is the union of the $p\mid\sigma'(i)$ for $\sigma'{\upharpoonright}i=\sigma{\upharpoonright}i$ that lie on $(p,F)$.
Suppose ${\mathbb{P}}_n$ is a finite iteration of perfect posets.
1. Determined conditions $(p,F)$ are dense in ${\mathbb{Q}}({\mathbb{P}}_n)$.
2. Given a determined condition $(p,F)$, the finite set $$\{p\mid \sigma\mid \sigma\text{ lies on }(p,F)\}$$ is a maximal antichain below $p$ in ${\mathbb{P}}_n$.
\[prop:InsideDenseOpenSet\] Suppose ${\mathbb{P}}_n$ is a finite iteration of perfect posets and $D\subseteq {\mathbb{P}}_n$ is open dense. For any determined $(p,F)\in{\mathbb{Q}}({\mathbb{P}}_n)$, there is a determined $(q,F)\leq (p,F)$ satisfying $$q\mid\sigma\in D\text{ whenever }\sigma\text{ lies on }(q,F).$$
For details of proofs, see [@abraham:jensenRealsIterations].
Since we will be working mainly with determined conditions, we will now introduce a kind of normal form for them.
\[def:sigmaAssigment\] Suppose ${\mathbb{P}}_n$ is a finite iteration of perfect posets and $\sigma:n\to {{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}2$. Let’s call a condition $p\in {\mathbb{P}}_n$ a $\sigma$-*condition* if
1. $p(0)\leq ({{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}2)_{\sigma(0)}$,
2. for all $1\leq i<n$, $p{\upharpoonright}i{\Vdash}p(i)\leq ({{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}2)_{\sigma(i)}$.
Suppose that $X_F$ is a collection of $\sigma:n\to{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}2$ lying on levels $F:n\to\omega$. An $X_F$-*assignment* is a function $\varphi:X_F\to {\mathbb{P}}_n$ such that each $\varphi(\sigma)$ is a $\sigma$-condition and $\varphi(\sigma){\upharpoonright}i=\varphi(\sigma'){\upharpoonright}i$ whenever $\sigma{\upharpoonright}i=\sigma'{\upharpoonright}i$.
To motivate these definitions, consider a determined condition $(p,F)\in {\mathbb{Q}}({\mathbb{P}}_n)$ and let $X_F^p$ be the set of all $\sigma$ that lie on it. In this case, the map $\varphi_p$ defined by $\varphi_p(\sigma)=p\mid\sigma$ for all $\sigma\in X_F^p$ is clearly an $X_F^p$-assignment. Thus, a determined condition $(p,F)$ gives us a natural $X_F$-assignment, and now conversely we would like to argue that any $X_F$-assignment has a naturally associated determined condition $(q,F)$.
Suppose $X_F$ and $\varphi$ are as in Definition \[def:sigmaAssigment\]. Observe that given any $\sigma,\sigma'\in X_F$, either $\sigma(0)=\sigma'(0)$ and so $\varphi(\sigma)(0)=\varphi(\sigma')(0)$, or $\sigma(0)=s\neq t=\sigma'(0)$ are two nodes on level $F(0)$ such that $\varphi(\sigma)(0)\leq ({{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}2)_s$ and $\varphi(\sigma')(0)\leq ({{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}2)_t$. More generally, if $\sigma\neq\sigma'$, then there is some least $i$ such that $\sigma{\upharpoonright}i=\sigma'{\upharpoonright}i$ and there are nodes $\sigma(i)=s\neq t=\sigma'(i)$ on level $F(i)$ such that $\varphi(\sigma){\upharpoonright}i{\Vdash}\varphi(\sigma)(i)\leq ({{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}2)_s$ and $\varphi(\sigma'){\upharpoonright}i{\Vdash}\varphi(\sigma')(i)\leq({{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}2)_t$. It follows, in particular, that for any $i<n$, the conditions $\varphi(\sigma){\upharpoonright}i$ for $\sigma\in X_F$ form an antichain.
\[prop:sigmaConditions\] Suppose $X_F$ is a collection of $\sigma:n\to {{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}2$ on level some $F:n\to\omega$ and $\varphi$ is an $X_F$-assignment. Then there is a determined condition $(q,F)\in {\mathbb{Q}}({\mathbb{P}}_n)$ such that
1. $\sigma$ lies on $(q,F)$ if and only if $\sigma\in X_F$,
2. for every $\sigma\in X_F$, $$q\mid\sigma(0)=\varphi(\sigma)(0),$$ and for all $i<n$, $(q\mid\sigma){\upharpoonright}i$ and $\varphi(\sigma){\upharpoonright}i$ extend each other and force $$q\mid\sigma(i)=\varphi(\sigma)(i).$$
Let $q(0)$ be the union of the $\varphi(\sigma)(0)$ for $\sigma\in X_F$, which is in the perfect poset ${\mathbb{Q}}_0$. Let $q(1)$ be the (canonical) ${\mathbb{P}}_1$-name for the tree which is the union of the collection of trees given by the interpretation of the name $$\{( \varphi(\sigma)(1),\varphi(\sigma)(0))\mid \sigma\in X_F\}.$$ Since each $\varphi(\sigma)(0){\Vdash}\varphi(\sigma)(1)\in\dot{\mathbb{Q}}_1$, and $\dot {\mathbb{Q}}_1$ is forced to be a perfect poset, it follows that $q(0){\Vdash}q(1)\in \dot{\mathbb{Q}}_1$. Let’s see what $q(1)_G$ looks like in a forcing extension $V[G]$ by ${\mathbb{P}}_1$. If $\varphi(\sigma)(0)$ and $\varphi(\sigma')(0)$ are in $G$, then $\sigma(0)=\sigma'(0)=s$ for some $s$, and so $\varphi(\sigma)(0)=\varphi(\sigma')(0)=p$ for some $p$. Thus, the interpretation $q(1)_G$ is the tree which is the union of the $\varphi(\sigma)(1)_G$ for $\sigma(0)=s$. Similarly, let $q(i)$ be the ${\mathbb{P}}_i$-name for the tree which is the union of the collection of trees given by the interpretation of the name $\{(\varphi(\sigma)(i), \varphi(\sigma){\upharpoonright}i)\mid \sigma\in X_F\}$. Again, we have $q{\upharpoonright}i{\Vdash}q(i)\in\dot{\mathbb{Q}}_i$. Let’s see now what $q(i)_G$ looks like in a forcing extension $V[G]$ by ${\mathbb{P}}_i$. Since $\varphi(\sigma){\upharpoonright}i$ for $\sigma\in X_F$ form an antichain, if $\varphi(\sigma){\upharpoonright}i$ and $\varphi(\sigma'){\upharpoonright}i$ are both in $G$, then $\sigma'{\upharpoonright}i=\sigma{\upharpoonright}i=\tau$ for some $\tau$. So the interpretation $q(i)_G$ is the union of the $\varphi(\sigma)(i)_G$ for $\sigma{\upharpoonright}i=\tau$.
First, we argue that every $\sigma\in X_F$ lies on $(q,F)$ and simultaneously show (2). So fix some $\sigma\in X_F$. By construction $\sigma(0)\in q(0)$ and $q(0)_{\sigma(0)}=\varphi(\sigma)(0)$. So assume inductively that for some $i<n$,
1. $\sigma{\upharpoonright}i$ lies on $q{\upharpoonright}i$,
2. $(q{\upharpoonright}i)\mid(\sigma{\upharpoonright}i)\leq\varphi(\sigma){\upharpoonright}i$,
3. $\varphi(\sigma){\upharpoonright}i\leq (q{\upharpoonright}i)\mid(\sigma{\upharpoonright}i)$.
Suppose that $G\subseteq {\mathbb{P}}_i$ is a $V$-generic filter containing $(q{\upharpoonright}i)\mid(\sigma{\upharpoonright}i)$. Then also, by assumption, $\varphi(\sigma){\upharpoonright}i\in G$. By definition of $q$, we have $$(q(i)_G)_{\sigma(i)}=\varphi(\sigma)(i)_G.$$ So $\sigma{\upharpoonright}i+1$ lies on $q{\upharpoonright}i+1$ and we have $$(q{\upharpoonright}i+1)\mid(\sigma{\upharpoonright}i+1)\leq \varphi(\sigma){\upharpoonright}i+1\text{ and }\varphi(\sigma){\upharpoonright}i+1\leq (q{\upharpoonright}i+1)\mid(\sigma{\upharpoonright}i+1).$$
Now suppose that $\tau:n\to{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}2$ lies on $(q,F)$. By definition of $q$, it follows that $\tau(0)=\sigma(0)$ for some $\sigma\in X_F$. So suppose inductively that for some $i<n$, there is $\sigma\in X_F$ such that $\tau{\upharpoonright}i=\sigma{\upharpoonright}i$. Since $\tau$ lies on $q$, it follows that $(q{\upharpoonright}i)\mid (\sigma{\upharpoonright}i){\Vdash}\tau(i)\in q(i)$. But then there is some $\sigma'$ such that $\sigma{\upharpoonright}i=\sigma'{\upharpoonright}i$ and $\tau(i)=\sigma'(i)$. So in the last step, we will obtain $\sigma\in X_F$ such that $\sigma=\tau$.
We shall call the condition $q$ constructed in Proposition \[prop:sigmaConditions\] the *amalgamation* of $\varphi$.
Suppose $(p,F)$ is determined and $q$ is the amalgamation of the $X_F^p$-assignment $\varphi_p$. Then $q\leq p$ and $p\leq q$.
Clearly $p(0)=q(0)$. So let’s suppose inductively that for some $i<n$, we have $q{\upharpoonright}i\leq p{\upharpoonright}i$ and $p{\upharpoonright}i\leq q{\upharpoonright}i$. We will argue that both $p{\upharpoonright}i$ and $q{\upharpoonright}i$ force $p(i)=q(i)$. So suppose $G\subseteq {\mathbb{P}}_i$ is a $V$-generic filter containing $p{\upharpoonright}i$, and hence also $q{\upharpoonright}i$. Since the conditions $(p\mid\sigma){\upharpoonright}i$ for $\sigma\in X_F^p$ form a maximal antichain below $p{\upharpoonright}i$, it follows that one such $(p\mid\sigma){\upharpoonright}i$ is in $G$. Thus, $p(i)_G$ is the union of the $(p\mid\sigma'(i))_G$ for $\sigma'\in X_F^p$ with $\sigma'{\upharpoonright}i=\sigma{\upharpoonright}i$. Similarly, there is some $\tau\in X_F^p$ such that $(q\mid\tau){\upharpoonright}i\in G$, so we must have $\tau{\upharpoonright}i=\sigma{\upharpoonright}i$. Also, $q(i)_G$ is the union of the $(q\mid\sigma'(i))_G$ for $\sigma'\in X_F^p$ with $\sigma'{\upharpoonright}i=\sigma{\upharpoonright}i$. Now, by Proposition \[prop:sigmaConditions\], $(q\mid\sigma){\upharpoonright}i{\Vdash}q\mid\sigma'(i)=p\mid\sigma'(i)$ for all $\sigma'\in X_F^p$ with $\sigma'{\upharpoonright}i=\sigma{\upharpoonright}i$. So $p(i)_G=q(i)_G$.
Let us say that a map $\tau:n\to{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}2$ *extends* another map $\sigma:n\to{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}2$ if $\tau(i)$ extends $\sigma(i)$ for every $i<n$. Now we would like to argue that given a determined condition $(q,F)$ and strengthenings $q_\sigma\leq q\mid\sigma$ for every $\sigma$ lying on $(q,F)$, we can obtain a determined condition $(\bar q,\bar F)\leq (q,F)$ such that for every $\sigma$ lying on $(q,F)$, there is some $\tau$ lying on $(\bar q,\bar F)$ extending $\sigma$ with $\bar q\mid \tau\leq q_\sigma$. Indeed, we will get the following stronger statement.
\[prop:strengthenDependentAssigment\] Suppose $(q,F)\in {\mathbb{Q}}({\mathbb{P}}_n)$ is determined and for every $\sigma$ which lies on $(q,F)$, there is a finite set $\mathscr X_\sigma$ of conditions $p\leq q\mid\sigma$. Then there is a condition $(\bar q,\bar F)\leq (q,F)$, also determined, such that for every $\sigma$ which lies on $(q,F)$ and $p\in \mathscr X_\sigma$, there is $\tau$ which lies on $(\bar q,\bar F)$ and extends $\sigma$ having $\bar q\mid\tau\leq p$. Moreover, for every $\tau$ which lies on $(\bar q,\bar F)$, $\bar q\mid\tau\leq p$ for some $p\in \mathscr X_\sigma$ with $\tau$ extending $\sigma$.
First, suppose $n=1$. Let $(T,n)$ be a condition in ${\mathbb{Q}}({\mathbb{P}})$ and suppose that for every node $t$ on level $n$ of $T$, we have a finite set $\mathscr X_t$ of trees $S\leq T_t$. Let $T'$ be the tree we get by replacing $T_t$ with the union of $S$ in $\mathscr X_t$ for every $t$ on level $n$ of $T$. Fix $t$ on level $n$ of $T'$. Find a large enough level $n_t\geq n$ such that for every $s$ on level $n _t$ of $T'$, we can thin out $T'_s$ to $P^s$ with the property that $P^s\leq S$ for some $S\in \mathscr X_t$, and moreover for every $S\in \mathscr X_t$, there is some node $s$ such that $P^s\leq S$. Let $\bar T$ be the tree we obtain by replacing $T'_s$ with $P^s$ in $T'$. Let $\bar n\geq n_t$ for every $t$ on level $n$ of $T'$. Clearly the condition $(\bar T,\bar n)$ has all the desired properties.
Next, let’s consider the case $n=2$. Let $(q,F)$ be a condition in ${\mathbb{Q}}({\mathbb{P}}_2)$ and suppose that for every $\sigma$ which lies on $(q,F)$, we have a finite set $\mathscr X_\sigma$ of conditions $p\leq q\mid\sigma$. For every node $t$ on level $F(0)$ of $q(0)$, let $\mathscr X_t$ be the set of all trees $p(0)$ with $p\in \mathscr X_\sigma$ for $\sigma(0)=t$. Using the case $n=1$, let $(\bar T,\bar n)$ be the condition for the sets $\mathscr X_t$. Fix a node $s$ on level $\bar n$ of $\bar T$ and let $t_s$ be the node on level $F(0)$ of $\bar T$ which $s$ extends. Since $(q,F)$ was determined, $\bar T_s$ decides the $F(1)$-th level of $q(1)$. So let $\dot T^s$ be a ${\mathbb{P}}_1$-name for the tree we get by replacing, for every node $t$ on level $F(1)$ of $q(1)$, each $q(1)_t$ with the union of $p(1)$ for $p\in \mathscr X_\sigma$ with $\sigma=\{(0,t_s),(1,t)\}$ and $\bar T_s\leq p(0)$. Strengthen $\bar T_s$ to $P^s$ deciding a level $m_s\geq F(1)$ such that for every node $u$ on level $m_s$ of $\dot T^s$ above a node $t$ on level $F(1)$, we can thin out $\dot T^s_u$ to $\dot P^{s,u}$ such that $\dot P^{s,u}\leq p(1)$ for some $p\in \mathscr X_\sigma$ with $\sigma=\{(0,t_s),(1,t)\}$ and $\bar T_s\leq p(0)$, and moreover for every such $p(1)$, there is some node $u$ such that $\dot P^{s,u}\leq p(1)$. We can assume, with some more thinning out, that all $m_s=m$ for a fixed $m$. Let $\bar T'$ be the tree we get by replacing each $\bar T_s$ with $P^s$ in $\bar T$. Let $\dot T^{'s}$ be a ${\mathbb{P}}_1$-name for the tree we get by replacing, for every node $u$ on level $m$ of $\dot T^s$, each $\dot T^s_u$ with $\dot P^{s,u}$. Let $\bar F=\{(0,\bar n),(1,m)\}$. Let $X_{\bar F}$ consist of $\sigma:2\to{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}2$ such that $\sigma(0)=s$ on level $\bar n$ of $\bar T'$ and $\sigma(1)=u$ on level $m$ of $\dot T^{'s}$. Let $\varphi$ be the $X_{\bar F}$-assignment such that $\varphi(\sigma)=\{(0,\bar T'_{\sigma(0)}),(1,\dot T^{'\sigma(0)}_{\sigma(1)})\}$. Let $\bar q$ be the amalgamation of $\varphi$. Clearly the condition $(\bar q,\bar F)$ has all the desired properties.
Finally, assuming that the statement holds for $n$, let’s argue that it holds for $n+1$ by mimicking how the argument moves from $n=1$ to $n=2$. Let $(q,F)$ be a condition in ${\mathbb{Q}}({\mathbb{P}}_{n+1})$ and suppose that for every $\sigma$ which lies on $(q,F)$, we have a finite set $\mathscr X_\sigma$ of conditions $p\leq q\mid\sigma$. For every $\tau$ which lies on $(q{\upharpoonright}n,F{\upharpoonright}n)$, let $\mathscr X_\tau$ be the set of all conditions $p{\upharpoonright}n$ with $p\in \mathscr X_\sigma$ for $\sigma{\upharpoonright}n=\tau$. Using the inductive hypothesis for $n$, let $(r,J)$ be the condition for the sets $\mathscr X_\tau$. Fix $\rho$ lying on $(r,J)$, and find $\tau_\rho$ which lies on $(q{\upharpoonright}n,F{\upharpoonright}n)$ such that $\rho$ extends $\tau_\rho$. Since $(q,F)$ was determined, $r\mid\rho$ decides the $F(n)$-th level of $q(n)$. So let $\dot T^\rho$ be a ${\mathbb{P}}_n$-name for the tree we get by replacing, for every node $t$ on level $F(n)$ of $q(n)$, each $q(n)_t$ with the union of $p(n)$ for $p\in \mathscr X_\sigma$ with $\sigma=\tau_\rho{\cup}\{(n,t)\}$ and $r\mid\rho\leq p{\upharpoonright}n$. Strengthen $r\mid \rho$ to $r'_\rho$ deciding a level $m_\rho\geq F(n)$ such that for every node $u$ on level $m_\rho$ of $\dot T^\rho$ above a node $t$ on level $F(n)$, we can thin out $\dot T^\rho_u$ to $\dot P^{\rho,u}$ such that $\dot P^{\rho,u}\leq p(n)$ for some $p\in \mathscr X_\sigma$ with $\sigma=\tau_\rho{\cup}\{(n,t)\}$ and $r\mid\rho\leq p{\upharpoonright}n$, and moreover for every such $p(n)$, there is some node $u$ with $\dot P^{\rho,u}\leq p(n)$. We can assume, with some more thinning out, that all $m_\rho=m$ for a fixed $m$. Let $(r',J')$ be the condition we get by applying the inductive hypothesis to $(r,J)$ and $\mathscr Y_\rho=\{r'_\rho\}$. Let $\dot T^{'\rho}$ a ${\mathbb{P}}_n$-name for the tree we get by replacing, for every node $u$ on level $m$ of $\dot T^\rho$, each $\dot T^\rho_u$ with $\dot P^{\rho,u}$. Let $\bar F=J'{\cup}\{(n,m)\}$. Let $X_{\bar F}$ consist of $\sigma:n+1\to{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}2$ such that $\sigma=\tau{\cup}\{( n,u)\}$ for some $\tau$ which lies on $(r',J')$ with $\tau$ extending $\rho$ lying on $(r,J)$ and $u$ on level $m$ of $\dot T^{'\rho}$. Let $\varphi$ be the $X_{\bar F}$-assignment such that $\varphi(\sigma)=r'\mid(\sigma{\upharpoonright}n){\cup}\{( n,T^{'\rho}_{\sigma(n)})\}$, where $\sigma{\upharpoonright}n$ extends $\rho$ lying on $(r,J)$. Let $\bar q$ be the amalgamation of $\varphi$. Clearly the condition $(\bar q,\bar F)$ has all the desired properties.
Growing finite iterations of perfect posets {#sec:GrowingIterations}
===========================================
In the construction of Jensen’s poset ${\mathbb{P}}^J$, at nontrivial stages $\alpha$, we used the $\omega$-many perfect trees obtained from a partially generic filter for ${\mathbb{Q}}({\mathbb{P}}_\alpha)^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$ to grow the perfect poset ${\mathbb{P}}_\alpha$ to ${\mathbb{P}}_{\alpha+1}$. What we would like to do now is to find an appropriate generalization of this construction for growing a finite iteration ${\mathbb{P}}_n$ of perfect posets using partially generic filters for the associated fusion poset ${\mathbb{Q}}({\mathbb{P}}_n)$. More precisely, we would like the following.
Given a finite iteration ${\mathbb{P}}_n={\mathbb{Q}}_0*\dot {\mathbb{Q}}_1*\cdots*\dot {\mathbb{Q}}_{n-1}$ of perfect posets, we would like to be able to extend it to a finite iteration ${\mathbb{P}}_n^*={\mathbb{Q}}_0^**\dot{\mathbb{Q}}_1^**\cdots*\dot{\mathbb{Q}}_{n-1}^*$ of perfect posets, constructed in a generic extension of some suitable model $M$, with the following properties:
1. ${\mathbb{Q}}_0\subseteq {\mathbb{Q}}_0^*$,
2. For all $1\leq i<n$, ${\mathop{1\hskip-2.5pt {\rm l}}}_{{\mathbb{P}}_i^*}$ forces that $\dot{\mathbb{Q}}_i$ is a perfect poset and $\dot{\mathbb{Q}}_i^*$ extends it.
3. ${\mathbb{P}}_n\subseteq {\mathbb{P}}_n^*$.
4. Every maximal antichain $\mathcal A\in M$ of ${\mathbb{P}}_n$ remains maximal in ${\mathbb{P}}_n^*$.
The next theorem from [@abraham:jensenRealsIterations] holds the main idea for constructing ${\mathbb{P}}_n^*$. The set-up for the theorem is left intentionally vague with the details forthcoming in the next section.
Fix a suitable model $M$ with ${\mathbb{P}}_n\in M$. We carry out the construction of ${\mathbb{P}}_n^*$ in $n$-steps, at each step $i$, constructing a ${\mathbb{P}}_i^*$-name $\dot{\mathbb{Q}}_i^*$ for a perfect poset extending $\dot{\mathbb{Q}}_i$. We extend ${\mathbb{Q}}_0$ to ${\mathbb{Q}}_0^*$ as before in a (carefully chosen) forcing extension of $M$ by ${\mathbb{Q}}({\mathbb{Q}}_0)^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$. So suppose inductively that we already extended ${\mathbb{P}}_i$ to ${\mathbb{P}}_i^*$ satisfying requirements (1)-(4). Thus, in particular, every maximal antichain $\mathcal A\in M$ of ${\mathbb{P}}_i$ remains maximal in ${\mathbb{P}}_i^*$, and so every $V$-generic filter $H^*$ for ${\mathbb{P}}_i^*$ restricts to an $M$-generic filter $H$ for ${\mathbb{P}}_i$. From this it follows that $\dot {\mathbb{Q}}_i$ is forced to be a perfect poset by ${\mathbb{P}}_i^*$ because $M[H]$ satisfies that $(\dot {\mathbb{Q}}_i)_H$ is a perfect poset, and this statement is clearly absolute between $M[H]$ and $V[H^*]$. Fix a carefully chosen $M$-generic filter $G\subseteq {\mathbb{Q}}({\mathbb{P}}_{i+1})$. Let $\tau(G)$ be a ${\mathbb{P}}_i^*$-name for a subset of ${\mathbb{Q}}(\dot {\mathbb{Q}}_i)$ such that in any forcing extension $V[H^*]$ by ${\mathbb{P}}_i^*$ $\tau(G)$ gets interpreted as the collection of all pairs $(p(i)_{H^*},F(i))$ for $(p,F)\in G$. Provided that the poset ${\mathbb{P}}_i^*$ contains a kind of master condition for $G$, we will be able to conclude that $\tau(G)$ is $M[H]$-generic for ${\mathbb{Q}}((\dot{\mathbb{Q}}_i)_H)$, where $H$ is the restriction of the $V$-generic filter $H^*\subseteq {\mathbb{P}}_i^*$.
\[th:genericCondition\] Suppose $\bar p\in {\mathbb{P}}^*_i$ is such that for every $(p,F)\in G$, $\bar p\leq p{\upharpoonright}i$. Then
$\bar p{\Vdash}``\tau(G)$ is an $M[\dot H]$-generic filter for ${\mathbb{Q}}(\dot{\mathbb{Q}}_i)$",
where $\dot H$ is the canonical name for the restriction of the generic filter to ${\mathbb{P}}_i$.
Suppose $H^*\subseteq {\mathbb{P}}^*_i$ is $V$-generic with $\bar p\in H^*$. Let $H$ be the restriction of $H^*$ to ${\mathbb{P}}_i$. Let $$K=\tau(G)_{H^*}=\{(p(i)_{H^*},F(i))\mid p\in G\}$$ and ${\mathbb{Q}}_i=(\dot{\mathbb{Q}}_i)_{H}$. First, we argue that $K$ is a filter on ${\mathbb{Q}}({\mathbb{Q}}_i)$. Suppose for $(p,F)\in G$, $(p(i)_{H},F(i))\leq (T,b)$ in ${\mathbb{Q}}({\mathbb{Q}}_i)$, so that $b\leq F(i)$ and $p(i)_{H}\cap{{}^{b}}2=T\cap{{}^{b}}2$. It follows that there is a ${\mathbb{P}}_i$-name $\dot T$ for $T$ such that
${\mathop{1\hskip-2.5pt {\rm l}}}_{{\mathbb{P}}_i}{\Vdash}``p(i)\leq \dot T\text{ and } p(i)\cap{{}^{b}}2=\dot T\cap{{}^{b}}2"$.
Let $p'=p{\upharpoonright}i{\cup}\{(i,\dot T)\}$ and $F'=F{\cup}\{(i,b)\}$. Clearly $(p,F)\leq (p',F')$, which means that $(p',F')\in G$. It follows that $p'(i)_{H}=T$, and so $(T,b)\in K$.
Next, we fix $(p,F)$ and $(p',F')$ both in $G$ and argue that $(p(i)_{H},F(i))$ and $(p'(i)_{H},F'(i))$ are compatible in $K$. There is $(q,J)\in G$ below both $(p,F)$ and $(p',F')$. It follows that $J(i)\geq F(i),F'(i)$, $q{\upharpoonright}i{\Vdash}_{{\mathbb{P}}_i} q(i)\leq p(i),p'(i)$,
$q{\upharpoonright}i{\Vdash}_{{\mathbb{P}}_i} ``q(i)\cap {{}^{F(i)}}2=p(i)\cap{{}^{F(i)}}2\text{ and }q(i)\cap {{}^{F'(i)}}2=p'(i)\cap{{}^{F'(i)}}2"$,
and $(q(i)_{H},J(i))\in K$. Since $\bar p\leq q{\upharpoonright}i$ and $\bar p\in H^*$, we have $q{\upharpoonright}i\in H^*$. Now observe that $q{\upharpoonright}i$ must force the statements mentioned above over ${\mathbb{P}}^*_i$ as well because any $V$-generic filter for ${\mathbb{P}}_i^*$ restricts to an $M$-generic filter for ${\mathbb{P}}_i$ and the statements in question are absolute. Thus, $$q(i)_{H}\leq p(i)_{H},p'(i)_{H}$$ and $$q(i)_{H}\cap {{}^{F(i)}}2=p(i)_{H}\cap{{}^{F(i)}}2\text{ and }q(i)_{H}\cap {{}^{F'(i)}}2=p'(i)_{H}\cap{{}^{F'(i)}}2.$$ So $(q(i)_{H},J(i))\leq (p(i)_{H},F(i)),\,(p'(i)_{H},F'(i))$.
Finally, we have to see that $K$ is $M[H]$-generic. So suppose $D\in M[H]$ is dense open in ${\mathbb{Q}}({\mathbb{Q}}_i)$. Let $\dot D\in M$ be a ${\mathbb{P}}_i$-name for $D$ such that
${\mathop{1\hskip-2.5pt {\rm l}}}_{{\mathbb{P}}_i}{\Vdash}_{{\mathbb{P}}_i} ``\dot D$ is dense open in ${\mathbb{Q}}(\dot {\mathbb{Q}}_i)"$.
In $M$, define $$E=\{(p,F)\in {\mathbb{Q}}({\mathbb{P}}_{i+1})\mid p{\upharpoonright}i{\Vdash}(p(i),F(i))\in \dot D\}.$$ We claim that $E$ is dense open in ${\mathbb{Q}}({\mathbb{P}}_{i+1})$. It is easy to see that $E$ is open, so let’s argue that it is dense. Fix some $(q,J)\in {\mathbb{Q}}({\mathbb{P}}_{i+1})$ and assume without loss that $(q,J)$ is determined. There must be some pair $(\dot T,\dot k)$ such that $\dot T$ is a ${\mathbb{P}}_i$-name for an element of $\dot{\mathbb{Q}}_i$, $\dot k$ is a ${\mathbb{P}}_i$-name for a natural number, and $$q{\upharpoonright}i{\Vdash}``(\dot T,\dot k)\in\dot D\text{ and }(\dot T,\dot k)\leq (q(i),J(i))".$$ The set of conditions which decide the value of $\dot k$ is dense open below $q{\upharpoonright}i$ in ${\mathbb{P}}_i$. So, by Proposition \[prop:InsideDenseOpenSet\], there a determined condition $(p',F')\leq (q{\upharpoonright}i,J{\upharpoonright}i)$ in ${\mathbb{Q}}({\mathbb{P}}_i)$ such that for every $\sigma$ which lies on it, $p'\mid \sigma$ decides that $\dot k=k(\sigma)$. Let $k\in\omega$ be above all the $k(\sigma)$. Then $$p'{\Vdash}``(\dot T, k)\leq (\dot T,\dot k)\leq (q(i),J(i))\text{ and }(\dot T,k)\in \dot D".$$ Let $p=p'{\cup}\{(i,\dot T)\}$ and $F=F'{\cup}\{(i,k)\}$. Clearly $(p,F)\in E$. Let’s argue that $(p,F)\leq (q,J)$. By construction $p\leq q$ and each $k(\sigma)\geq J(i)$, so $k\geq J(i)$. Finally, $p'=p{\upharpoonright}i{\Vdash}(\dot T,k)\leq (q(i),J(i))$ and so $p{\upharpoonright}i{\Vdash}\dot T\cap {{}^{J(i)}}2=q(i)\cap {{}^{J(i)}}2$.
Fix some $(p,F)\in E\cap G$. Since $\bar p\leq p{\upharpoonright}i$, it follows that $p{\upharpoonright}i\in H^*$. Thus, $(p(i)_{H},F(i))\in D\cap K$, completing the argument that $K$ is an $M[H]$-generic filter for ${\mathbb{Q}}({\mathbb{Q}}_i)$.
Next, we are going to obtain a stronger version of Theorem \[th:genericCondition\] that tells us how to get an $M[H]$-generic filter for ${\mathbb{Q}}({\mathbb{Q}}_i)^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$, which is really what we need to extend $\dot {\mathbb{Q}}_i$ to $\dot {\mathbb{Q}}_i^*$. For this, we will need to enlarge our fusion poset ${\mathbb{Q}}({\mathbb{P}}_{i+1})$. Let $\bar{\mathbb{Q}}_i({\mathbb{P}}_{i+1})$ be the following modification of ${\mathbb{Q}}({\mathbb{P}}_{i+1})$. Conditions in $\bar{\mathbb{Q}}({\mathbb{P}}_{i+1})$ are pairs $(p,F)$ such that $(p{\upharpoonright}i,F{\upharpoonright}i)\in {\mathbb{Q}}({\mathbb{P}}_i)$, $p(i)$ is some finite tuple ${\langle}\dot T_0,\dot T_1,\ldots,\dot T_{k-1}{\rangle}$ with $p{\upharpoonright}i{\Vdash}\dot T_j\in \dot{\mathbb{Q}}_i$ for all $j<k$, and $F(i)=f:k\to\omega$. The ordering is $(p',F')\leq (p,F)$ whenever
1. $(p'{\upharpoonright}i,F'{\upharpoonright}i)\leq (p{\upharpoonright}i, F{\upharpoonright}i)$, and\
for $j<k$,
2. $F'(i)(j)\geq F(i)(j)$,
3. $p'{\upharpoonright}i{\Vdash}p'(i)(j)\cap {{}^{F(i)(j)}}2=p(i)(j)\cap {{}^{F(i)(j)}}2$.
The point is that if $H^*\subseteq {\mathbb{P}}_i^*$ is $V$-generic, then $${\langle}((\dot T_0)_{H^*},f(0)),\ldots,((\dot T_{k-1})_{H^*},f(k-1)){\rangle}$$ is a condition in ${\mathbb{Q}}((\dot {\mathbb{Q}}_i)_H)^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$. Suppose now that $G\subseteq \bar {\mathbb{Q}}({\mathbb{P}}_{i+1})$ is $M$-generic. Let $\tau(G)$ be a ${\mathbb{P}}_i^*$-name for a subset of ${\mathbb{Q}}({\mathbb{Q}}_i)^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$ such that in any forcing extension $V[H^*]$ by ${\mathbb{P}}_i^*$, $\tau(G)$ gets interpreted as the collection of all $${\langle}((\dot T_0)_{H^*},f(0)),\ldots,((\dot T_{k-1})_{H^*},f(k-1)){\rangle}$$ for $(p,F)\in G$. If it so happens that ${\mathbb{P}}_i^*$ has a master condition $\bar p$ for $G$, then $\tau(G)$ will be $M[H]$-generic for ${\mathbb{Q}}(({\mathbb{Q}}_i)_H)^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$, where $H$ is the restriction of the $V$-generic filter $H^*\subseteq {\mathbb{P}}_i^*$.
\[th:genericConditionProduct\] Suppose $\bar p\in {\mathbb{P}}^*_i$ is such that for every $(p,F)\in G$, $\bar p\leq p{\upharpoonright}i$. Then
$\bar p{\Vdash}``\tau(G)$ is an $M[\dot H]$-generic filter for ${\mathbb{Q}}(\dot{\mathbb{Q}}_i)^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$",
where $\dot H$ is the canonical name for the restriction of the generic filter to ${\mathbb{P}}_i$.
The proof is essentially the same as of Theorem \[th:genericCondition\]. Using Theorem \[th:genericConditionProduct\], we can let $\dot{\mathbb{Q}}_i^*$ be the canonical ${\mathbb{P}}_i^*$-name for the extension of $\dot {\mathbb{Q}}_i$ formed in $M[\dot H][\tau(G)]$, where $\dot H$ is the restriction of the generic filter to ${\mathbb{P}}_i$. In the next section, we will show how to obtain the required $M$-generic filters $G$ so that the inductive assumptions hold for ${\mathbb{P}}_i^*$.
Tree iterations of perfect posets {#sec:treeIterations}
=================================
Let’s define that an *$\omega$-iteration of perfect posets* is a sequence $$\vec P={\langle}{\mathbb{P}}_n\mid n<\omega{\rangle},$$ where ${\mathbb{P}}_0=\{\emptyset\}$ is a trivial poset and each ${\mathbb{P}}_n$, for $n\geq 1$, is a finite iteration of perfect posets with the coherence requirement that for $0<m<n$, ${\mathbb{P}}_n{\upharpoonright}m={\mathbb{P}}_m$. The initial poset ${\mathbb{P}}_0$ is included in the sequence to make the subsequent definitions more uniform. For this reason, we will also make the ad hoc definition that ${\mathbb{Q}}({\mathbb{P}}_0)=\{\emptyset\}$. Note that an $\omega$-iteration of perfect posets is not itself an iteration, rather it is a coherent sequence of finite iterations.
A *tree iteration* is a non-linear forcing iteration along some tree. Given a tree of height $\omega$, the tree iteration of perfect posets will use an $\omega$-iteration $\vec P={\langle}{\mathbb{P}}_n\mid n<\omega{\rangle}$ of perfect posets with conditions assigned to level $n$ nodes of the tree coming from the poset ${\mathbb{P}}_n$. Conditions will be assigned to the nodes coherently so that if a node $s$ on level $n$ extends a node $t$ on level $m$, then the condition $p$ on node $s$ will be such that $p{\upharpoonright}m$ is on node $t$.
Let $\vec P={\langle}{\mathbb{P}}_n\mid n<\omega{\rangle}$ be an $\omega$-iteration of perfect posets and let $\mathscr T$ be a tree of height $\omega$. A *$\mathscr T$-iteration of perfect posets* is the following partial order ${\mathbb{P}}(\vec P,\mathscr T)$. Conditions in ${\mathbb{P}}(\vec P,\mathscr T)$ are functions $f_X$ with domain some finite subtree $X$ of $\mathscr T$ such that:
1. For every node $s$ on level $n$ of $X$, $f_X(s)\in {\mathbb{P}}_n$.
2. Whenever $s\leq t$ are two nodes in $X$, then $f_X(t){\upharpoonright}{{\rm lev}}(s)=f_X(s)$.
The ordering is $f_Y\leq f_X$ whenever $Y$ extends $X$ and for every node $s\in X$, $f_Y(s)\leq f_X(s)$.
The analogue of the fusion posets ${\mathbb{Q}}({\mathbb{P}})$ and ${\mathbb{Q}}({\mathbb{P}}_n)$ for ${\mathbb{P}}(\vec P,\mathscr T)$ will be the fusion poset ${\mathbb{Q}}(\vec P,\mathscr T)$ whose conditions are functions $f_X$ with domain some finite subtree $X$ of $\mathscr T$ such that:
1. For every node $s$ on level $n$ of $X$, $f_X(s)\in {\mathbb{Q}}({\mathbb{P}}_n)$.
2. Whenever $s\leq t$ are two nodes in $X$, with $f_X(s)=(p_s,F_s)$ and $f_X(t)=(p_t,F_t)$, then $p_t{\upharpoonright}{{\rm lev}}(s)=p_s$ and $F_t{\upharpoonright}{{\rm lev}}(s)=F_s$.
The ordering is $f_Y\leq f_X$ whenever $Y$ extends $X$ and for every node $s\in X$, $f_Y(s)\leq f_X(s)$.
\[prop:strengthenTreeCondition\] $\,$
1. Suppose $f_X\in {\mathbb{P}}(\vec P, \mathscr T)$, $f_X(s)=p$, and $q\leq p$. Then there is a condition $g_X\leq f_X$ in ${\mathbb{P}}(\vec P,\mathscr T)$ with $g_X(s)=q$.
2. Suppose $f_X\in {\mathbb{Q}}(\vec P,\mathscr T)$, $f_X(s)=(p,F)$, and $(q,F')\leq (p,F)$. Then there is $g_X\leq p_X$ such that $g_X(s)=(q,F')$.
We will only prove (1) because the proof of (2) is analogous. Define $g_X$ as follows. Fix a node $t\in X$. If $t\leq s$, then let $g_X(t)=q{\upharpoonright}{{\rm lev}}(t)$. If $s\leq t$, then let $g_X(t)$ be $q$ concatenated with the tail of $f_X(t)$. Otherwise, let $t'$ be the largest node that is compatible with both $t$ and $s$. Let $g_X(t)$ be $q{\upharpoonright}{{\rm lev}}(t')$ concatenated with the tail of $f_X(t)$.
Let us call a condition $f_X\in {\mathbb{Q}}(\vec P,\mathscr T)$ *determined* if every $f_X(s)$ is determined. Clearly a condition $f_X$ is determined if and only if conditions on the terminal nodes are determined.
\[prop:determinedInTreeDense\] The set of all determined conditions is dense in ${\mathbb{Q}}(\vec P,\mathscr T)$.
Fix $f_X\in {\mathbb{Q}}(\vec P,\mathscr T)$. Let ${\langle}s_i\mid i\leq m{\rangle}$ be an enumeration of the terminal nodes of $X$. Using the construction in the proof of Proposition \[prop:strengthenTreeCondition\], strengthen $f_X$ to $f^0_X$ such that $f^0_X(s_0)$ is determined. Inductively, given $f^i_X$, we let $f^{i+1}_X$ be the strengthening of $f^i_X$ constructed as in the proof of Proposition \[prop:strengthenTreeCondition\] such that $f^{i+1}_X(s_{i+1})$ is determined. Let’s argue that $f^m_X$ is determined. We can assume inductively that the conditions $f^{m-1}_X(s_i)$ for $i<m$ are determined. By construction $f^m_X(s_m)$ is determined. So fix $i<m$. Let $s$ be the node where the branch of $s_i$ and the branch of $s_m$ split. By construction, $f^m_X(s_i)$ is $f^m_X(s_m){\upharpoonright}{{\rm lev}}(s)$ concatenated with the tail of $f^{m-1}_X(s_i)$. Clearly $f^m_X(s_m){\upharpoonright}{{\rm lev}}(s)$ is determined and $f^m_X(s_m){\upharpoonright}{{\rm lev}}(s)\leq f^{m-1}_X(s_i){\upharpoonright}{{\rm lev}}(s)$. Now using the definition of what it means to be determined, it is easy to see that the whole condition is determined.
We will initially consider tree iterations along the countable tree ${{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega$, and then later extend our results to tree iterations along the uncountable tree ${{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1$.
It is easy to see that the poset ${\mathbb{Q}}({\mathbb{Q}}_0)^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$ completely embeds into ${\mathbb{Q}}(\vec P,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$ via the map sending a condition to the corresponding tree of height $\leq 2$.
\[prop:completeEmbeddingProductTreeLevel1\] The poset ${\mathbb{Q}}({\mathbb{Q}}_0)^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$ completely embeds into ${\mathbb{Q}}(\vec P,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$ via the following map $\varphi$:
1. $\varphi({\mathop{1\hskip-2.5pt {\rm l}}}_{{\mathbb{Q}}({\mathbb{Q}}_0)})=f_X$, where $X$ consists of the root node $s$ and $f_X(s)=\emptyset$.\
For $p$ with some non-trivial $p(i)$,
2. $\varphi(p)=f_X$, where $X$ consists of the root node together with nodes ${\langle}i{\rangle}$ for non-trivial $p(i)$, such that $f_X({\langle}i{\rangle})=p(i)$.
More generally, for each node $s$ on level $n$ of ${{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega$, $\bar{\mathbb{Q}}({\mathbb{P}}_{n+1})$ completely embeds into ${\mathbb{Q}}(\vec P,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$ via the map sending a condition to the corresponding tree of height $\leq n+2$ whose stem stretches up to $s$.
\[prop:completeEmbeddingProductTree\] Fix a node $s$ on level $n$ of ${{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}{\omega}$. The poset $\bar{\mathbb{Q}}({\mathbb{P}}_{n+1})$ completely embeds into ${\mathbb{Q}}(\vec P,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$ via the following map $\varphi_s$:
1. $\varphi_s({\mathop{1\hskip-2.5pt {\rm l}}}_{\bar {\mathbb{Q}}({\mathbb{P}}_{n+1})})=f_X$, where $X$ is the branch ending in $s$, such that $f_X(s)={\mathop{1\hskip-2.5pt {\rm l}}}_{{\mathbb{Q}}({\mathbb{P}}_n)}$.\
For $(p,F)$ with $\text{dom}(p(n))=k$,
2. $\varphi_s((p,F))=f_X$, where $X$ consists of the branch ending in $s$ together with nodes ${\langle}s{\mathbin{{}^\smallfrown}}i{\rangle}$ for $i<k$, such that $f_X(s)=(p{\upharpoonright}n,F{\upharpoonright}n)$ and
$f_X(s{\mathbin{{}^\smallfrown}}i)=(p{\upharpoonright}n{\mathbin{{}^\smallfrown}}p(n)(i),F{\upharpoonright}n{\mathbin{{}^\smallfrown}}F(n)(i))$.
Suppose $G\subseteq {\mathbb{Q}}(\vec P,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$ is $V$-generic and fix some node $s$ on level $n$ of ${{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}{\omega}$. We will use the notation $G_s$ for the $V$-generic filter for $\bar{\mathbb{Q}}({\mathbb{P}}_{n+1})$ added by $G$ via the embedding $\varphi_s$ and we will use the notation $G_\emptyset$ for the $V$-generic filter for ${\mathbb{Q}}({\mathbb{Q}}_0)^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$ added by $G$ via the embedding $\varphi$.
Suppose now that $\vec P={\langle}{\mathbb{P}}_n\mid n<\omega{\rangle}$ is an $\omega$-iteration of perfect posets that is an element of a suitable model $M$. Let $G\subseteq {\mathbb{Q}}(\vec P,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$ be $M$-generic. We will argue that in $M[G]$, we can grow each iteration ${\mathbb{P}}_n$ to an iteration ${\mathbb{P}}_n^*$ satisfying requirements (1)-(4) from Section \[sec:GrowingIterations\]:
1. ${\mathbb{Q}}_0\subseteq {\mathbb{Q}}_0^*$,
2. For all $1\leq i<n$, ${\mathop{1\hskip-2.5pt {\rm l}}}_{{\mathbb{P}}_i^*}$ forces that $\dot{\mathbb{Q}}_i$ is a perfect poset and $\dot{\mathbb{Q}}_i^*$ extends it.
3. ${\mathbb{P}}_n\subseteq {\mathbb{P}}_n^*$.
4. Every maximal antichain $\mathcal A\in M$ of ${\mathbb{P}}_n$ remains maximal in ${\mathbb{P}}_n^*$.
It is straightforward to extend ${\mathbb{Q}}_0$ to ${\mathbb{Q}}^*_0$. By Proposition \[prop:completeEmbeddingProductTreeLevel1\], $G$ adds an $M$-generic filter $G_\emptyset$ for ${\mathbb{Q}}({\mathbb{Q}}_0)^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$. Let ${\mathcal T}_i^0$ for $i<\omega$ be the generic perfect trees added by $G_\emptyset$ and construct ${\mathbb{Q}}_0^*$ as before. Recall that ${\langle}{\mathcal T}_i^0\mid i<\omega{\rangle}$ is a maximal antichain in ${\mathbb{Q}}_0^*$ and every maximal antichain $\mathcal A\in M$ of ${\mathbb{Q}}_0$ remains maximal in ${\mathbb{Q}}_0^*$.
Now let’s show how to extend $\dot{\mathbb{Q}}_1$ to a ${\mathbb{P}}_1^*$-name $\dot{\mathbb{Q}}^*_1$ for a perfect poset. By Proposition \[prop:completeEmbeddingProductTree\], for each node $s$ on level 1 of the tree ${{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega$, $G$ adds an $M$-generic filter $G_s$ for $\bar{\mathbb{Q}}({\mathbb{P}}_2)$. Observe that each ${\mathcal T}_i^0\leq p{\upharpoonright}1$ for all $p$ with $(p,F)\in G_{{\langle}i{\rangle}}$. Thus, by Theorem \[th:genericConditionProduct\], whenever $H^*\subseteq {\mathbb{P}}_1^*$ is a $V$-generic filter containing ${\mathcal T}_i^0$, then the interpretation $\tau(G_{{\langle}i{\rangle}})_{H^*}$ is an $M[H]$-generic filter for ${\mathbb{Q}}((\dot{\mathbb{Q}}_1)_{H})^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$, where $H$ is the restriction of $H^*$ to ${\mathbb{P}}_1$. Let $\tau$ be a mixed ${\mathbb{P}}_1^*$-name that is interpreted as $\tau(G_{{\langle}i{\rangle}})_{H^*}$, whenever ${\mathcal T}_i^0\in H^*$. Since the conditions ${\mathcal T}_i^0$ form a maximal antichain in ${\mathbb{P}}_1^*$,
${\mathop{1\hskip-2.5pt {\rm l}}}_{{\mathbb{P}}_1^*}{\Vdash}``\tau$ is an $M[\dot H]$-generic filter for ${\mathbb{Q}}(\dot{\mathbb{Q}}_1)^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}$,"
where $\dot H$ is the canonical name for the restriction of the generic filter to ${\mathbb{P}}_1$. So let $\dot {\mathbb{Q}}_1^*$ be a ${\mathbb{P}}_1^*$-name for the perfect poset constructed as usual from $\dot {\mathbb{Q}}_1$ and $\tau$.
For each $j<\omega$, we can choose a ${\mathbb{P}}_1^*$-name $\dot {\mathcal T}_j^1$ for the perfect tree on coordinate $j$ of $\tau$. The pairs $({\mathcal T}_i^0,\dot {\mathcal T}_j^1)$ form a maximal antichain in ${\mathbb{P}}_2^*$ because the collection of ${\mathcal T}_i^0$ is a maximal antichain and each ${\mathcal T}_i^0$ forces that the trees $\dot T_j^1$ form a maximal antichain in $\dot{\mathbb{Q}}_1^*$. Also, clearly for every $(p,F)\in G_{{\langle}i, j{\rangle}}$, we have $({\mathcal T}_i^0,\dot {\mathcal T}_j^1)\leq p{\upharpoonright}2$. This will allow us to apply Theorem \[th:genericConditionProduct\] to grow $\dot {\mathbb{Q}}_2$ to $\dot {\mathbb{Q}}_2^*$. Finally, let’s argue that ${\mathbb{P}}_2$ is actually a subset of ${\mathbb{P}}_2^*$. Suppose $p$ is a condition in ${\mathbb{P}}_2$. Then $p(0)\in {\mathbb{P}}_1$ and hence $p(0)\in {\mathbb{P}}_1^*$ as well. Also, clearly $p(1)$ is a ${\mathbb{P}}_1^*$-name. So we need to argue that $p(0){\Vdash}_{{\mathbb{P}}_1^*}p(1)\in \dot{\mathbb{Q}}_1^*$. Fix a $V$-generic filter $H^*\subseteq {\mathbb{P}}_1^*$ with $p(0)\in H^*$ and consider $M[H]$ where $H$ is the restriction of $H^*$ to ${\mathbb{P}}_1$. Since $p(0){\Vdash}_{{\mathbb{P}}_1} p(1)\in \dot {\mathbb{Q}}_1$, in $M[H]$, we have $p(1)_H\in (\dot {\mathbb{Q}}_1)_H\subseteq (\dot {\mathbb{Q}}_1^*)_{H^*}$. It remains to show (4), that every maximal antichain of ${\mathbb{P}}_2$ from $M$ remains maximal in ${\mathbb{P}}_2^*$. We will provide an inductive proof of this later in Lemma \[lem:antichainSuccessorCase\].
For now to finish the construction, we assume that properties (1)-(4) hold for ${\mathbb{P}}_n^*$. We will additionally assume that:
1. For each node $s$ on level $n$ of ${{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega$ there is a condition $$({\mathcal T}_{s(0)}^0,\dot{\mathcal T}_{s(1)}^1,\ldots,\dot {\mathcal T}_{s(n-1)}^{n-1})\in {\mathbb{P}}_n^*$$ which is below all $p{\upharpoonright}n$ with $(p,F)\in G_s$.
2. The collection of all such conditions $({\mathcal T}_{s(0)}^0,\dot{\mathcal T}_{s(1)}^1,\ldots,\dot {\mathcal T}_{s(n-1)}^{n-1})$ is a maximal antichain in ${\mathbb{P}}_n^*$.
With this set-up, we extend $\dot {\mathbb{Q}}_n$ to $\dot {\mathbb{Q}}_n^*$ identically to the case $n=1$ above, using Theorem \[th:genericConditionProduct\]. It is also easy to see inductively that ${\mathbb{P}}_n$ is a subset of ${\mathbb{P}}_n^*$.
To get some intuition for the construction, let’s fix an $L$-generic filter $H^*\subseteq{\mathbb{P}}_1^*$ and see what a condition ${\mathcal T}^1_j=(\dot{\mathcal T}^1_j)_{H^*}$ looks like. Since ${\langle}{\mathcal T}^0_i\mid i<\omega{\rangle}$ is a maximal antichain of ${\mathbb{P}}_1^*$, there is a unique $i$ such that ${\mathcal T}^0_i\in H^*$. Let $$K=\{(p(1)_{H^*},F(1))\mid f_X\in G\text{ and }f_X({\langle}i,j{\rangle})=(p,F)\}.$$ Then ${\mathcal T}_j^1$ is the union of $T\cap {{}^{n}}2$ for $(T,n)\in K$.
To give the promised argument that every maximal antichain of ${\mathbb{P}}_n$ from $M$ remains maximal in ${\mathbb{P}}_n^*$, we first need to define the analogue of ${\mathbb{U}}\subseteq {\mathbb{P}}^*$ from Section \[sec:PerfectPosets\] for ${\mathbb{P}}_n^*$.
Let ${\mathbb{U}}_n$ be the subset of ${\mathbb{P}}_n^*$ consisting of conditions $p=(p_0,\dot p_1,\ldots,\dot p_{n-1})$ such that $p_0={\mathcal T}^0_{j_0}\wedge S_0$ for some $j_0<\omega$ and $S_0\in {\mathbb{Q}}_0$, and for $i<n$, $p{\upharpoonright}i{\Vdash}\dot p_i=\dot{\mathcal T}^i_{j_i}\wedge \dot S_i$ for some $j_i<\omega$ and a ${\mathbb{P}}_i$-name $\dot S_i$ such that ${\mathop{1\hskip-2.5pt {\rm l}}}_{{\mathbb{P}}_i}{\Vdash}\dot S_i\in\dot{\mathbb{Q}}_i$.
\[prop:UnDenseInP\*n\] ${\mathbb{U}}_n$ is dense in ${\mathbb{P}}^*_n$.
We argue by induction on $n$. By Proposition \[prop:propertiesOfP\*\], ${\mathbb{U}}_1$ is dense in ${\mathbb{P}}_1^*$ and every maximal antichain of ${\mathbb{P}}_1$ from $M$ remains maximal in ${\mathbb{P}}_1^*$. So let’s suppose inductively that ${\mathbb{U}}_n$ is dense in ${\mathbb{P}}_n^*$ and every maximal antichain of ${\mathbb{P}}_n$ from $M$ remains maximal in ${\mathbb{P}}_n^*$. This argument is meant to take place simultaneously with the inductive proof of Lemma \[lem:antichainSuccessorCase\], where we use the density of ${\mathbb{U}}_n$ to argue that maximal antichains stay maximal.
Fix a condition $(q_0',\dot q_1',\ldots, \dot q_n')$ in ${\mathbb{P}}_{n+1}^*$. In particular, we have $(q_0',\dot q_1',\ldots,\dot q_{n-1}'){\Vdash}\dot q_n'\in {\mathbb{Q}}_n^*$. Consider any forcing extension $V[H^*]$ by ${\mathbb{P}}_n^*$ with $(q_0',\dot q_1',\ldots,\dot q_{n-1}')\in H^*$. In $V[H^*]$, there is $j_n<\omega$ and $S\in (\dot {\mathbb{Q}}_n)_{H^*}$ such that $(\dot {\mathcal T}_{j_n}^n)_{H^*}\wedge S\leq (q_n')_{H^*}$. So $S\in M[H]$, where $H$ is the restriction of $H^*$ to ${\mathbb{P}}_n$. Let $\dot S$ be a ${\mathbb{P}}_n$-name in $M$ such that $(\dot S)_H=S$ and let $p\in H$ force that $\dot S\in \dot{\mathbb{Q}}_n$. Let $\mathcal A\in M$ be any maximal antichain of ${\mathbb{P}}_n$ extending $\{p\}$ and let $\dot S_n$ be the mixed name such that $p{\Vdash}\dot S=\dot S_n$ and every other $q\in \mathcal A$ forces that $\dot S_n={{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}2$. So ${\mathop{1\hskip-2.5pt {\rm l}}}_{{\mathbb{P}}_n}{\Vdash}\dot S_n\in \dot{\mathbb{Q}}_n$. Now let $(q_0,\dot q_1,\ldots,\dot q_{n-1})\leq (q_0',\dot q_1',\ldots,\dot q_{n-1}')$ be a condition in $H^*$ forcing that $\dot{{\mathcal T}}^n_{j_n}\wedge \dot S_n\leq \dot q_n'$. We have just argued that there is a condition $(q_0,\dot q_1,\ldots,\dot q_{n-1})\leq (q_0',\dot q_1',\ldots,\dot q_{n-1}')$ and a ${\mathbb{P}}_n$-name $\dot S_n$ such that ${\mathop{1\hskip-2.5pt {\rm l}}}_{{\mathbb{P}}_n}{\Vdash}\dot S_n\in \dot{\mathbb{Q}}_n$ and $$(q_0,\dot q_1,\ldots,\dot q_{n-1}){\Vdash}\dot{{\mathcal T}}^n_{j_n}\wedge \dot S_n\leq \dot q_n'$$ for some $j_n<\omega$.
Now, by the inductive assumption, there is a condition $$(p_0,\dot p_1,\ldots,\dot p_{n-1})\leq (q_0,\dot q_1,\ldots, \dot q_{n-1})$$ in ${\mathbb{U}}_n$. So clearly the condition $(p_0,\dot p_1,\ldots,\dot p_{n-1},\dot{{\mathcal T}}^n_{j_n}\wedge \dot S_n)\in{\mathbb{U}}_{n+1}$.
We will usually abuse notation by writing conditions in ${\mathbb{U}}_n$ in the form $$({\mathcal T}^0_{j_0}\wedge S_0,\dot {\mathcal T}^1_{j_1}\wedge \dot S_1,\ldots,\dot {\mathcal T}^{n-1}_{j_{n-1}}\wedge \dot S_{n-1}).$$ The next lemma is a generalization of Proposition \[prop:subtreeContainedInS\].
\[lem:subtreeContainedinSiteration\] Suppose $f_X\in {\mathbb{Q}}(\vec P,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$ forces that $$({\mathcal T}^0_{j_0}\wedge S_0,\dot {\mathcal T}_{j_1}^1\wedge \dot S_1,\ldots,\dot {\mathcal T}_{j_{n-1}}^{n-1}\wedge \dot S_{n-1})\in {\mathbb{U}}_n.$$ Then there is a condition $g_Y\leq f_X$ such that $g_Y({\langle}j_0,j_1,\ldots,j_{n-1}{\rangle})=(q,F)$ is determined and $\tau:n\to {{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}2$ lies on $(q,F)$ such that
$q(0)_{\tau(0)}\leq S_0$, and for all $1\leq i<n$, $(q{\upharpoonright}i)\mid(\tau{\upharpoonright}i){\Vdash}q(i)_{\tau(i)}\leq \dot S_i$.
The case $n=1$ follows by Proposition \[prop:subtreeContainedInS\]. So suppose $n=2$. Recall that every maximal antichain of ${\mathbb{P}}_1$ from $M$ remains maximal in ${\mathbb{P}}_1^*$. By strengthening $f_X$, we can assume that ${\langle}j_0,j_1{\rangle}\in X$. Let $f_X({\langle}j_0,j_1{\rangle})=(\bar q,\bar F)$. By strengthening further, using the case $n=1$, we can assume that there is a node $s_0$ on level $\bar F(0)$ such that $\bar q(0)_{s_0}\leq S_0$. By strengthening some more, we can assume that $(\bar q,\bar F)$ is determined.
Let’s argue that $\bar q(0)_{s_0}$ forces that for some node $s_1$ on level $\bar F(1)$ of $\bar q(1)$, $\bar q(1)_{s_1}\wedge \dot S_1\neq\emptyset$. If this is not the case, then there is some $T\leq\bar q(0)_{s_0}$ which forces that there is no such node. Let $T'$ be the tree we get by replacing $\bar q(0)_{s_0}$ with $T$ in $\bar q(0)$. Let $\bar q'=(T',\bar q(1))$ and let $f'_X$ be a condition strengthening $f_X$ so that $f'_X({\langle}j_0,j_1{\rangle})=(\bar q',\bar F)$. Let $G$ be an $M$-generic filter for ${\mathbb{Q}}(\vec P,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$ containing $f'_X$ and consider $M[G]$. The tree $R={\mathcal T}_{j_0}^0\wedge S_0\wedge\bar q'(0)_{s_0}(=T)$ is a condition in ${\mathbb{P}}_1^*$, and so we can fix some $V$-generic filter $H^*\subseteq {\mathbb{P}}^*_1$ containing $R$. In particular, $H^*$ contains ${\mathcal T}_{j_0}^0\wedge S_0$ and $T$. Thus, in $M[H]$ (where $H$ is the restriction of $H^*$ to ${\mathbb{P}}_1$), $\bar q(1)_{H}$ does not have any subtree on level $F(1)$ which has a common subtree with $(\dot S_1)_{H}$. But then $(\dot{\mathcal T}_{j_1}^1)_{H^*}\wedge (\dot S_1)_{H}=\emptyset$ contradicting our assumption that $f_X$ forced $({\mathcal T}_{j_0}^0\wedge S_0,\dot {\mathcal T}_{j_1}^1\wedge \dot S_1)\in {\mathbb{U}}_2$.
Let’s strengthen $\bar q(0)_{s_0}$ to some $T_0$, so that there is a node $s_1$ on level $\bar F(1)$ such that $T_0{\Vdash}\bar q(1)_{s_1}\wedge\dot S_1\neq\emptyset$. Let $\dot T_1$ be a ${\mathbb{P}}_1$-name such that $$T_0{\Vdash}\dot T_1 =\bar q(1)_{s_1}\wedge\dot S_1.$$ Since $(\bar q,\bar F)$ is determined, it follows that $\bar q(0)_{s_0}{\Vdash}s_1\in \bar q(1)$. Thus, $\tau=\{(0, s_0), (1,s_1)\}$ lies on $(\bar q,\bar F)$ and $(T_0,\dot T_1)\leq \bar q\mid\tau$. By Proposition \[prop:strengthenDependentAssigment\], there is a determined condition $(q,F)\leq (\bar q,\bar F)$ with $\rho$ lying on $(\bar q,\bar F)$ and extending $\tau$ such that $q\mid\rho\leq (T_0,\dot T_1)$. So finally, we strengthen $f_X$ to a condition $g_X$ with $g_X({\langle}j_0,j_1{\rangle})=(q,F)$ using Proposition \[prop:strengthenTreeCondition\] (2).
Suppose inductively that the statement holds for $n$ and every maximal antichain of ${\mathbb{P}}_n$ from $M$ remains maximal in ${\mathbb{P}}_n^*$. Let’s argue that the statement holds for $n+1$. This will basically be a generalization of the argument passing from $n=1$ to $n=2$. By strengthening $f_X$, we can assume that ${\langle}j_0,\ldots,j_n{\rangle}\in X$ and that $f_X$ is determined. Let $f_X({\langle}j_0,\ldots,j_n{\rangle})=(\bar q,\bar F)$. By using our inductive assumption for $n$ and strengthening further, we can assume that $\tau:n\to {{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}2$ lies on $(\bar q{\upharpoonright}n,\bar F{\upharpoonright}n)$ such that
$q(0)_{\tau(0)}\leq S_0$, and for all $1\leq i<n$, $(q{\upharpoonright}i)\mid(\tau{\upharpoonright}i){\Vdash}q(i)_{\tau(i)}\leq \dot S_i$.
Let’s argue that $(\bar q{\upharpoonright}n)\mid \tau$ forces that for some node $s_n$ on level $\bar F(n)$ of $\bar q(n)$, $\bar q(n)_{s_n}\wedge \dot S_n\neq\emptyset$. If this is not the case, then there is a condition $p\leq (\bar q{\upharpoonright}n)\mid \tau$ which forces that there is no such node. By Proposition \[prop:strengthenDependentAssigment\], there is a determined condition $(p',F')\leq (\bar q{\upharpoonright}n,F{\upharpoonright}n)$ with $\rho$ lying on $(p',F')$ and extending $\tau$ such that $p'\mid\rho\leq p$. Let $\bar q'=p'{\cup}\{(n,\bar q(n))\}$ and $\bar F'=F'{\cup}\{(n,\bar F(n))\}$. Let $f'_X$ be the strengthening of $f_X$ in which $f'_X({\langle}j_0,\ldots,j_n{\rangle})=(\bar q',\bar F')$. Let $G$ be an $M$-generic filter for ${\mathbb{Q}}(\vec P,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$ containing $f'_X$ and consider $M[G]$. The condition $$R=({\mathcal T}_{j_0}^0\wedge S_0\wedge \bar q'(0)_{\rho(0)},\ldots,\dot T_{j_{n-1}}^{n-1}\wedge \dot S_{n-1}\wedge \bar q'(n-1)_{\rho(n-1)})$$ is in ${\mathbb{P}}^*_n$ (where $(\bar q'(0)_{\rho(0)},\ldots,\bar q'(n-1)_{\rho(n-1)})\leq p$ by construction), and so we can fix a $V$-generic filter $H^*$ containing $R$, and argue as in the case $n=2$ towards a contradiction.
Thus, there is a condition $$(T_0,\dot T_1,\ldots,\dot T_{n-1})\leq (\bar q{\upharpoonright}n)\mid \tau,$$ a ${\mathbb{P}}_n$-name $\dot T_n$ for a perfect tree, and a node $s_n$ on level $\bar F(n)$ such that $$(T_0,\dot T_1,\ldots,\dot T_{n-1}){\Vdash}\dot T_n=\bar q(n)_{s_n}\wedge \dot S_n.$$ Let $\sigma=\tau\cup \{( n,s_n)\}$. So $(T_0,\dot T_1,\ldots,\dot T_{n-1},\dot T_n)\leq \bar q\mid \sigma$, and it remains to construct the required condition using Proposition \[prop:strengthenDependentAssigment\].
We are now ready to prove that every maximal antichain of ${\mathbb{P}}_n$ from $M$ remains maximal in ${\mathbb{P}}_n^*$.
\[lem:antichainSuccessorCase\] Every maximal antichain of ${\mathbb{P}}_n$ from $M$ remains maximal in ${\mathbb{P}}_n^*$.
The statement is true for $n=1$. So we can assume inductively that ${\mathbb{U}}_n$ is dense in ${\mathbb{P}}_n^*$. Let $\mathcal A\in M$ be a maximal antichain in ${\mathbb{P}}_n$. It suffices to show that every condition in ${\mathbb{U}}_n$ is compatible with an element of $\mathcal A$. So let $f_X$ be some condition which forces that $$({\mathcal T}_{j_0}^0\wedge S_0,\dot T^1_{j_1}\wedge \dot S_1,\ldots, \dot T_{j_{n-1}}^{n-1}\wedge \dot S_{n-1})\in {\mathbb{U}}_n.$$ By Proposition \[lem:subtreeContainedinSiteration\], we can strengthen $f_X$ to a condition $g_Y$ such that $g_Y({\langle}j_0,\ldots,j_{n-1}{\rangle})=(q,F)$ is determined and $\tau:n\to {{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}2$ lies on $(q,F)$ such that $q(0)_{\tau(0)}\leq S_0$, and for all $1\leq i<n$, $(q{\upharpoonright}i)\mid (\tau{\upharpoonright}i){\Vdash}q(i)_{\tau(i)}\leq\dot S_i$. Since $\mathcal A$ is maximal in ${\mathbb{P}}_n$, $q\mid\tau$ is compatible with some $p\in \mathcal A$. So let $q'\leq q\mid\tau, p$. By Proposition \[prop:strengthenDependentAssigment\], there is a determined condition $(\bar q,\bar F)\leq (q,F)$ with $\rho$ lying on $(\bar q,\bar F)$ and extending $\tau$ such that $\bar q\mid\rho\leq q'$. Strengthen $g_Y$ to $\bar g_Y$ with $(q,F)$ strengthened to $(\bar q,\bar F)$ as in Proposition \[prop:strengthenTreeCondition\] (2). The condition $\bar g_Y$ forces that $({\mathcal T}_{j_0}^0\wedge S_0,\dot T^1_{j_1}\wedge \dot S_1,\ldots, \dot T_{j_{n-1}}^{n-1}\wedge \dot S_{n-1})$ is compatible with $p\in\mathcal A$.
Let $\vec P^*={\langle}{\mathbb{P}}_n^*\mid n<\omega{\rangle}$ be the $\omega$-iteration made up of the extended iterations ${\mathbb{P}}_n^*$. Let ${\mathbb{U}}$ be the subset of ${\mathbb{P}}(\vec P^*,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$ consisting of conditions $f_X$ such that for all $s\in X$ on level $n$, $f_X(s)\in {\mathbb{U}}_n$.
\[prop:UisDenseInP\*\] ${\mathbb{U}}$ is dense in ${\mathbb{P}}(\vec P^*,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$.
Next, we will prove the analogue of Proposition \[prop:propertiesOfProductP\*\].
\[lem:antichainTreeIteration\] Every maximal antichain of ${\mathbb{P}}(\vec P,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$ from $M$ remains maximal in ${\mathbb{P}}(\vec P^*,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$.
Fix a maximal antichain $\mathcal A\in M$ of ${\mathbb{P}}(\vec P,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$. By Proposition \[prop:UisDenseInP\*\], it suffices to show that every condition $f_X\in {\mathbb{U}}$ is compatible with some element of $\mathcal A$. So fix $f_X\in {\mathbb{U}}$. For a node $s\in X$ on level $n$ of ${{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega$, let $$f_X(s)=({\mathcal T}_{j_{0,s}}^0\wedge S_{0,s},\ldots,\dot {\mathcal T}_{j_{n-1,s}}^{n-1}\wedge \dot S_{n-1,s}).$$ Let ${\langle}j_{0,s},\ldots,j_{n-1,s}{\rangle}=\vec j_s$, and note that we can have $\vec j_s=\vec j_{s'}$ for $s'\neq s$. Fix a condition $g_Y\in G$ forcing that $f_X\in {\mathbb{U}}$. Repeatedly using the construction in the proof of Lemma \[lem:subtreeContainedinSiteration\] on terminal nodes, we can find a determined condition $\bar g_{Y}\leq g_Y$ with $\bar g_{Y}(\vec j_s)=(\bar q_{\vec j_s},\bar F_{\vec j_s})$ such that for every node $s\in X$, some $\tau_s:n\to {{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}2$ lies on $(\bar q_{\vec j_s},\bar F_{\vec j_s})$ with
$\bar q_{\vec j_s}(0)_{\tau_s(0)}\leq S_{0,s}$, and for all $1\leq i<n$, $(\bar q_{\vec j_s}{\upharpoonright}i)\mid(\tau_s{\upharpoonright}i){\Vdash}\bar q_{\vec j_s}(i)_{\tau_s(i)}\leq \dot S_{i,s}$.
The construction gives $\tau_s$ satisfying that if $s'$ extends $s$, then $\tau_{s'}$ extends $\tau_s$.
Let $\bar f_X$ be a condition in ${\mathbb{P}}(\vec P,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$ such that for every node $s\in X$, we have $\bar f_X(s)=\bar q_{\vec j_s}\mid \tau_s$. Since $\mathcal A$ is maximal in ${\mathbb{P}}(\vec P,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$, there is $a_Z\in \mathcal A$ compatible with $\bar f_X$. So let $r_W\leq a_Z,\bar f_X$.
We then carry out the construction from the proof of Proposition \[prop:strengthenDependentAssigment\], working our way up the tree instead of up the iteration, to obtain a condition $h_Y\leq \bar g_Y$ with $h_Y(\vec j_s)=(q_{\vec j_s},F_{\vec j_s})$ such that for every $s\in X$, some $\sigma$ lies on $(q_{\vec j_s},F_{\vec j_s})$ having $q_{\vec j_s}\mid \sigma\leq r_W(s)$. The condition $h_Y$ forces that $f_X$ and $a_Z$ are compatible. By density, some such $h_{Y}$ must be in $G$.
We now summarize our results in the following lemma, which will serve as the analogue of Proposition \[prop:propertiesOfP\*\].
Suppose $M$ is a suitable model and $\vec P={\langle}{\mathbb{P}}_n\mid n<\omega{\rangle}\in M$ is an $\omega$-iteration of perfect posets. If $G\subseteq {\mathbb{Q}}(\vec P,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$ is $M$-generic and $\vec P^*={\langle}{\mathbb{P}}_n^*\mid n<\omega{\rangle}$ is constructed in $M[G]$ as above, then
1. $\vec P^*$ is an $\omega$-iteration of perfect posets,
2. ${\mathbb{Q}}_0\subseteq {\mathbb{Q}}_0^*$ and for all $n<\omega$, ${\mathop{1\hskip-2.5pt {\rm l}}}_{{\mathbb{P}}_n^*}$ forces that $\dot {\mathbb{Q}}_n$ is a perfect poset and $\dot {\mathbb{Q}}_n^*$ extends it,
3. ${\mathbb{P}}_n\subseteq {\mathbb{P}}_n^*$,
4. every maximal antichain of ${\mathbb{P}}_n$ from $M$ remains maximal in ${\mathbb{P}}_n^*$,
5. every maximal antichain of ${\mathbb{P}}(\vec P,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$ from $M$ remains maximal in ${\mathbb{P}}(\vec P^*,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$.
Now we will define an $\omega$-iteration $\vec P^J={\langle}{\mathbb{P}}_n^J\mid n<\omega{\rangle}$ of perfect posets such that the tree iteration ${\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$ is going to have the property that in a forcing extension $L[G]$ of $L$ by ${\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$, the only $L$-generic filters for the iteration ${\mathbb{P}}_n^J$ are the restrictions of $G$ to a level $n$ node. This will be the generalization of Jensen’s uniqueness of generic filters property to tree iterations.
Let ${\langle}S_\alpha\mid \alpha<\omega_1{\rangle}$ be a canonically defined $\diamondsuit$-sequence. We will construct $\vec P^J$ in $\omega_1$-many steps using $\diamondsuit$ to seal maximal antichains along the way. Let $\vec P_0$ be the $\omega$-iteration of perfect posets where ${\mathbb{Q}}_0={\mathbb{P}}_{\text{min}}$ and each $\dot {\mathbb{Q}}_n=\check {\mathbb{P}}_{\text{min}}$. Suppose the $\omega$-iteration $\vec P_\alpha={\langle}{\mathbb{P}}_n^\alpha\mid n<\omega{\rangle}$ of perfect poset has been defined. We let $\vec P_\alpha=\vec P_{\alpha+1}$, unless the following happens. Suppose $S_\alpha$ codes a well-founded binary relation $E\subseteq \alpha\times\alpha$ such that the collapse of $E$ is a suitable model $M_\alpha$ with $\vec P_\alpha\in M_\alpha$ and $\alpha=\omega_1^{M_\alpha}$. In this case, we take the $L$-least $M_\alpha$-generic filter $G\subseteq {\mathbb{Q}}(\vec P_\alpha,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$ and let $\vec P_{\alpha+1}=\vec P_\alpha^*$ as constructed in $M_\alpha[G]$. At limit stages $\lambda$, to obtain the $\omega$-iteration $\vec P_\lambda$, we let ${\mathbb{Q}}^\lambda_0$ be the union of the ${\mathbb{Q}}^\xi_0$ for $\xi<\lambda$, and given that we have defined ${\mathbb{P}}_n^\lambda$, we let $\dot {\mathbb{Q}}_n^\lambda$ be a ${\mathbb{P}}^\lambda_n$-name for the poset that is the union of the $\dot {\mathbb{Q}}_n^\xi$ for $\xi<\lambda$. In order for this limit definition to make sense, we need to verify that each $\dot{\mathbb{Q}}^\xi_n$ is a ${\mathbb{P}}^\lambda_n$-name for a perfect poset. So let’s argue that this is indeed the case.
Clearly ${\mathbb{P}}_1^\lambda$ makes sense, ${\mathbb{P}}_1^\xi\subseteq {\mathbb{P}}_1^\lambda$ for every $\xi<\lambda$, and every maximal antichain of ${\mathbb{P}}_1^\xi$ from $M_\xi$ remains maximal in ${\mathbb{P}}_1^\lambda$. So we can assume inductively that we have defined ${\mathbb{P}}_n^\lambda$ so that ${\mathbb{P}}_n^\xi\subseteq{\mathbb{P}}_n^\lambda$ for every $\xi<\lambda$, and every maximal antichain of ${\mathbb{P}}_n^\xi$ in $M_\xi$ remains maximal in ${\mathbb{P}}_n^\lambda$. Let $H^*\subseteq {\mathbb{P}}_n^\lambda$ be $V$-generic. By our assumption, the filter $H^*$ restricts to an $M_\xi$-generic filter $H$ for ${\mathbb{P}}^\xi_n$. The model $M_\xi[H]$ satisfies that $(\dot{\mathbb{Q}}^\xi_n)_H$ is a perfect poset and so this must be the case in $V[H^*]$ as well since this is absolute. So we can extend the definition to ${\mathbb{P}}_{n+1}^\lambda$. It is easy to see that ${\mathbb{P}}_{n+1}^\xi\subseteq {\mathbb{P}}_{n+1}^\lambda$. It remains to argue that every maximal antichain $\mathcal A\in M_\xi$ of ${\mathbb{P}}_{n+1}^\xi$ remains maximal in ${\mathbb{P}}_{n+1}^\lambda$. Fix $p\in {\mathbb{P}}_{n+1}^\lambda$. We can assume inductively that we have already showed for every $\mu<\nu<\lambda$, that every maximal antichain of ${\mathbb{P}}_{n+1}^{\mu}$ in $M_\mu$ remains maximal in ${\mathbb{P}}_{n+1}^{\nu}$. Then, by definition of ${\mathbb{P}}^\lambda_{n+1}$, there is some $\xi<\eta<\lambda$ such that $p(0)\in {\mathbb{Q}}_0^{\eta}$, and $p{\upharpoonright}i{\Vdash}p(i)\in \dot {\mathbb{Q}}^{\eta}_i$ for $1\leq i<n$. Since $\mathcal A$ is maximal in ${\mathbb{P}}^\eta_{n+1}$, it follows that there is some $q\in\mathcal A$ that is compatible with $p\in {\mathbb{P}}^\eta_{n+1}\subseteq {\mathbb{P}}^\lambda_{n+1}$. In particular, we have just shown the following.
\[lem:antichainLimitCase\] If $\lambda$ is a limit ordinal and $\xi<\lambda$, then ${\mathbb{P}}_n^\xi\subseteq {\mathbb{P}}_n^\lambda$ and every maximal antichain of ${\mathbb{P}}_n^\xi$ from $M_\xi$ remains maximal in ${\mathbb{P}}_n^\lambda$.
\[lem:antichainTreeIteration\] If $\alpha<\beta<\omega_1$, then every maximal antichain of ${\mathbb{P}}(\vec P_{\alpha},{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$ from $M_\alpha$ remains maximal in ${\mathbb{P}}(\vec P_{\beta},{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$.
The proof uses Lemma \[lem:antichainSuccessorCase\] for successor stages and follows easily for limit stages.
For $n<\omega$, let ${\mathbb{P}}_n^J={\mathbb{P}}_n^{\omega_1}$ be constructed as all other limit stages and let $\vec {\mathbb{P}}^J={\langle}{\mathbb{P}}_n^J\mid n<\omega{\rangle}$ be the corresponding $\omega$-iteration.
It will be useful for future arguments to assume that conditions in ${\mathbb{P}}_n^J$ are coded by reals.
\[th:countableTreeIterationCCC\] The poset ${\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$ has the ccc.
Fix a maximal antichain $\mathcal A$ of ${\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$. Choose some transitive $M\prec L_{\omega_2}$ of size $\omega_1$ with $\mathcal A\in M$. We can decompose $M$ as the union of an elementary chain of countable substructures $$X_0\prec X_1\prec\cdots\prec X_\alpha\prec\cdots\prec M$$ with $\mathcal A\in X_0$. By the properties of $\diamondsuit$, there is some $\alpha$ such that $\alpha=\omega_1\cap {X_\alpha}$, ${\mathbb{P}}_n^\alpha={\mathbb{P}}_n^J\cap X_\alpha$ for all $n<\omega$, and $S_\alpha$ codes $X_\alpha$. Let $M_\alpha$ be the transitive collapse of $X_\alpha$. Then ${\mathbb{P}}_n^\alpha$ is the image of ${\mathbb{P}}_n^J$ under the collapse and $\alpha$ is the image of $\omega_1$. Let $\bar {\mathcal A}=\mathcal A\cap X_\alpha$ be the image of $\mathcal A$. So at stage $\alpha$ in the construction of $\vec P^J$, we chose a forcing extension $M_\alpha[G]$ of $M_\alpha$ by ${\mathbb{Q}}(\vec P_{\alpha},{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$ and let ${\mathbb{P}}^{\alpha+1}_n=({\mathbb{P}}_n^\alpha)^*$ be constructed in $M_\alpha[G]$. By Lemma \[lem:antichainTreeIteration\], $\bar {\mathcal A}$ remains maximal in all further ${\mathbb{P}}(\vec P_\beta,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$ for $\beta>\alpha$, and hence $\bar{\mathcal A}$ remains maximal in ${\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$. It follows that $\bar {\mathcal A}=\mathcal A$ is countable.
The generalized Kanovei-Lyubetsky Theorem {#sec:KanoveiLyubetskyGeneralized}
=========================================
Now that we have constructed the $\omega$-iteration $\vec P^J={\langle}{\mathbb{P}}_n^J\mid n<\omega{\rangle}$ and the corresponding tree iteration ${\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$ of Jensen’s forcing, we would like to generalize the Kanovei-Lyubetsky argument of Section \[sec:KanoveiLyubetsky\] to prove the analogue of the “uniqueness of generics" property for tree iterations. Namely, we will show that in a forcing extension $L[G]$ by ${\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}})$ the only $L$-generic filters for ${\mathbb{P}}_n^J$ are the restrictions of $G$ to a level $n$ node.
Suppose that $\vec P={\langle}{\mathbb{P}}_n\mid n<\omega{\rangle}$ is an $\omega$-iteration of perfect posets and $H$ is a generic filter for ${\mathbb{P}}(\vec P,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$. Given a node $s$ on level $n$ of ${{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega$, let $x_s$ be the $n$-length sequence of generic reals added by $H$ on node $s$ and let $\dot x_s$ be the canonical name for $x_s$.
For the next lemma, suppose that $\vec P={\langle}{\mathbb{P}}_n\mid n<\omega{\rangle}$ is an $\omega$-iteration of perfect posets that is an element of a suitable model $M$. We should think of $\vec P$ as one of the $\omega$-iterations $\vec P_\alpha$ arising at stage $\alpha$ in the construction of the $\omega$-iteration $\vec P^J$ and we should think of $M$ as the model $M_\alpha$ from that stage.
\[le:denseToAvoidBranchesInTreeIteration\] In $M$, suppose that $\dot r$ is a ${\mathbb{P}}(\vec P,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$-name for an $n$-length sequence of reals such that for all nodes $s$ on level $n$ of ${{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega$, $${\mathop{1\hskip-2.5pt {\rm l}}}{\Vdash}_{{\mathbb{P}}(\vec P,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)} \dot r\neq \dot x_s.$$ Then in a forcing extension $M[G]$ by ${\mathbb{Q}}(\vec P,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$, for every node $s$ on level $n$ of ${{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega$, the set of conditions forcing the statement
$\Phi(s):=$“If $\dot r$ is $M[G]$-generic for ${\mathbb{P}}^*_n$, then ${\langle}{\mathcal T}^0_{s(0)},\dot {\mathcal T}^1_{s(1)},\ldots,\dot {\mathcal T}^{n-1}_{s(n-1)}{\rangle}$ is not in the filter determined by $\dot r$."
is dense in ${\mathbb{P}}(\vec P^*,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$.
Fix a condition $f_X\in {\mathbb{P}}(\vec P^*,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$ and a node $d$ on level $n$ of ${{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega$. We need to find a condition $f'_{X'}\leq f_X$ such that $f'_{X'}{\Vdash}\Phi(d)$. Since ${\mathbb{U}}$ is dense in ${\mathbb{P}}(\vec P^*,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$, we can assume without loss that $f_X\in{\mathbb{U}}$. For a node $s$ on level $m$ of $X$, let $$f_X(s)=({\mathcal T}_{j_{0,s}}^0\wedge S_{0,s},\ldots,\dot {\mathcal T}_{j_{m-1,s}}^{m-1}\wedge \dot S_{m-1,s}).$$ Let ${\langle}j_{0,s},\ldots,j_{m-1,s}{\rangle}=\vec j_s$, and note that we can have $\vec j_s=\vec j_{s'}$. By strengthening if necessary, we can assume that there is a node $s\in X$ with $\vec j_s=d$. Repeatedly using the construction in the proof of Lemma \[lem:subtreeContainedinSiteration\] on terminal nodes, we can find a determined condition $g_Y\in G$ with $g_Y(\vec j_s)=(q_{\vec j_s},F_{\vec j_s})$ such that for every node $s\in X$, some $\tau_s:m\to {{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}2$ lies on $(q_{\vec j_s},F_{\vec j_s})$ with
$q_{\vec j_s}(0)_{\tau_s(0)}\leq S_{0,s}$, and for all $1\leq i<m$, $(q_{\vec j_s}{\upharpoonright}i)\mid(\tau_s{\upharpoonright}i){\Vdash}q_{\vec j_s}(i)_{\tau_s(i)}\leq \dot S_{i,s}$.
The construction gives $\tau_s$ satisfying that if $s'$ extends $s$, then $\tau_{s'}$ extends $\tau_s$.
Now we are going to construct a condition $a_A^g\in {\mathbb{P}}(\vec P,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$ with $a_A^g(s)=q_s$ associated to $g_Y$, satisfying the following properties:
1. $X\subseteq A$ and for every node $s\in X$, $q_s\leq q_{\vec j_s}\mid \tau_s$.\
There is $X\subseteq \bar A\subseteq A$ such that:
2. For every $\sigma$ which lies on $(q_d,F_d)$, there is a node $s_\sigma\in \bar A$ such that $q_{s_\sigma}\leq q_d\mid\sigma$.
3. For every node $s$ on level $n$ of $\bar A$, $a_A^g$ forces over ${\mathbb{P}}(\vec P,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$ the statement:
“There is $i<n-1$ such that $\dot r{\upharpoonright}i$ is (ground model) generic for ${\mathbb{P}}_i$ and $\dot r(i)\notin[(a_A^g(s)(i))_{\dot r{\upharpoonright}i}]$."
Let $a_{\bar A}$ be a condition in ${\mathbb{P}}(\vec P,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$ with $X\subseteq \bar A$ be such that $a_{\bar A}(s)=q_{\vec j_s}\mid \tau_s$ for every node $s\in X$, and for every $\sigma$ which lies on $(q_d,F_d)$, there is a node $s_\sigma\in\bar A$ such that $a_{\bar A}(s_\sigma)=q_d\mid\sigma$. Fix a node $s$ on level $n$ of $\bar A$ and consider a forcing extension $M[H]$ by ${\mathbb{P}}(\vec P,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$. By assumption, we have $r=\dot r_H\neq x_s$. So there is $i<n-1$ such that $r{\upharpoonright}i=x_s{\upharpoonright}i$ and $r(i)\neq x_s(i)$. So we can strengthen $a_{\bar A}$ to a condition $a_{A'}$ forcing that there is $i<n-1$ such that $\dot r{\upharpoonright}i$ is $M$-generic and $\dot r(i)\notin[(a_{A'}(s)(i))_{\dot r(i)}]$. By repeating this for all the finitely many nodes on level $n$, we obtain the required condition $a_A^g$.
Next, we can carry out the construction in the proof of Proposition \[prop:strengthenDependentAssigment\], moving up the tree instead of up the iteration, to obtain a condition $\bar g_Y\leq g_Y$ with $\bar g_Y(\vec j_s)=(\bar q_{\vec j_s},\bar F_{\vec j_s})$ such that for every $s\in X$, some $\sigma$ lies on $(\bar q_{\vec j_s},\bar F_{\vec j_s})$ having $\bar q_{\vec j_s}\mid\sigma\leq a_A^g(s)$. We can also ensure that for every $\tau$ which lies on $(\bar q_d,\bar F_d)$, if $\tau$ extends $\sigma$ which lies on $(q_d,F_d)$, then $\bar q_d\mid \tau\leq a^g_A(s_\sigma)$. By density, some such $\bar g_{Y}$ must be in $G$.
By construction, $\bar g_{Y}$ forces that $f_X$ is compatible with $a_A^g$. So let $\bar f_{\bar X}\leq f_X,a_A^g$. We will be done if we can show that $a^g_A$ forces the statement $\Phi(d)$ over ${\mathbb{P}}(\vec P^*,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$.
Suppose $H^*\subseteq {\mathbb{P}}(\vec P^*,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$ is a $V$-generic filter containing $a^g_A$ ($V=M[G]$). Now let’s suppose towards a contradiction that $p={\langle}{\mathcal T}^0_{d(0)},\dot {\mathcal T}^1_{d(1)},\ldots, \dot {\mathcal T}^{n-1}_{d(n-1)}{\rangle}$ is in the filter determined by $r$. Thus, there is some $\sigma$ which lies on $p$ such for for all $i<n$, $r(i)$ is a branch through $(p\mid\sigma(i))_{r{\upharpoonright}i}$. By construction, we have that $p\mid\sigma\leq a^g_A(s)$ for some $s$ on level $n$ of ${{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega$. Let $H$ be the restriction of $H^*$ to an $M$-generic filter for ${\mathbb{P}}(\vec P,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$, and note that $a^g_A\in H$. Thus, there is some $i<n-1$ such that $r(i)$ is not a branch through $(a^g_A(s)(i))_{r{\upharpoonright}i}$, which is the desired contradiction.
\[th:uniquenessOfGenericsTreeIterationCountable\] Suppose $H\subseteq {\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$ is $L$-generic. If an $n$-length sequence of reals $r\in L[H]$ is $L$-generic for ${\mathbb{P}}_n$, then $r=x_s$ for some node $s$ on level $n$ of ${{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega$.
Let’s suppose that $r$ is not one of the $x_s$ for $s$ on level $n$ of ${{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega$. Let $\dot r$ be a nice ${\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$-name for $r$ such that for all nodes $s$ on level $n$ of ${{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega$, ${\mathop{1\hskip-2.5pt {\rm l}}}{\Vdash}_{{\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)} \dot r\neq \dot x_s$.
Choose some transitive $M\prec L_{\omega_2}$ of size $\omega_1$ with $\dot r\in M$. We can decompose $M$ as the union of a continuous elementary chain of countable substructures $$X_0\prec X_1\prec\cdots\prec X_\alpha\prec\cdots\prec M$$ with $\dot r\in X_0$. By the properties of $\diamondsuit$, there is some $\alpha$ such that $\alpha=\omega_1\cap{X_\alpha}$, ${\mathbb{P}}_n^\alpha={\mathbb{P}}_n^J\cap X_\alpha$ for all $n<\omega$, and $S_\alpha$ codes $X_\alpha$. Let $M_\alpha$ be the transitive collapse of $X_\alpha$. Then, for every $n<\omega$, ${\mathbb{P}}_n^\alpha$ is the image of ${\mathbb{P}}_n^J$ under the collapse, and $\alpha$ is the image of $\omega_1$. The name $\dot r$ is fixed by the collapse by our assumption that we can always code conditions in ${\mathbb{P}}(\vec {\mathbb{P}}^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$ by subsets of $\omega$ and because all antichains of ${\mathbb{P}}(\vec {\mathbb{P}}^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$ are countable. So at stage $\alpha$ in the construction of $\vec P^J$, we chose a forcing extension $M_\alpha[G]$ of $M_\alpha$ by ${\mathbb{Q}}(\vec P^{\alpha},{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$ and let $\vec P_{\alpha+1}=\vec P_\alpha^*$ be constructed in $M_\alpha[G]$.
By elementarity, $M_\alpha$ satisfies that ${\mathop{1\hskip-2.5pt {\rm l}}}_{{\mathbb{P}}(\vec P_\alpha,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)}{\Vdash}\dot r\neq \dot x_s$ for all $s$ on level $n$ of ${{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega$. Thus, by Lemma \[le:denseToAvoidBranchesInTreeIteration\], for every $s$ on level $n$ of ${{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega$, ${\mathbb{P}}(\vec {\mathbb{P}}_{\alpha+1},{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$ has a maximal antichain $\mathcal A_s$ consisting of conditions $f_X$ forcing the statement:
$\Phi(s):=$“If $\dot r$ is $M_\alpha[G]$-generic for ${\mathbb{P}}^*_n$, then ${\langle}{\mathcal T}^0_{s(0)},\dot {\mathcal T}^1_{s(1)},\ldots,\dot {\mathcal T}^{n-1}_{s(n-1)}{\rangle}$ is not in the filter determined by $\dot r$."
It follows, by an argument analogous to the proof of Proposition \[prop:MaxAntiMForcingExtension\], that every $\mathcal A_s$ remains maximal in ${\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$. So let’s argue that if $f_X\in \mathcal A_s$, then $f_X$ forces in ${\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$ that if $\dot r$ is $L$-generic, then ${\langle}{\mathcal T}^0_{s(0)},\dot {\mathcal T}^1_{s(1)},\ldots,\dot {\mathcal T}^{n-1}_{s(n-1)}{\rangle}$ is not in the filter determined by $\dot r$.
Let $\bar H^*\subseteq {\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$ be an $L$-generic filter containing $f_X$ and let $\bar H$ be the restriction of $\bar H^*$ to ${\mathbb{P}}(\vec P_{\alpha+1},{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$. Let $r=\dot r_{\bar H^*}$ and suppose that it is $L$-generic for ${\mathbb{P}}_n^*$. Since $f_X\in \bar H$, it follows that $M_{\alpha}[G][\bar H]$ satisfies that ${\langle}{\mathcal T}^0_{s(0)},\dot {\mathcal T}^1_{s(1)},\ldots,\dot {\mathcal T}^{n-1}_{s(n-1)}{\rangle}$ is not in the filter determined by $r$. But then this is true in $L$ as well.
Since $H$ must meet every $\mathcal A_s$, it holds in $L[H]$ that if $r$ is $L$-generic for ${\mathbb{P}}_n^J$, then it does not meet the maximal antichain consisting of conditions ${\langle}{\mathcal T}^0_{s(0)},\dot {\mathcal T}^1_{s(1)},\ldots,\dot {\mathcal T}^{n-1}_{s(n-1)}{\rangle}$ for nodes $s$ on level $n$ of ${{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega$ (from level $\alpha$ of the construction), and so $r$, in fact, cannot be $L$-generic.
Iterating $\vec P^J$ along the tree ${{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1$
====================================================================================================================================================================================================================
We will now argue that the tree iteration ${\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1)$, where we iterate along the uncountable tree ${{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1$, shares all the key properties of the tree iteration ${\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$, namely it has the ccc and Theorem \[th:uniquenessOfGenericsTreeIterationCountable\], concerning the uniqueness of generic filters for ${\mathbb{P}}_n^J$, continues to holds.
The poset ${\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1)$ has the ccc.
Let’s suppose to the contrary that there is an uncountable antichain $\mathcal A$ in ${\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1)$. By a $\Delta$-system argument, there must be some finite subtree $X\subseteq {{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1$ and an uncountable subset $\mathcal A'\subseteq \mathcal A$ such that for any $f_Y$ and $g_Z$ in $\mathcal A'$, $Y\cap Z=X$. Given $f_Y\in \mathcal A'$, let $f_X$ be the restriction of $f_Y$ to $X$. Clearly the collection of all such $f_X$ is an uncountable antichain in ${\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1)$. But then since $X$ is finite, there must be a corresponding uncountable antichain in ${\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$, which is impossible by Theorem \[th:countableTreeIterationCCC\].
\[th:uniquenessOfGenericsTreeIterationUncountable\] Suppose $H\subseteq {\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1)$ is $L$-generic. If an $n$-length sequence of reals $r\in L[H]$ is $L$-generic for ${\mathbb{P}}_n^J$, then $r=x_s$ for some node $s$ on level $n$ of ${{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1$.
Suppose that $r$ is an $n$-length sequence of reals in $L[H]$ that is $L$-generic for ${\mathbb{P}}_n^J$. Let $\dot r$ be a nice ${\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1)$-name for $r$. Since ${\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1)$ has the ccc by Theorem \[th:uniquenessOfGenericsTreeIterationUncountable\], it follows that conditions in the name $\dot r$ use only countably many nodes of ${{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1$. Thus, $\dot r$ is a ${\mathbb{P}}(\vec P,\mathscr T)$-name, where $\mathscr T$ is a countable subtree of ${{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1$. We may assume that $\mathscr T$ is isomorphic to ${{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega$, and so ${\mathbb{P}}(\vec P,\mathscr T)$ is isomorphic to ${\mathbb{P}}(\vec P,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$. Let $\bar H$ be the restriction of $H$ to ${\mathbb{P}}(\vec P,\mathscr T)$, which is $L$-generic for it. Thus, $r\in L[\bar H]$, from which it follows, by Theorem \[th:uniquenessOfGenericsTreeIterationCountable\], that $r=x_s$ for some $s\in\mathscr T$.
Symmetric models of ${{\rm ZF}}+{{\rm AC}}_\omega +\neg {{\rm DC}}$ {#sec:symmetricModels}
===================================================================
We will construct a symmetric submodel of a forcing extension $L[G]$ by ${\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1)$ in which ${{\rm ZF}}+{{\rm AC}}_\omega$ holds, but ${{\rm DC}}$ fails. The reals of this model will be a model of second-order arithmetic in which ${{\rm Z}}_2+\Sigma^1_\infty$-${{\rm AC}}$ holds, but $\Pi^1_2$-${{\rm DC}}$ fails. Let’s start with a brief discussion of the method of constructing symmetric submodels of a forcing extension, which goes all the way back to Cohen’s pioneering work on forcing.
Suppose that ${\mathbb{P}}$ is a forcing notion. Recall that if $\pi$ is an automorphism of ${\mathbb{P}}$, then we can apply $\pi$ recursively to conditions in a ${\mathbb{P}}$-name $\sigma$ to obtain the ${\mathbb{P}}$-name $\pi(\sigma)$. It is not difficult to see by induction on complexity of formulas, that for every formula $\varphi$ and condition $p\in {\mathbb{P}}$, $p{\Vdash}\varphi(\sigma)$ if and only if $\pi(p){\Vdash}\varphi(\pi(\sigma))$. Fix some group $\mathcal G$ of automorphisms of ${\mathbb{P}}$. Recall that a filter $\mathscr F$ on subgroups of a group $\mathcal G$ is *normal* if whenever $g\in \mathcal G$ and $\mathcal K\in \mathscr F$, then $g\mathcal K g^{-1}\in \mathscr F$. Let’s fix a normal filter $\mathscr F$ on the subgroups of $\mathcal G$. The subgroup of $\mathcal G$ fixing a particular ${\mathbb{P}}$-name $\sigma$, consisting of automorphisms $\pi$ such that $\pi(\sigma)=\sigma$, is called ${\text{sym}}(\sigma)$. If ${\text{sym}}(\sigma)$ is in $\mathscr F$, then we say that $\sigma$ is a *symmetric* ${\mathbb{P}}$-name. We recursively define that a ${\mathbb{P}}$-name is *hereditarily symmetric* when it is symmetric and all names inside it are hereditarily symmetric. Let ${{\rm HS}}$ be the collection of all hereditarily symmetric ${\mathbb{P}}$-names. Let $G\subseteq {\mathbb{P}}$ be $V$-generic. The *symmetric model* $$N=\{\sigma_G\mid \sigma\in {{\rm HS}}\}$$ associated to the group of autmorphisms $\mathcal G$ and the normal filter $\mathscr F$ consists of the interpretations of all hereditarily symmetric ${\mathbb{P}}$-names. It is a standard result that $N\models{{\rm ZF}}$.
We would like to review now a classical construction of a symmetric model in which countable choice holds but dependent choice fails, which is due to Jensen [@jensen:ACplusNotDC]. Let ${\mathbb{P}}$ be the forcing adding a set of Cohen subsets of $\omega_1$, indexed by the tree ${{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1$, with countable conditions. We will call the poset adding a Cohen subset to $\omega_1$ with countable conditions ${{\rm Add}}(\omega_1,1)$. So, more precisely, conditions in ${\mathbb{P}}$ are functions $f_X:X\to {{\rm Add}}(\omega_1,1)$, where $X$ is some countable subtree of ${{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1$, ordered so that $g_Y\leq f_X$ whenever $X\subseteq Y$ and for all $t\in X$, $g_Y(t)\leq f_X(t)$ in ${{\rm Add}}(\omega_1,1)$. Note that ${\mathbb{P}}$ is countably closed. Next, we need to select an appropriate group of automorphisms of ${\mathbb{P}}$.
Let ${{\rm Aut}}({{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1)$ be the automorphism group of the tree ${{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1$. Every automorphism $\pi\in {{\rm Aut}}({{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1)$ induces an automorphism $\pi^*$ of ${\mathbb{P}}$, so that $\pi^*(f_X)=f_{\pi{\mathbin{\hbox{\tt\char'42}}}X}$ with $f_{\pi{\mathbin{\hbox{\tt\char'42}}}X}(t)=f_X(\pi^{-1}(t))$. Let $$\mathcal G=\{\pi^*\mid \pi\in{{\rm Aut}}({{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1)\}.$$ Next, we need to select an appropriate filter on the subroups of $\mathcal G$. Let’s call a subtree $T$ of the tree ${{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1$ *useful* if it is countable and doesn’t have an infinite branch. Given a useful tree $T$, let $H_T$ be the subgroup of $\mathcal G$ consisting of all automorphisms $\pi^*$ such that $\pi$ point-wise fixes $T$. Let $\mathscr F$ be the filter on the subgroups of $\mathcal G$ generated by all such subgroups $H_T$ with $T$ a useful tree. To see that $\mathscr F$ is normal, observe that if $T$ is a useful tree and $H_T\subseteq K\in\mathscr F$, then $\pi{\mathbin{\hbox{\tt\char'42}}}T$ is useful and $H_{\pi{\mathbin{\hbox{\tt\char'42}}}T}\subseteq \pi^*K\pi^{*-1}$.
Now, let $G\subseteq{\mathbb{P}}$ be $V$-generic and let $N=\{\sigma_G\mid \sigma\in {{\rm HS}}\}$ be the symmetric model associated to $\mathcal G$ and $\mathscr F$. In $V[G]$, let ${\mathcal T}$ be the tree isomorphic to ${{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1$ whose nodes are the Cohen subsets of $\omega_1$ added by $G$. Given a node $t\in{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1$, let $\sigma_t$ be the canonical name for the Cohen subset of $\omega_1$ added on node $t$ by $G$. Let $\dot {\mathcal T}=\{( {\rm op}(\sigma_s,\sigma_t),{\mathop{1\hskip-2.5pt {\rm l}}}_{{\mathbb{P}}})\mid s\leq t\text{ in }{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1\}$, where ${\rm op}(\sigma_s,\sigma_t)$ is the canonical ${\mathbb{P}}$-name for the ordered pair of $\sigma_s$ and $\sigma_t$. Clearly $\dot {\mathcal T}_G={\mathcal T}$. Fix any $\pi^*\in \mathcal G$, and observe that $\pi^*(\dot {\mathcal T})=\{({\rm op}(\sigma_{\pi(s)},\sigma_{\pi(t)}),{\mathop{1\hskip-2.5pt {\rm l}}}_{{\mathbb{P}}})\mid s\leq t\text{ in }{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1\}=\dot {\mathcal T}$. Also, any automorphism $\pi^*$ with $\pi(s)=s$ fixes $\sigma_s$. This shows that $\dot {\mathcal T}\in {{\rm HS}}$, and hence ${\mathcal T}$ is in the symmetric model $N$. Using that a symmetric name must be fixed by a group of automorphisms point-wise fixing a tree without an infinite branch, we can show that no infinite branch through $\mathcal T$ can have a symmetric name, and so ${{\rm DC}}$ fails in $N$. Using that ${\mathbb{P}}$ is countably closed, we can show that ${{\rm AC}}_\omega$ holds in $N$.
We are now ready to construct a symmetric model $N$ of a forcing extension $L[G]$ by ${\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1)$ with the same properties. First, observe that every $\pi\in{{\rm Aut}}({{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1)$ induces an automorphism $\pi^*$ of ${\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1)$ so that $\pi^*(f_X)=f_{\pi{\mathbin{\hbox{\tt\char'42}}}X}$ with $f_{\pi{\mathbin{\hbox{\tt\char'42}}}X}(t)=f_X(\pi^{-1}(t))$. As above, let $\mathcal G$ be the group of automorphisms $\pi^*$ for $\pi\in {{\rm Aut}}({{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1)$ and let $\mathscr F$ be the normal filter generated by the subgroups $H_T$ of $\mathcal G$, consisting of all automorphisms point-wise fixing some useful subtree $T$ of $\omega_1{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}$. Note that in the present situation the domains of conditions $f_X$ are finite subtrees $X$, while we use countable trees $T$ to determine which names are symmetric. Let $N\subseteq L[G]$ be the symmetric model determined by the group of automorphisms $\mathcal G$ and the filter $\mathscr F$. In $L[G]$, consider the tree $\mathcal T$, isomorphic to ${{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1$ whose nodes are the generic sequences of reals added by $G$. The argument we gave above generalizes in a straightforward way to show that $\dot{\mathcal T}$, the canonical name for ${\mathcal T}$, is hereditarily symmetric, and hence ${\mathcal T}\in N$.
Suppose $T$ is any subtree of $\omega_1{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}$. If $f_X$ is a condition in ${\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1)$, we will denote by $f_{X\cap T}$, the restriction of $f_X$ to nodes in $T$, i.e. $f_{X\cap T}$ has domain $X\cap T$ and $f_{X\cap T}(t)=f_X(t)$. We let $$G_T=\{f_X\in G\mid X\subseteq T\}.$$ Note that if $T$ is useful, then the canonical name for $G_T$ is fixed by all elements of the subgroup $H_T$, and therefore $G_T\in N$. Thus, each $L[G_T]\subseteq N$.
If $\sigma$ is a ${\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1)$-name and $H_T\subseteq {\text{sym}}(\sigma)$ for a useful tree $T$, we will say that $T$ *witnesses* that $\sigma$ is symmetric.
\[prop:countableTreeDecides\] Suppose that $\sigma$ is a symmetric ${\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1)$-name, as witnessed by a useful tree $T$, and $f_X{\Vdash}\varphi(\sigma,\check a)$. Then $f_{X\cap T}{\Vdash}\varphi(\sigma,\check a)$.
Suppose it is not the case that $f_{X\cap T}{\Vdash}\varphi(\sigma,\check a)$. Then there is a condition $g_Y\leq f_{X\cap T}$ which forces $\neg\varphi(\sigma,\check a)$. So $g_Y$ is incompatible with $f_X$ and the incompatibility must occur on nodes outside of $T$. Let $\pi\in{{\rm Aut}}({{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1)$ be an automorphism which switches these nodes in $Y\setminus T$ with some nodes outside both $T$ and $X$. Clearly $\pi^*\in H_T$, and so $\pi^*(\sigma)=\sigma$. It follows that $\pi^*(g_Y){\Vdash}\neg\varphi(\sigma,\check a)$. But, by construction, $\pi^*(g_Y)$ is compatible with $f_X$, which is the desired contradiction. Thus, we have shown that $f_{X\cap T}{\Vdash}\varphi(\sigma,\check a)$.
\[prop:subsetsOfOrdinalsSymmetricModels\] Suppose that $\sigma$ is a symmetric ${\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1)$-name, as witnessed by a useful tree $T$, and $A=\sigma_G$ is a set of ordinals. Then $A\in L[G_T]$.
Let $f_X$ be some condition in $G$ forcing that $\sigma$ is a set of ordinals. Define a name $$\sigma^*=\{(\xi,g_{Y\cap T})\mid g_Y\leq f_X, g_Y{\Vdash}\xi\in\sigma\}.$$ We will argue that $f_X{\Vdash}\sigma=\sigma^*$. Let $H\subseteq {\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1)$ be some $L$-generic filter containing $f_X$. Suppose $\xi\in\sigma_H$. Then there is $g_Y\in H$ such that $g_Y\leq f_X$ and $g_Y{\Vdash}\xi\in \sigma$, from which it follows that $(\xi,g_{Y\cap T})\in \sigma^*$ and $g_{Y\cap T}\in H$. So we have $\xi\in \sigma^*_H$. Next, suppose that $\xi\in\sigma^*_H$. Then there is a condition $g_Y{\Vdash}\xi\in \sigma$ such that $(\xi,g_{Y\cap T})\in \sigma^*$ and $g_{Y\cap T}\in H$. But by Proposition \[prop:countableTreeDecides\], it follows that $g_{Y\cap T}{\Vdash}\xi\in\sigma$, and so $\xi\in\sigma_H$.
${{\rm DC}}$ fails in $N$.
We will argue that $\mathcal T$ does not have a infinite branch in $N$, and hence ${{\rm DC}}$ fails. Suppose to the contrary that $N$ has an infinite branch $b$ through $\mathcal T$, which we can view via coding as a subset of natural numbers (since it is an $\omega$-length sequence of reals). Fix a symmetric name $\dot b$ for $b$, as witnessed by a useful tree $T$. By Proposition \[prop:subsetsOfOrdinalsSymmetricModels\], we can assume that the name $\dot b$ mentions only conditions with domains contained in $T$. Recall that for a node $s\in {{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1$, $x_s$ is the $L$-generic sequence of reals for ${\mathbb{P}}_{\text{len}(s)}^J$ added by $G$ on node $s$ and $\dot x_s$ is the canonical ${\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1)$-name for $x_s$. Let $\bar b$ be the branch through the tree ${{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1$ which corresponds to $b$ via the obvious isomoprhism. Since $T$ doesn’t have infinite branches by assumption, $\bar b$ cannot be a branch through $T$. Thus, there is some natural number $n$ such that $\bar b{\upharpoonright}n\subseteq T$ and $\bar b(n)=s$ is outside $T$.
Fix a condition $$f_X{\Vdash}\dot b(n)=\dot x_s,$$ and assume without loss that $s\in X$. It is easy to see that conditions $g_Y$ having some $t\in Y\setminus (X{\cup}T)$ with $g_Y(t)=f_X(s)$ are dense below $f_X$. So fix a condition $g_Y\in G$ with $X\subseteq Y$ and $t\in Y\setminus (X{\cup}T)$ with $g_Y(t)=f_X(s)$. Let $\pi$ be the automorphism of the tree ${{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1$ which swaps $s$ with $t$. Let $H=\pi{\mathbin{\hbox{\tt\char'42}}}G$, which is also $L$-generic for ${\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1)$. Observe that $\dot b_H=b$ since the name $\dot b$ only mentioned conditions with domain in $T$. Observe also that $f_X\in H$ since it is above $\pi(g_Y)=g_{\pi{\mathbin{\hbox{\tt\char'42}}}Y}$. So it must be the case that $b(n)=(\dot x_s)_H$. But this is impossible because $(\dot x_s)_H=x_t$ and $(\dot x_s)_G=x_s$ and, by genericity, $x_s\neq x_t$.
The proof of the next lemma relies mainly on the fact that ${\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1)$ has the ccc and uses very little else about what the conditions in ${\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1)$ look like. So to make the notation nicer, we will switch away from the previous convention and call conditions in ${\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1)$ standard names like $p$ and $q$.
${{\rm AC}}_\omega$ holds in $N$.
Suppose $F\in N$ is a countable family of non-empty sets. Let $\dot F$ be a hereditarily symmetric name for $F$ with the useful tree $S$ witnessing that $\dot F$ is symmetric, and let $q\in G$ force that $\dot F$ is a countable family of non-empty sets. We would like to build a name $\dot C\in {{\rm HS}}$ such that $q$ forces that $\dot C$ is a choice function for $\dot F$. We will adapt the following strategy. First, we will build a mixed name $\dot C_0\in {{\rm HS}}$ (over an antichain below $q$) and a useful tree $T_0$ extending $S$, witnessing that $\dot C_0$ is symmetric, such that $q{\Vdash}\dot C_0\in \dot F(0)$. Next, we will build a mixed name $\dot C_1\in{{\rm HS}}$ and a useful tree $T_1$ extending $T_0$, witnessing that $\dot C_1$ is symmetric, such that $q{\Vdash}\dot C_1\in\dot F(1)$. Proceeding in this fashion, we will build names $\dot C_n\in{{\rm HS}}$ and an increasing sequence of useful trees $T_n$ such that $q{\Vdash}\dot C_n\in \dot F(n)$. Provided we can ensure in the course of the construction that $T={\bigcup}_{n\in\omega}T_n$ does not have an infinite branch, we will be able to build from the names $\dot C_n$ a hereditarily symmetric name $\dot C$, witnessed by $T$ to be symmetric, that is forced by $q$ to be a choice function for $\dot F$.
Let $D_0$ be the dense set below $q$ of conditions $p$ such that for some name $\dot c_p\in {{\rm HS}}$, $p{\Vdash}\dot c_p\in\dot F(0)$. We will thin out $D_0$ to a maximal antichain over which we can mix the names $\dot c_p$ to get the desired name $\dot C_0\in{{\rm HS}}$. At the beginning, we choose some condition $p_0\in D_0$ and a name $\dot c_{p_0}\in{{\rm HS}}$, witnessed by a useful tree $S_0'$ to be symmetric, such that $p_0{\Vdash}\dot c_{p_0}\in \dot F(0)$. By extending $S_0'$, if necessary, we can assume that ${\text{dom}}(p_0)\subseteq S_0'$. Let $S_0$ be the union of $S$ and $S_0'$. Next, we choose some $p_1'\in D_0$, incompatible to $p_0$, and a name $\dot c_{p_1'}\in {{\rm HS}}$, witnessed by a useful tree $S_1'$, with ${\text{dom}}(p_1')\subseteq S_1'$, to be symmetric, such that $p_1'{\Vdash}\dot c_{p_1'}\in \dot F(0)$. At this point, it is not a good idea to union up $S_0$ and $S'_1$ because if we keep doing this we might end up with an infinite branch in the final tree. So instead, we first consider those nodes in $S_0\setminus S$ which have some new node of $S_1'$ sitting directly above them, meaning that we are extending an old branch that was already extended once from $S_0$. Call these nodes in $S_0\setminus S$ *bad* and the nodes of $S_1'$ sitting directly above them *bad successors*. Let $\pi$ be an automorphism of the tree ${{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1$ which moves every bad node $t$ (and its bad successors along with it) to some node outside $S_0$. Since $S_0$ is countable and each node in ${{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1$ has $\omega_1$-many successors, there are many nodes to which the nodes $t$ can move. Since $\pi$ fixes $S$, we have $\pi^*\in H_S$. So we have $\pi^*(p_1'){\Vdash}\pi^*(\dot c_{p_1'})\in \dot F(0)$. We can safely union up $S_0$ and $\pi{\mathbin{\hbox{\tt\char'42}}}S_1'$ without growing any branches in $S_0$ that have extended beyond $S$. But the problem now is that $\pi^*(p_1')$ may no longer be incompatible with $p_0$, since the incompatibility could have come from some node $t$ that got moved. This, however, can be easily remedied by adding back to $\pi^*(p_1')$ one of the nodes from $S_0$ which caused the incompatibility. Let $p_1$ be the condition $\pi^*(p_1')$ together with possibly one extra node from $S_1'\cap S_0$, let $\dot c_{p_1}=\pi^*(\dot c_{p_1'})$, and let $S_1$ be the union of $S_0$ and $\pi{\mathbin{\hbox{\tt\char'42}}}S_1'$. Using that the filter $\mathscr F$ is normal, it is not difficult to see that the image of a name $\sigma$ in ${{\rm HS}}$ under an automorphism $\pi^*$ from $\mathcal G$ is also a name in ${{\rm HS}}$, and moreover if some useful tree $T$ witnesses that $\sigma$ is symmetric, then $\pi{\mathbin{\hbox{\tt\char'42}}}T$ witnesses that $\pi^*(\sigma)$ is symmetric. So $\dot c_{p_1}\in {{\rm HS}}$ and $S_1$ witnesses that it is symmetric. Finally, note that ${\text{dom}}(p_1)\subseteq S_1$. In the next step, we choose some condition in $D_0$ incompatible with both $p_0$ and $p_1$ together with some name in ${{\rm HS}}$ it forces to be in $\dot F(0)$. We use an automorphism from $H_S$ to transform the witnessing tree into a tree that can be unioned up with $S_1$ without growing any branches by moving over nodes (outside $S$) that have new nodes sitting directly above them. We continue this process transfinitely for as long as we can find a condition in $D_0$ that is incompatible to all previously chosen conditions. At limit stages, we simply take the union of the increasing sequence of trees constructed up to that stage. The process must terminate by some countable ordinal stage $\alpha$ since ${\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1)$ has the ccc. Thus, the final union tree $T_0={\bigcup}_{\xi<\alpha}S_\xi$ must be countable and by construction, since we never allow any branch to grow more than once, $T_0$ cannot have an infinite branch.
Let’s argue that the tree $T_0$, because it contains ${\bigcup}_{\xi<\alpha}{\text{dom}}(p_\xi)$, witnesses that the mixed name $\dot C_0$ of the names $\dot c_{p_\xi}$ for $\xi<\alpha$ over the antichain ${\langle}p_\xi\mid \xi<\alpha{\rangle}$ is symmetric. Recall that $$\dot C_0={\bigcup}_{\xi<\alpha}\{( \tau,r)\mid r\leq p_\xi,\,r{\Vdash}\tau\in \dot c_{p_\xi},\tau\in\text{{\text{dom}}}(\dot c_{p_\xi})\}.$$ Fix an automorphism $\pi$ point-wise fixing $T_0$. It suffices to argue that whenever $(\tau,r)\in \dot C_0$, then $(\pi^*(\tau),\pi^*(r))\in \dot C_0$. So suppose $(\tau,r)\in\dot C_0$ and fix $p_\xi$ witnessing this. Since $r\leq p_\xi$, it follows that $\pi^*(r)\leq p_\xi$; since $r{\Vdash}\tau\in \dot c_{p_\xi}$, it follows that $\pi^*(r){\Vdash}\pi^*(\tau)\in \dot c_{p_\xi}$; and finally, since $\dot c_{p_\xi}$ is symmetric and $\tau\in\text{dom}(\dot c_{p_\xi})$, it follows that $\pi^*(\tau)\in\text{dom}(\dot c_{p_\xi})$. Now observe that, since each $\dot c_{p_\xi}\in {{\rm HS}}$, we have $\dot C_0\in {{\rm HS}}$.
Next, we let $D_1$ be the dense set below $q$ of conditions $p$ such that for some name $\dot c_p\in {{\rm HS}}$, $p{\Vdash}\dot c_p\in\dot F(1)$. We thin out $D_1$ to a maximal antichain over which we can mix to obtain the desired symmetric name $\dot C_1$ by the same process as above building the new trees as extensions of $T_0$, where branches that have extended beyond $S$ are no longer allowed to grow. Once, $\dot C_1$ and $T_1$ have been obtained, we proceed to obtain $\dot C_2$ and $T_2$ extending $T_1$, and continue in this manner for $\omega$-many steps. The final tree $T={\bigcup}_{n<\omega}T_n$ is still countable, cannot have an infinite branch by construction, and witnesses that the canonical name for the sequence of the $\dot C_n$ is symmetric.
A second-order arithmetic model of $\Sigma^1_{\infty}$-${{\rm AC}}$+$\neg\Pi^1_2$-${{\rm DC}}$ {#sec:SOArithmetic}
==============================================================================================
We will now argue that the reals of the symmetric submodel $N$ of the forcing extension $L[G]$ by ${\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega)$, constructed in Section \[sec:symmetricModels\], is a second-order arithmetic model of ${{\rm Z}}_2$ together with $\Sigma^1_\infty$-${{\rm AC}}$ in which $\Pi^1_2$-${{\rm DC}}$ fails.
\[th:SOArithmeticModelACnegDC\] There is a $\beta$-model of second-order arithmetic ${{\rm Z}}_2$ together with $\Sigma^1_\infty$-${{\rm AC}}$ in which $\Pi^1_2$-${{\rm DC}}$ fails.
Let $G\subseteq {\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1)$ be $L$-generic and let $N\subseteq L[G]$ be the symmetric submodel of ${{\rm ZF}}+{{\rm AC}}_\omega+\neg {{\rm DC}}$ constructed in Section \[sec:symmetricModels\]. Let ${\mathcal M}$ be the model of ${{\rm Z}}_2$ whose collection of sets is $P(\omega)^N$. Since ${{\rm AC}}_\omega$ holds in $N$, it immediately follows that $\Sigma^1_{\infty}$-${{\rm AC}}$ holds in ${\mathcal M}$. Thus, it remains to show that $\Pi^1_2$-${{\rm DC}}$ fails in ${\mathcal M}$. The result will follow if we can show that the tree ${\mathcal T}$ is $\Pi^1_2$-definable in ${\mathcal M}$. By Theorem \[th:uniquenessOfGenericsTreeIterationUncountable\], a sequence ${\langle}r_0,\ldots,r_{n-1}{\rangle}$ of reals is in the domain of ${\mathcal T}$ if and only if it is $L$-generic for ${\mathbb{P}}_n^J$. So we need to see that the statement ${\langle}r_0,\ldots,r_{n-1}{\rangle}$ is $L$-generic for ${\mathbb{P}}_n^J$ is $\Pi^1_2$-expressible in ${\mathcal M}$. This will suffice because the sequences in ${\mathcal T}$ are ordered simply by extension.
What we would like to say is that all $L_\alpha$, with $\alpha$ a limit ordinal, should satisfy the following. Given an ordinal $\beta$ we check whether stage $\beta$ was nontrivial in the construction of $\vec P^J={\langle}{\mathbb{P}}_n^J\mid n<\omega{\rangle}$. In this case, we check whether ${\langle}r_0,\ldots,r_{n-1}{\rangle}$ generates an $M_\beta$-generic filter for ${\mathbb{P}}_n^\beta$. Let $H_0$ consist of all trees in ${\mathbb{P}}_1^\beta$ which have the branch $r_0$. If $H_0$ is not $M_\beta$-generic, we are done. Otherwise, let $H_1$ consist of all conditions $p\in{\mathbb{P}}_2$ such that $p(0)\in H_0$ and $r_1$ is a branch through $p(1)_{H_0}$. If $H_1$ is not $M_\beta$-generic, we are done. Otherwise, we continue. So ${\langle}r_0,\ldots,r_{n-1}{\rangle}$ is ${\mathbb{P}}_n^J$-generic if whenever $\alpha$ is a limit ordinal, then $L_\alpha$ satisfies that for every nontrivial stage $\beta$ in the construction of ${\mathbb{P}}_n^J$, ${\langle}r_0,\ldots,r_{n-1}{\rangle}$ generates an $M_\beta$-generic filter. Now it is not difficult to check that the complexity of this statement is $\Pi^1_2$, because it says that for every set $X$, if $X$ codes a limit ordinal $\alpha$, then there is another set $Y$ coding $L_\alpha$ and a set coding the filters $H^\beta$ needed to verify for every nontrivial stage $\beta\in L_\alpha$ that $H^\beta$ is $M_\beta$-generic. The statement that $X$ codes an ordinal is $\Pi^1_1$ and the remaining statements are $\Sigma^1_1$.
Reflection can fail in models of ${{\rm ZFC}}$ without powerset {#sec:reflectionPrinciple}
===============================================================
Let’s call the *Reflection Principle* the statement that for every set $a$ and formula $\varphi(\bar x,a)$, there is a transitive set model $M$ containing $a$ such that $\varphi(\bar x,a)$ is absolute between $M$ and the universe. By the [Lévy]{}-Montague reflection, ${{\rm ZFC}}$ implies the Reflection Principle, namely every formula is reflected by some $V_\alpha$. It is natural to wonder whether the Reflection Principle holds in models of ${{\rm ZFC}}^-$, which may not have the $V_\alpha$-hierarchy.[^5] It has been suspected for a long time that the Reflection Principle can fail in such models. The question was first asked by Zarach in [@Zarach1996:ReplacmentDoesNotImplyCollection], and considered again in [@zfcminus:gitmanhamkinsjohnstone]. We will show that the Reflection Principle fails in the model $H_{\omega_1}^N$, where $N$ is the symmetric submodel of the forcing extension $L[G]$ by ${\mathbb{P}}(\vec P^J,{{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}\omega_1)$ that we constructed in Section \[sec:symmetricModels\].
The argument requires first seeing that the Reflection Principle is equivalent to a version of ${{\rm DC}}$ for definable relations. Following [@zfcminus:gitmanhamkinsjohnstone], let’s define that the *Dependent Choice Scheme*, abbreviated ${{\rm DC}}$-scheme, asserts for every formula $\varphi(x,y,z)$ and parameter $a$ that if for every $x$, there is $y$ such that $\varphi(x,y,a)$ holds, then there is an $\omega$-sequence ${\langle}x_n\mid n<\omega{\rangle}$ such that for all $n$, $\varphi(x_n,x_{n+1},a)$ holds; in other words, if a definable relation has no terminal nodes, then we can make $\omega$-many dependent choices according to it.
\[lem:RSequivDCScheme\] The Reflection Principle is equivalent over ${{\rm ZFC}}^-$ to the ${{\rm DC}}$-scheme.[^6]
First, let’s assume that the Reflection Principle holds. Fix a relation $\varphi(x,y,a)$ without terminal nodes. Let $M$ be a transitive set which reflects $\varphi(x,y,a)$ and let $R$ be the set relation on $M$ derived from $\varphi(x,y,a)$. Now fix a well-ordering $W$ of $M$ and use it to define an $\omega$-sequence of choices according to $R$.
Next, suppose that the ${{\rm DC}}$-scheme holds. Fix a formula $\varphi(\bar x,a)$. Observe that given any set $A$, using collection, we can argue that there is a set $\bar A\supseteq A$ which is closed under existential witnesses for subformulas of $\varphi(\bar x, a)$ with parameters from $A$. By taking the transitive closure we can assume that $\bar A$ is transitive. So consider the definable relation $R$ which says that $A$ is related to $\bar A$, whenever $A\subseteq\bar A$, $\bar A$ is transitive, and $\bar A$ is closed under existential witnesses for subformulas of $\varphi(\bar x,a)$ with parameters from $A$. We just argued that $R$ has no terminal nodes. Thus, we can make a sequence ${\langle}M_n\mid n<\omega{\rangle}$ of dependent choices in $R$. But then clearly $M={\bigcup}_{n<\omega}M_n$ reflects $\varphi(\bar x,a)$.
The theory ${{\rm ZFC}}^-$ does not imply the Reflection Principle.
Let $M=H_{\omega_1}^N$, where $N$ is the ${{\rm ZF}}$-model constructed in Section \[sec:symmetricModels\]. Clearly $M\models{{\rm ZF}}^-$ and since $N\models{{\rm AC}}_{\omega}$, choice holds in $M$, so $M\models{{\rm ZFC}}^-$. Since ${\mathcal T}$ is definable in $M$ (by the argument given in the proof of Theorem \[th:SOArithmeticModelACnegDC\]), it follows that the ${{\rm DC}}$-scheme fails in $M$. Thus, by Lemma \[lem:RSequivDCScheme\], the Reflection Principle fails as well.
Open Questions
==============
We end with some open questions related to the results of this article.
To produce our $\beta$-model of ${{\rm Z}}_2+\Sigma^1_\infty$-${{\rm AC}}$ in which $\Pi^1_2$-${{\rm DC}}$ fails we used the consistency of (a fragment of) ${{\rm ZFC}}$.
Does the consistency of ${{\rm Z}}_2 + \Sigma^1_\infty\text{-}{{\rm AC}}+ \neg\Pi^1_2\text{-}{{\rm DC}}$ follow just from the consistency of ${{\rm Z}}_2$? Does this implication hold with “consistency” replaced by “existence of a $\beta$-model”?
The third author showed in [@kanovei:ACnotDC] that it is consistent to have a model of ${{\rm Z}}_2+\Sigma^1_\infty$-${{\rm AC}}+\Pi^1_n$-${{\rm DC}}$ in which $\Pi^1_{n+1}$-${{\rm DC}}$ fails. The construction used countable-support iterations of Jensen’s forcing.
Can we obtain a new proof of the result using techniques of this article involving finite iterations?
Finally, we showed that the Reflection Principle can fail in a model of ${{\rm ZFC}}^-$ in which every set is hereditarily countable.
\[ques:KM\] Can the Reflection Principle fail in a model of ${{\rm ZFC}}^-$ with a largest cardinal $\kappa$ that is inaccessible in the model?
The theory ${{\rm ZFC}}^-$ together with the assertion that there is a largest cardinal $\kappa$ that is inaccessible is bi-interpretable with the second-order set theory Kelley-Morse together with an appropriate version of the choice scheme. Kelley-Morse is the set-theoretic analogue of ${{\rm Z}}_2$ in that it contains the comprehension scheme for all second-order assertions. The strategy for answering Question \[ques:KM\] involves defining a version of Jensen’s forcing for an inaccessible cardinal $\kappa$ and carrying out the rest of the construction to produce a symmetric model $N$ such that $(V_\kappa^N,V_{\kappa+1}^N)$ is a model of Kelley-Morse together with the choice scheme in which the dependent choice scheme fails. We will undertake this project in an upcoming article.
[^1]: Indeed, the much weaker theory ${\rm ATR}_0$ suffices for everything we have said so far. See [@simpson:second-orderArithmetic] (Section VII.3 and VII.4) for details.
[^2]: Note that this is not a standard definition. In the literature, a *perfect poset* is usually defined to be a collection ${\mathbb{P}}$ of perfect trees such that ${{}^{{{{\mathrel{\mathchoice{\raise2pt\hbox{$\scriptstyle<$}}{\raise1pt\hbox{$\scriptstyle<$}}{\raise0pt\hbox{$\scriptscriptstyle<$}}{\scriptscriptstyle<}}}}\omega}}}2\in{\mathbb{P}}$ and whenever $T\in{\mathbb{P}}$ and $s\in T$, then $T_s\in{\mathbb{P}}$.
[^3]: Abraham [@abraham:jensenRealsIterations] uses the terminology $\sigma$ is *consistent* with $p$.
[^4]: Abraham [@abraham:jensenRealsIterations] uses the terminology $\sigma$ is *bounded* by $F$.
[^5]: In the absence of powerset, the axiom of choice, defined as the existence of choice functions, is not equivalent the assertion that every set can be well-ordered [@zarach:unions_of_zfminus_models]. Here the theory ${{\rm ZFC}}^-$ is assumed to include the assertion that every set can be well-ordered.
[^6]: This equivalence holds regardless of whether we include the existence of choice functions or well-orderings as the choice axiom in our theory.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: 'We calculate the diameter and chirality dependences of the binding energies, sizes, and bright-dark splittings of excitons in semiconducting single-wall carbon nanotubes (SWNTs). Using results and insights from [*ab initio*]{} calculations, we employ a symmetry-based, variational method based on the effective-mass and envelope-function approximations using tight-binding wavefunctions. Binding energies and spatial extents show a leading dependence with diameter as $1/d$ and $d$, respectively, with chirality corrections providing a spread of roughly 20% with a strong family behavior. Bright-dark exciton splittings show a $1/d^2$ leading dependence. We provide analytical expressions for the binding energies, sizes, and splittings that should be useful to guide future experiments.'
author:
- 'Rodrigo B. Capaz$^{1,2,5,6}$'
- 'Catalin D. Spataru$^{3,5,6}$'
- 'Sohrab Ismail-Beigi$^4$'
- 'Steven G. Louie$^{5,6}$'
bibliography:
- 'var-condmat.bib'
title: Diameter and Chirality Dependence of Exciton Properties in Carbon Nanotubes
---
Diameter and chirality trends are one of the most useful concepts in nanotube science. Often, new physics arises when the diameter and chirality dependences of a given property are fully disclosed. A classic example is the analysis of “family patterns” in optical transitions combined with the diameter dependence of vibrational frequencies that paved the way to reliable $(n,m)$ assignments of single-wall carbon nanotubes (SWNTs) [@jorio013] and posed the fundamental “ratio problem” [@bachilo02]. Therefore, a reliable determination of diameter and chirality trends of a given nanotube property, even when this is accomplished by simplified models, is often as important as determining accurately that property for a limited number of tubes. Moreover, when a reliable model for trends is coupled with an accurate [*ab initio*]{} theory that determines its parameters, the model acquires quantitative and predictive powers.
The exciton concept solved the “ratio problem” [@spataru041], and it is now widely accepted that the optical spectra of carbon nanotubes are dominated by exciton features [@spataru041; @ando97; @chang04; @perebeinos04; @zhao04]. Recent experiments based on two-photon spectroscopy [@wang05; @dukovic05; @maultzsch05] and Raman spectroscopy on electrochemically doped samples [@wang06] have provided the first experimental evaluations of exciton binding energies for a few single-wall carbon nanotubes (SWNTs). However, a full description of diameter and chirality dependences of exciton properties in SWNTs has not yet been provided, neither experimentally nor theoretically. [*Ab initio*]{} calculations are restricted to a few small-diameter tubes [@spataru041; @chang04]. Perebeinos [*et al.*]{} [@perebeinos04] have extracted scaling relations of binding energies and sizes with diameter from model calculations, but the chirality dependence has not been addressed. Semi-empirical calculations have also been done for a larger variety of tubes [@zhao04], but again systematic diameter and chirality trends have not been extracted. Finally, the important issue of bright-dark exciton splittings have been addressed in detail by considerably fewer calculations [@spataru05; @perebeinos051; @chang06].
In this work, we calculate the full diameter and chirality dependences of exciton properties in SWNTs. We employ a symmetry-based, variational, tight-binding method, based on the effective mass and envelope function approximations [@knox; @brown87]. Since we explictly impose the symmetry of the exciton wavefunction, we can calculate properties of bright and dark excitons. Our model is parametrized by [*ab initio*]{} results. We calculate binding energies and sizes for the lowest-energy bright excitons (those usually associated with the E$_{11}$ singularity in the single particle joint density of states), as well as dark-bright exciton splittings for a large number of SWNTs. From these results, we extract reliable analytical expressions for the diameter and chirality dependences of such properties.
Our variational exciton wavefunction is written as: $$\psi(\vec{r}_e,\vec{r}_h) = C\sum_{v,c}A_{vc}\phi_{c}(\vec{r}_e)\phi_{v}^*(\vec{r}_h)e^{\frac{(z_e-z_h)^2}{2\sigma^2}},\label{exc:psi}$$ where $\phi_{c}(\vec{r}_e)$ and $\phi_{v}(\vec{r}_h)$ are conduction (electron) and valence (hole) single-particle states. The sum is restricted to the four band-edge states $c=\pm m$ and $v=\pm m$ from the top of the valence and bottom of the conduction bands. Note that both valence and conduction band edges are 2-fold degenerate for both zigzag and chiral tubes, once time-reversal symmetry is considered [@vukovic02; @barros06]. The single-particle wavefunctions are labeled by their quasi-angular momentum quantum numbers $+m$ and $-m$ and they are taken from properly symmetrized wavefunctions of graphene expanded in a $\pi$-orbital tight-binding basis. [@vukovic02] The coefficients $A_{vc}$, responsible for the quantum interference between pair excitations, are then completely determined by symmetry, as described in Table \[tab:1\].
Zigzag
-------------------------------------------------------- ------ ---------- ---------- ---------- ---------- ----------
Symmetry Deg. Activity $A_{++}$ $A_{--}$ $A_{+-}$ $A_{-+}$
$A_{1u}(_0B_0^-)$ 1 dark 1 -1 0 0
$A_{2u}(_0A_0^-)$ 1 bright 1 1 0 0
$E_{m',u}(_0E_{m'}^-)$ 2 dark 0 0 $\pm $1 $\mp $1
Chiral
Symmetry Deg. Activity $A_{++}$ $A_{--}$ $A_{+-}$ $A_{-+}$
$A_1(_0A_0^+)$ 1 dark 1 1 0 0
$A_2(_0A_0^-)$ 1 bright 1 -1 0 0
$\mathbb{E}_{m'}(k) + \mathbb{E}_{-m'}(-k) (_kE_{m'})$ 2 dark 0 0 $\pm $1 $\mp $1
: Symmetries, degeneracies, optical activities and coefficients $A_{vc}$ for excitons in zigzag and chiral tubes. The symmetries are described by the irreducible representations in both group of the wave vector [@barros06] and line group [@vukovic02](in parenthesis) notations. The label $m'$ is the quasi-angular momentum quantum number of the double-degenerate exciton. \[tab:1\]
We choose a gaussian envelope function. This choice is justified both by a fit of the [*ab initio*]{} exciton wavefunctions [@spataru041], as shown in Fig. \[fig:gaussian\], and by an analogy with the regularized Coulomb potential problem in 1D [@loudon59; @banyai87], for which the ground-state Whittaker function closely resembles a gaussian. The gaussian width or exciton size $\sigma$ is the only variational parameter in the problem. The constant $C$ normalizes the exciton wavefunction: $\int\int |\psi(\vec{r}_e,\vec{r}_h)|^2 d\vec{r}_e d\vec{r}_h = 1$.
![(Color online) Lowest-energy singlet exciton wavefunction of the (11,0) tube. Black thin line: [*Ab initio*]{} $|\psi|^2$ after integrating out on the coordinates perpendicular to the tube. Red thick dashed line: Envelope fit using a gaussian. Blue thick line: Envelope fit using a Whittaker function. Notice that the two fits are almost indistinguishable.[]{data-label="fig:gaussian"}](fig-gaussian.eps){width="7.5cm"}
We minimize the exciton energy that is composed of three terms: direct, exchange and kinetic energies. Here, we treat singlet excitons only. The direct term is written as: $$\begin{aligned}
<K^d> =\int \psi^*(\vec{r}_e,\vec{r}_h) V_C^{scr}(\vec{r}_e-\vec{r}_h) \psi(\vec{r}_e,\vec{r}_h) d\vec{r}_e d\vec{r}_h \nonumber \\
=C^2\sum_{vc,v'c'}A_{vc}^*A_{v'c'}\sum_{\vec{R}_1,\vec{R}_2}c_v(\vec{R}_1)c_{v'}^*(\vec{R}_1)c_c^*(\vec{R}_2)c_{c'}(\vec{R}_2)\times \nonumber \\
\times e^{-\frac{(Z_1-Z_2)^2}{\sigma^2}}U_{Ohno}^{scr}(|\vec{R_1}-\vec{R_2}|),
\label{eq:dir}\end{aligned}$$ where $V_C^{scr}$ is the screened Coulomb interaction and we wrote the direct energy in terms of the tight-binding expansion coefficients of the single-particle wavefunctions in a $p_z$-orbital basis $\varphi (\vec r - \vec R_i)$ centered in the atomic positions $\vec R_i$: $$\phi_n(\vec r)=\sum_i c_n(\vec R_i) \varphi (\vec r - \vec R_i).$$ The Coulomb integrals between sites are parametrized by the Ohno formula [@ohno64]: $$U_{Ohno}^{scr}(R)=\frac{U_0}{\epsilon\sqrt{(\frac{4\pi\epsilon_0}{e^2}U_0R)^2 + 1}}.
\label{eq:ohno}$$ The onsite Coulomb repulsion $U_0=16$ eV and the dielectric constant $\epsilon=1.846$ are chosen to reproduce the [*ab initio*]{} values for the binding energy and bright-dark exciton splittings for the (11,0) tube and kept constant for all other tubes. The exchange energy is given by: $$\begin{aligned}
<K^{x}> = 2 \int \psi^*(\vec{r}_e,\vec{r}_e) V_C(\vec{r}_e-\vec{r}_h) \psi(\vec{r}_h,\vec{r}_h) d\vec{r}_e d\vec{r}_h \nonumber \\
=2C^2\sum_{vc,v'c'}A_{vc}^*A_{v'c'}\sum_{\vec{R}_1,\vec{R}_2}c_v(\vec{R}_1)c_c(\vec{R}_1)c_{v'}^*(\vec{R}_2)c_{c'}^*(\vec{R}_2)\times \nonumber \\
\times U_{Ohno}(|\vec{R}_1-\vec{R}_2|). \label{eq:exc}\end{aligned}$$ In this case, the unscreened Coulomb interaction $V_C$ is parametrized by taking $\epsilon=1$ in Eq.(\[eq:ohno\]). Finally, the kinetic energy associated with the exciton relative coordinate is simply that of a gaussian envelope: $$<T>=\frac{\hbar^2}{4m^*\sigma^2}, \label{eq:kinetic}$$ where the exciton reduced mass $m^*$ is given by $1/m^* = 1/m_e + 1/m_h$. We use the diameter- and chirality-dependent electron ($m_e$) and hole ($m_h$) effective masses obtained from tight-binding calculations [@jorio05].
To test our model, we compare in Table \[tab:2\] our variational binding energies with [*ab initio*]{} ones obtained from solving the Bethe-Salpeter equation [@spataru041] for a few zigzag tubes. The agreement is excellent, except for the E$_{22}$ exciton in the (7,0) SWNT. This discrepancy can be understood: For the small-diameter (7,0) tube, the E$_{22}$ exciton size becomes extremely small ($\sigma = 5.3$ Å) and therefore the envelope-function approximation is not expected to be valid in this regime. Notice also that our model correctly captures the family oscillations in the binding energy.
-------- ----------------- --------------- ----------------- -------
Tube [*Ab Initio*]{} Model [*Ab Initio*]{} Model
(7,0) 0.89 0.87 1.13 1.61
(8,0) 0.99 1.03 0.86 0.92
(10,0) 0.76 0.68 0.95 1.09
(11,0) 0.76 0.76 (fitted) 0.72 0.75
-------- ----------------- --------------- ----------------- -------
: [*Ab initio*]{} and model binding energies for bright E$_{11}$ and E$_{22}$ excitons for a few small-diameter SWNTs. \[tab:2\]
Fig. \[fig:binding\] shows the binding energy of the E$_{11}$ bright exciton as a function of diameter for 38 SWNTs covering the full range of chiralities. The $(2n+m)$ family indices are indicated in the figure. One clearly sees the well-known family pattern reminiscent from the so-called Kataura plots for optical transition energies. [@jorio05] As expected, binding energies decrease with tube diameter. Chirality effects are also strong, contributing to about 20% spread in the binding energies for a given diameter. It is clear that excitons in $(2n+m)$ mod $3=1$ (MOD1) tubes have generally larger binding energies than in $(2n+m)$ mod $3=2$ (MOD2) tubes.
Fig. \[fig:size\] shows the exciton sizes as a function of diameter. Again, as expected, exciton sizes increase with diameter and they show the opposite MOD1-MOD2 trends as compared to binding energies. Notice that even for tubes as small as 0.5 nm in diameter the E$_{11}$ exciton sizes are already several times larger than the carbon-carbon bond, thus justifying the use of the envelope-function approximation.
![Binding energies for the lowest-energy bright excitons in 38 SWNTs with varying diameter and chirality. The dots are our model results and the red lines represent the analytical fit using Eq.(\[eq:analyt\]). The labels indicate the $(2n+m)$ families.[]{data-label="fig:binding"}](fig-binding.eps){width="7.5cm"}
![Sizes of the bright lowest-energy excitons for 38 SWNTs with varying diameter and chirality. The dots are our model results and the red lines represent the analytical fit using Eq.(\[eq:analyt\]). The labels indicate the $(2n+m)$ families.[]{data-label="fig:size"}](fig-sigma.eps){width="7.5cm"}
Analytical expressions for diameter and chirality dependences, although sometimes lacking a deeper physical justification, can be extremely useful for a quick evaluation of a variety of nanotube properties. We succeeded in finding simple yet very accurate analytical approximations for both binding energies and sizes:
$$\begin{aligned}
E_b=\frac{1}{d}\left(A+\frac{B}{d}+C\xi+D\xi^2\right) \nonumber \\
\sigma=d(E+F\xi+G\xi^2), \label{eq:analyt}\end{aligned}$$
where $d$ is the tube diameter in nm and $\xi=(-1)^{\nu}cos3\theta/d$ captures the chirality dependence [@capaz05]. The best fits are given by $A=0.6724$ eV.nm, $B=-4.910\times10^{-2}$ eV.nm$^2$, $C=4.577\times 10^{-2}$ eV.nm$^2$, $D=-8.325\times 10^{-3}$ eV.nm$^3$, $E=1.769$, $F=-2.490\times 10^{-1}$ nm and $G=9.130\times10^{-2}$ nm$^2$. These analytical fits are plotted in red lines in Figs. \[fig:binding\] and \[fig:size\], together with the numerical results. The agreement is nearly perfect.
Our theory also allows for an estimation of chirality and diameter dependences of exciton splittings among exciton states of the ground-state complex of the same E$_{ii}$. These splittings are fundamental to understand a variety of optical properties of carbon nanotubes, such as the quantum efficiency for light emission and the exciton radiative lifetime [@spataru05; @perebeinos051]. [*We find that the lowest-energy exciton for all SWNTs is the singly-degenerate dark state*]{}, due to its vanishing exchange energy [@spataru05]. Defining the exciton splittings from the lowest-energy exciton to the bright exciton and to the double-degenerate dark exciton as $\delta_1$ and $\delta_2$, we find the following dependence on diameter and chirality: $$\delta_i=\frac{1}{d^2}\left(A_i+B_i\xi+C_id\xi^2\right), \label{eq:delta}$$ with $A_1 = 18.425$ meV.nm$^2$, $B_1=12.481$ meV.nm$^3$, $C_1=-0.715$ meV.nm$^3$, $A_2=32.332$ meV.nm$^2$, $B_2=7.465$ meV.nm$^3$ and $C_2=-2.576$ meV.nm$^3$. So, in disagreement with Perebeinos [*et al.*]{} [@perebeinos051], we find the leading dependence of bright-dark splittings to be $1/d^2$. This is precisely the dependence of the exchange energy $<K^x>$ on diameter.
It is instructive to explain on physical grounds the leading dependences on diameter of the exciton sizes, binding energies and the bright-dark splittings. The exciton sizes $\sigma$ scale like $d$ because the 1D Coulomb potential is smoothed out or regularized over the scale of the tube diameter $d$ and this sets the length scale of the bound state (recall that in a pure 1D system with no lateral size, the Coulomb potential gives a delta-function ground state with infinite binding energy). The binding energies go like $1/d$ because $\sigma$ scales like $d$ and Coulomb interactions go like inverse distance [@perebeinos04]. The scaling of dark-bright splittings mirrors the scaling of the exchange energy $<K^x>$ which goes like $1/d^2$ because $<K^x>$ is the self-interaction of a neutral charge distribution with dipole moments: The long-range part (from distances larger than $d$) can be written as $\int_d^\infty dx / x^3 \sim 1/d^2$.
We now compare our results to the available experimental determinations of the exciton binding energies to date. Two-photon spectroscopy have been performed for SWNTs in a polymeric matrix [@wang05; @dukovic05] and in D$_2$O solution wrapped by a surfactant [@maultzsch05]. These environments should provide extra screening, so these results should not be directly compared with [*ab initio*]{} theory for isolated tubes. However, in our variational scheme, it is very easy to investigate the influence of screening and to adjust the dielectric constant $\epsilon$ to match the experimental results. In fact, we find that binding energies follow very nicely the scaling $E_b \propto \epsilon^{-1.4}$ proposed by Perebeinos [*et al.*]{} [@perebeinos04]. Therefore, it is straightforward to apply Eq.(\[eq:analyt\]) for SWNTs in [*any*]{} environment, provided that one scales the binding energies by using the appropriate phenomenological dielectric constant. For instance, taking $\epsilon = 3.049$ gives binding energies in excellent agreement (standard deviation of 0.02 eV) for all 13 SWNTs measured by Dukovic [*et al.*]{} [@dukovic05]. Similarly, the results of Maultzsch [*et al.*]{} [@maultzsch05] for 6 different SWNTs are reproduced with a standard deviation of 0.03 eV using a slightly larger dielectric constant $\epsilon = 3.208$.
In another recent experiment, Raman spectroscopy under electrochemical doping was used in nanotubes coated with a surfactant to give 0.62 eV and 0.49 eV for the binding energies of excitons associated with E$_{22}$ transitions in the (7,5) and (10,3) SWNTs, respectively. [@wang06] We have also calculated the binding energies of those excitons. It should be noted that for E$_{22}$ excitons, the MOD1-MOD2 oscillations in the binding energies are inverted, i.e., MOD2 tubes have larger binding energies than MOD1 tubes of similar diameter. In fact, in discrepancy with experiment, we find that (7,5) and (10,3) tubes should have E$_{22}$ excitons with similar binding energies, even though the latter has a larger diameter. By using $\epsilon = 2.559$ we find the best possible “average” agreement with experiment: 0.54 eV for the (7,5) and 0.55 eV for the (10,3) nanotube.
In conclusion, we have determined the full diameter and chirality dependence of exciton binding energies, sizes, and splittings in semiconducting SWNTs. All these exciton properties have strong diameter and chirality dependences, with a distinct family behavior. Comparisons between theoretical and experimental binding energies should be exercised with care, by ackowledging environmental screening effects. Our results should provide an useful guide to the interpretation of recent and future experimental determinations of exciton binding energies and other properties.
We acknowledge useful discussions with A. Jorio and T. G. Rappoport. RBC acknowledges financial support from the John Simon Guggenheim Memorial Foundation and Brazilian funding agencies CNPq, FAPERJ, Instituto de Nanoci[ê]{}ncias, FUJB-UFRJ and MCT. Work partially supported by NSF Grant No. DMR04-39768 and DOE Contract No. DE-AC02-05CH11231.
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
---
abstract: |
A Time Projection Chamber (TPC) is a powerful detector for 3-dimensional tracking and particle identification for ultra-high multiplicity events. It is the central tracking device of many experiments, e.g. the ALICE experiment at CERN. The necessity of a switching electrostatic gate, which prevents ions produced in the amplification region of MWPCs from entering the drift volume, however, restricts its application to trigger rates of the order of $1\,\kHz$.
Charge amplification by Gas Electron Multiplier (GEM) foils instead of proportional wires offers an intrinsic suppression of the ion backflow, although not to the same level as a gating grid. Detailed Monte Carlo simulations have shown that the distortions due to residual space charge from back-drifting ions can be limited to a few $\Cm$, and thus can be corrected using standard calibration techniques. A prototype GEM-TPC has been built with the largest active volume to date for a detector of this type. It has been commissioned with cosmics and particle beams at the FOPI experiment at GSI, and was employed for a physics measurement with pion beams.
For future operation of the ALICE TPC at the CERN LHC beyond 2019, where Pb-Pb collision rates of $50\,\kHz$ are expected, it is planned to replace the existing MWPCs by GEM detectors, operated in a continuous, triggerless readout mode, thus allowing an increase in event rate by a factor of 100. As a first step of the R&D program, a prototype of an Inner Readout Chamber was equipped with large-size GEM foils and exposed to beams of protons, pions and electrons from the CERN PS.
In this paper, new results are shown concerning ion backflow, spatial and momentum resolution of the FOPI GEM-TPC, detector calibration, and $\diff{E}/\diff{x}$ resolution with both detector prototypes. The perspectives of a GEM-TPC for ALICE with continuous readout will be discussed.
address: 'Technische Universität München, Physik Department, 85748 Garching, Germany'
author:
- Bernhard Ketzer
- 'for the GEM-TPC and ALICE TPC Collaborations'
title: 'A Time Projection Chamber for High-Rate Experiments: Towards an Upgrade of the ALICE TPC'
---
Gas Electron Multiplier ,GEM , Time Projection Chamber ,TPC ,Ion Backflow ,FOPI ,ALICE
Introduction
============
A Time Projection Chamber (TPC) [@Nygren:1978rx] is a high-resolution detector providing three-dimensional tracking of charged particles even in very high multiplicity environments such as heavy ion collisions, e.g. at the ALICE experiment at CERN’s LHC [@Alme:2010ke]. The large number of space points measured along each particle track considerably eases the task of pattern recognition for complex events, and in addition allows the identification of each particle by measuring its specific energy loss. Since track reconstruction requires a precise knowledge of the electric and magnetic fields inside the chamber, distortions have to be kept at a minimum. To this end, TPCs are normally used in a gated mode, where an electrostatic grid blocks the ions created in the amplification stage, usually an array of proportional wires, before they reach the drift volume. This introduces inevitable dead times and limits the maximum trigger rate to a few kHz. After the second long shutdown of the LHC in 2018, a luminosity for Pb-Pb collisions of $6\EE{27}\,\Cm^{-2}\,\s^{-1}$ is expected. In order to make full use of the increase in luminosity, ALICE plans to record all minimum bias events at a rate of about $50\,\kHz$ [@Musa:1475243], about two orders of magnitude higher than at present. As a result particle tracks from 5 events on average will be superimposed in the drift volume. A continuous, untriggered readout of the TPC is the obvious mode of operation, precluding the use of a gating grid.
A GEM-based TPC
===============
The use of Gas Electron Multiplier foils [@Sauli:97] instead of proportional wires offers an intrinsic suppression of the ion backflow [@Sauli:2003yf], although not to the level of $<10^{-4}$ as with a gating grid. Suppression factors around or below the percent level have been reached [@Blatt:2006nx], but strongly depend on the gas composition and the magnetic field. Together with low-noise readout electronics, allowing an operation at gas gains of the order of $10^{3}$, this results in about $10$ ions drifting back into the drift volume per electron arriving at the amplification stage. The drift distortions caused by the ion space charge have been shown in simulations to be of the order of a few $\Cm$, which can be corrected using standard calibration techniques [@Bohmer:2012wd; @Musa:1475243], provided the ion charge distribution is stable over time scales of the order of $\ms$.
In order to minimize distortions for the case of the ALICE TPC, which operates in a low solenoid field of $0.5\,\T$, measurements of the crucial ion backflow ($I\!B$), defined as the ratio of cathode to anode current, have been performed in triple GEM detectors with Ar- and Ne-based gas mixtures without magnetic field. Figure \[fig:IB-vs-Eind\_ar-70\_co2-30\] shows this quantity for an Ar/CO$_2$ (70/30) mixture as a function of the induction field, i.e. the field between the last GEM and the readout anode. Values of the $I\!B$ of about $0.8\%$ have been reached in a configuration with a low drift field, alternatingly high and low electric fields between GEMs, and the highest gain in the last GEM.
![Ion backflow, defined as the ratio of cathode to anode currents, in an Ar/CO$_2$ (70/30) mixture, as a function of induction field. The other fields are set to $E_\mathrm{drift}=250\,\V/\Cm$, $E_\mathrm{trans1}=6.0\,\kV/\Cm$, $E_\mathrm{trans2}=160\,\V/\Cm$, the potentials across the GEMs are $\Delta U_1=330\,\V$, $\Delta U_2=375\,\V$, $\Delta
U_1=450\,\V$, resulting in a gain of about $10^{4}$.[]{data-label="fig:IB-vs-Eind_ar-70_co2-30"}](Inductionfield-Ionbackflow.pdf){width="45.00000%"}
It is to be noted that such a configuration is not optimal for stability of the chamber against discharges in case of highly ionizing particles [@Bachmann:01e]. Moving to a Ne/CO$_2$ (90/10) mixture as currently used in the ALICE TPC requires a higher drift field ($400\,\V/\Cm$) and lower fields between GEMs ($<4\,\kV/\Cm$ in order not to enter a regime of charge amplification). As a consequence, the $I\!B$ is higher by about a factor of 5, as can be seen in Fig. \[fig:IB-vs-Et2\_ne-90\_co2-10\]. Steps to further decrease this value include adding small admixtures of N$_2$ to the gas and a fourth GEM, guided by microscopic simulations of charge transfer processes in GEMs [@GARFIELDPP:2012].
![Ion backflow in a Ne/CO$_2$ (90/10) mixture, as a function of transfer field 2. The other fields are set to $E_\mathrm{drift}=400\,\V/\Cm$, $E_\mathrm{trans1}=3.8\,\kV/\Cm$, $E_\mathrm{ind}=3.8\,\kV/\Cm$, the potentials across the GEMs are $\Delta U_1=225\,\V$, $\Delta U_2=235\,\V$, $\Delta
U_1=285\,\V$, resulting in a gain of about $10^{3}$.[]{data-label="fig:IB-vs-Et2_ne-90_co2-10"}](IB-vs-Et2_ne-90_co2-10_scale1_0.pdf){width="45.00000%"}
Strategies to eliminate the problem completely, e.g. by using foils with additional patterning on one side [@Lyashenko:09a] are under investigation, but are currently limited by the size of foils.
The FOPI GEM-TPC
================
In the framework of the PANDA experiment at FAIR a prototype GEM-TPC has been built which is the largest detector of this kind to date [@Fabbietti:2010fv; @Ball:2012xh].
![Cross section view of the FOPI GEM-TPC. A uniform electric field along the cylinder axis is provided by the cathode plane at the left end plane, the inner and outer field cage, and the top side of the first GEM at the right and plane.[]{data-label="fig:fopi.tpc"}](prototype.png){width="50.00000%"}
Figure \[fig:fopi.tpc\] shows a cross section of the detector, which has a drift length of $725\,\mm$, and an inner (outer) diameter of $105\,\mm$ ($300\,\mm$). Signals from the triple GEM amplification region are induced on 10254 hexagonal anode pads, chosen in order to optimize the spatial resolution for all possible track directions. The AFTER/T2K chip [@Baron:2008zz] is used to sample the analog signals at a frequency of $16\,\MHz$. The detector has been commissioned with cosmics and particle beams at the FOPI experiment at GSI, and was recently employed for physics measurements with pion beams. The performance of the GEM-TPC, operated with a gas mixture of Ar/CO$_2$ (90/10), fully matches the expectations with a spatial resolution of $\sim 230\,\upmu\m$ for small drift distances (Fig. \[fig:fopi.residual\]), improving the momentum resolution of the existing spectrometer by $30\%$ and providing more precise vertex coordinates with a resolution of better than $10\,\mm$ in $z$ direction. (Fig. \[fig:fopi.vertex\]).
![Residual width in $10\,\Cm$ bins of the drift length from cosmics tracks. The residual distributions have been fitted with a sum of two Gaussians; (red dots) sigma of the narrow Gaussian, (black dots) weighted mean of both Gaussians. Statistical error bars are included, but smaller than the data points. Systematic errors due to clustering and field distortions are not yet included. The dashed line is the transverse diffusion for single electrons.[]{data-label="fig:fopi.residual"}](res-vs-z_cosmics_v2.pdf){width="45.00000%"}
![Distribution of reconstructed vertices from $\pi$ interactions in a $10\,\mm$ thick C target, (red) using the central drift chamber (CDC) only, (blue) including the TPC.[]{data-label="fig:fopi.vertex"}](KalmanZ_nice.pdf){width="45.00000%"}
An important prerequisite for the measurement of the specific energy loss is the equalization of the gain across the active area of the detector. As in standard TPCs, this is achieved using decays of a $^{83\mathrm{m}}$Kr source introduced into the gas flow. Figure \[fig:Kr\_spectrum\] shows the corresponding pulse height spectrum before and after gain equalization. An energy resolution of $\Delta E/E=4.4\%$ for the main $41.55\,\keV$ peak is achieved in Ar/CO$_2$, and similar values are obtained with a Ne/CO$_2$ mixture. The first measurement of specific energy loss in a GEM-TPC over a wide momentum range is shown in Fig. \[fig:dEdx\_vs\_p\], which has been obtained from data taken with a pion beam hitting a C target. The energy resolution depends on the number of samples along a track, and reaches $15\%$ for the longest tracks, in agreement with expectations [@Allison:1980vw].
![Energy spectrum from decays of a $^{83m}$Kr source in the gas volume, obtained with the GEM-TPC in Ar/CO$_2$ (90/10) before and after equalization of the gain over the full active area of the GEM.[]{data-label="fig:Kr_spectrum"}](spectrum.pdf){width="45.00000%"}
![Specific energy loss versus particle momentum measured in the GEM-TPC in Ar/CO$_2$ (90/10) in a $0.6\,\T$ magnetic field. Here, a truncated mean algorithm cutting away the lowest $5\%$ and the highest $25\%$ of sample amplitudes was used.[]{data-label="fig:dEdx_vs_p"}](dEdx_vci.pdf){width="50.00000%"}
The ALICE TPC Prototype
=======================
A substantial upgrade of the ALICE detector is foreseen until 2018 in order to fully profit from the increase of luminosity for the investigation of hot and dense matter using rare probes [@Musa:1475243]. The upgrade also includes a replacement of the present gated MWPC-based amplification system of the TPC by GEM detectors, and a new continuous untriggered readout of the detector.
A first full-size prototype inner readout chamber (IROC) equipped with GEM foils has recently been designed and built (Fig. \[fig:alice.iroc\]). The GEM foils are of trapezoidal shape of $504\,\mm$ length and $292\,\mm$ ($467\,\mm$) width of the short (long) parallel sides, and have been manufactured at CERN using the newly developed single-mask technique [@DuartePinto:2009yq]. They consist of 18 individually powered sectors on one side in order to optimize the high-voltage stability of the chamber. They have been glued on one side onto $2\,\mm$ thick fiberglass frames with $400\,\mum$ thin spacer grids matching the sector boundaries. A new stretching tool making use of pneumatic frames for a homogeneous distribution of forces has been introduced for this purpose.
{width="45.00000%"} {width="25.00000%"}
The GEM IROC was equipped with three GEM foils with $2\,\mm$ transfer gaps and an induction gap of $4\,\mm$, matching the design of the FOPI GEM-TPC. For commissioning and testing the detector was placed in a test box equipped with a drift cathode and a field cage, providing a drift gap of $11.5\,\Cm$. After testing with sources and cosmics, the detector was equipped with $1000$ channels of readout electronics based on the PCA16 / ALTRO chips [@EsteveBosch:2003bj; @pca16:2009] and installed in a beam line at the CERN PS. Beam particles ($p$, $\pi^\pm$, $e^\pm$ of momenta between $1$ and $6\,\GeV/c$) were identified using a Cherenkov and a lead glass detector. Figure \[fig:alice.dEdx\] displays the energy loss distribution for $1\,\GeV/c$ electrons and negative pions measured in Ne/CO$_2$ (90/10) with the GEM IROC prototype. A resolution of $10.5\%$ is obtained for $1\,\GeV/c$ pion tracks with a most probable value of $60$ samples, nearly independent of the gain between $10^{3}$ and $5\EE{3}$, as can be seen in Fig. \[fig:alice.dEdx-vs-scale\]. For $1\,\GeV/c$ electrons, the resolution measured with the prototype is $9.5\%$ for all gain settings. These results are already in very good agreement with the value of $9.5\%$ measured for IROCs in the present TPC.
![Truncated ($5-70\%$) energy loss distributions for electrons and pions with $1\,\GeV/c$ momentum measured with the IROC GEM prototype at a gain of $5\cdot 10^{3}$.[]{data-label="fig:alice.dEdx"}](dE_dx.pdf){width="45.00000%"}
![Energy loss resolution for pions with $1\,\GeV/c$ as a function of the gain, represented by a scaling factor with respect to the settings mentioned in Fig. \[fig:IB-vs-Et2\_ne-90\_co2-10\].[]{data-label="fig:alice.dEdx-vs-scale"}](cGainRes_IBF_vs_Gain_Pio_m1.pdf){width="45.00000%"}
Conclusions
===========
A GEM-based TPC opens the possibility of using the powerful features of such a detector in high-luminosity environments, without the limits of a gating grid. It also presents challenges both to the detector and to the readout due to the extreme amount of data produced. A large GEM-TPC was built and successfully operated in the FOPI experiment at GSI, fully meeting the requirements such as improved momentum and vertex resolution, and yielding the first measurement of specific energy loss with such a device in a physics experiment. The ALICE experiment at CERN plans to replace the MWPC-based readout of the TPC with GEM detectors during the LHC shutdown in 2017/2018. A first prototype IROC was equipped with a triple GEM stack and performed well during a beam test, confirming the good energy resolution of such a device. Issues currently being studied in Ne/CO$_2$-based gas mixtures are the stability of a GEM detector in the harsh LHC environment, and the minimization of the ion backflow. The design of a first prototype GEM OROC (Outer Readout Chamber) is currently ongoing.
[10]{}
D. R. Nygren, et al., Phys. Today 31N10 (1978) 46.
J. Alme, et al., Nucl. Instr. Meth. A 622 (2010) 316.
ALICE Collaboration, L. Musa, et al., Letter of intent for the upgrade of the alice experiment, Tech. Rep. CERN-LHCC-2012-012. LHCC-I-022, CERN, Geneva (Aug 2012).
F. Sauli, Nucl. Instr. Meth. A 386 (1997) 531.
F. Sauli, et al., IEEE Trans. Nucl. Sci. 50 (2003) 803.
S. Blatt, et al., Nucl. Phys. B (Proc. Suppl.) 150C (2006) 155.
F. Böhmer, et al., [[arXiv:1209.0482 \[physics.ins-det\]]{}](http://arxiv.org/abs/1209.0482) (2012).
S. Bachmann, et al., Nucl. Instr. Meth. A 479 (2002) 294.
R. Veenhof, Garfield - simulation of tracking detectors, http://garfieldpp.web.cern.ch (2012).
A. Lyashenko, et al., Nucl. Instr. Meth. A 598 (2009) 116.
Collaboration, L. Fabbietti, et al., Nucl. Instr. Meth. A 628 (2011) 204.
GEM-TPC Collaboration, M. Ball, et al., [[ arXiv:1207.0013 \[physics.ins-det\]]{}](http://arxiv.org/abs/1207.0013) (2012).
P. Baron, et al., IEEE Trans. Nucl. Sci. 55 (2008) 1744.
W. Allison, et al., Ann. Rev. Nucl. Part. Sci. 30 (1980) 253.
S. Duarte Pinto, et al., Journal of Instrumentation 4 (2009) P12009.
R. Esteve Bosch, et al., IEEE Trans. Nucl. Sci. 50 (2003) 2460.
A. Kaukher, et al., EUDET-Memo-2009-08, [[EUDET-Memo-2009-08]{}](http://www.eudet.org/e26/e28/e42441/e78196/EUDET-MEMO-2009-08.pdf) (2009).
|
{
"pile_set_name": "ArXiv"
}
|
ArXiv
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.